The Implications Of Remote Work

Part 1:

>> Steven J. Davis: Welcome, everyone, to the Hoover Institution, Stanford Institute for Economic Policy Research, both at Stanford. They are our sponsors of this conference on the implications of remote work. I wanna welcome the audience that's here in person, as well as the sizable audience that we have joining us remotely.

Quite fittingly, given the topic. As everyone knows, the pandemic triggered a profound shift in working arrangements in many sectors, many jobs. Large numbers of people are now working from home or other remote location part of the week, or even all of the week in some cases. And that's had profound implications for worker well being, for how people manage their lives, for the economic fortunes of cities.

It's presented profound challenges to managers to deal with these new working arrangements and to organizations as they try to either adapt their workplace culture to these new working arrangements. Or, in some cases, to resist the new working arrangements, which present a different set of challenges. There's been an outpouring of research on these topics in the past three years on a huge variety of topics, and you'll see that reflected in the program for this workshop over the next three days.

Now I want to just very briefly thank our main sponsors, the Hoover Institution at Stanford, the Stanford Institute for Economic Policy Research, or SIEPR. They've provided generous funding for this conference and made it possible. They've also provided excellent staff support to make sure everything works behind the scenes.

So I'll thank them now cuz I know it's gonna go beautifully. My co-organizers for this conference, I wanna thank them as well. They are Nick Bloom, Emma Harrington, Natalia Emmanuel, Jose Maria Barrero. Sorry, Jose. He's only my co-author, but I still messed up his name. And Raj, Raj Chaudhary, okay?

So I think we'll probably hear from most of them over the course of the next couple days. So the way we're gonna start this off is Nick, Nick Bloom, arranged to interview Jack Nelis. We're gonna see a short video clip. For those of you don't know, Jack kicked off research on work from home, telework.

I think he coined the term telework only a mere 50 years ago, and the world is now catching up. And so let's hear from Jack and Nick, thanks.

>> Nicholas Bloom: I'm very happy to have Jack Nelis join us. Amazingly, it's 50 years since his groundbreaking project on telecommuting. That is like the seminal piece of research, so it's fantastic to have him.

And I wanted to quickly ask Jack three questions about his experience. So first, Jack, how did you get interested in it? You were so far ahead of your time. What made you look into telecommuting in the early 70s?

>> Jack Niles: Well, in the late 60s, I was basically a rocket scientist.

I was designing advanced reconnaissance satellites for the Air Force, and that got kind of boring, actually. So I started looking around to see how we might use this technology in the real world. And at one point, an urban planner says, if you guys can put man on the moon, and I had helped NASA find the landing places for the moon to get there.

He says, why can't you do something about traffic? I thought, whoa, that's a great idea. So I thought about it, and I said, there's amazing possibilities here. We look into it. So I talked to my execs at my corporation and said, hey, why don't we put together a project out, testing out this idea, getting people to work remotely?

And they said, what would it take? And I said, well, it's not just an engineering problem. Engineering is the least part of it. It's a lot of sociological things. So we'd probably have to hire a sociologist or two and some lawyers. And they said, forget about it. We're engineers, we don't do this touchy feely stuff.

So I got pretty mad at that, and I was grumbling to a friend of mine at the University of Southern California, and he said, why don't you do it here? So I invented a job at USC called Director of Interdisciplinary pProgram Development. And the key part of that was nobody knew what that meant, which is the key to my getting things to work.

But the idea was I would get faculty from different parts of the university to work on projects. Now, before I moved to USC, had also been working with the National Science Foundation to develop some of their applied research. And basically, the instant I moved to USC, I put in a grant to study the policy implications of the telecommunications transportation trade off.

That's sort of the title of the thing. We got the grant, and I said, I don't want this to be a thing where we just see if students can do it. I wanna try it in a real company. So we picked an actual company, got them to agree to set up a project.

Their reason was their turnover rate was too high. They're losing basically a third of their staff every year, and they had to recruit new ones and so forth. I said, if you put places near where they live so they don't have to deal with Los Angeles traffic, their turnover rates going to drop.

So let's try it, so we tried it. Ran it for six or eight months or so, turnover rate went to zero.

>> Nicholas Bloom: Wow.

>> Jack Niles: The productivity of the workers was up about 15%.

>> Nicholas Bloom: Wow.

>> Jack Niles: And said, hey, this works, guys. In the real world, it works, so this is great.

So you're gonna go ahead with this, right? Is it wrong? So why not? Well, because we're a non union company, and we feel that if we had our employees dispersed around the countryside, the unions could come in and pick them off site by site, and before we knew it, we'd be unionized, so forget about it.

A couple months later, I was talking to a strategic planner for the AFL CIO, and I explained telecommuting to him. He says, that's a terrible idea. I said, well, why is that? He says, well, If we have people scattered all over the place, how the hell will we ever organize them?

So turned down for two completely opposite reasons, this is basically the story of telecommuting ever since, until 2020.

>> Nicholas Bloom: That's amazing, we were talking earlier before we started recording that, I had heard you on podcasts and read some of your stuff. I mean, you're a legendary, and also, those findings are completely consistent with what we see now, that quit rates are down, it depends on the thing where productivity is definitely much more positive than people thought.

Actually, we had a paper last year in the remote work conferences showing that actually is harder to unionize remotely because people aren't connected up. That's astounding, so, I was gonna ask you, actually, we rolled on to the next question, what do you think of the pandemic, and what do you predict for the future?

 

>> Jack Niles: Well, all these years from 1973 to 2020, I was trying to figure out what's the lever I can use to get companies started in this, because over and over again, we get a company, they try a project and let it run two or three years in some cases, and then they'd get a new CEO who would say, I don't like this new stuff, and he'd shut it down.

And time after time, we were going to this, what is it I can use to get people to pay attention to the possibilities? And the tool was COVID-19 that was what made them actually pay attention. And all of a sudden, overnight, everybody had to work at home, because otherwise you kill each other with some devious virus or something.

So in a matter of weeks, all this happened I was thinking, man, this is not the way it should have been done, because ordinarily, when we set up a project, we'd spent three months figuring out what the implications could be in making sure not just cooking something up.

We spent a month or so training people to operate effectively in this regime, make sure that technology was right, and then try it and run it for several months. Changing from time to time, technology, whatever, to make it all smooth, at least as smooth as when you were in the office.

But in order to make it really stick, you had to have thousands of companies, as it turns out, forced into this full-time. So they had to pay attention for a couple of years, and lo and behold, this was no surprise, the employees loved it. We've never had trouble enticing employees to try this, they said, yeah, where do I start?

The problem is teaching managers to manage a different way than they're used to.

>> Nicholas Bloom: Yep.

>> Jack Niles: Management by looking around, I had arguments with Tom Peters. This was his motto, management by looking around, if anything's going wrong, you'll pick it up, said, no. You manage by deciding with the employees ahead of time what it is they're supposed to be doing, what their output's supposed to look like, when is the schedule, when's it going to be on time?

And then let them do it, it's that simple, and I've had managers come to me after they've been through the training and reading this for a while and says, I had this all wrong. I thought, you're making me manage telecommuters, it's my job, I'll never be able to do all this stuff, and instead, all the stuff I worried about went away.

 

>> Nicholas Bloom: Right.

>> Jack Niles: And the idea is the cup has changed from the guy with the title manager to the guy doing the work, it's his own responsibility to make sure the work's done because you all agreed on this. Surprise so that's where the big shift changes, I've done lots of forecast charts, including in the Telecast America 2000 report.

We had forecasts charts that shows the growth of this, but there's a big spike in now.

>> Nicholas Bloom: Yeah, I mean, it's amazing, everything you say, decades ahead is what I hear over and over again now and what we see in the research. And you're also right that there's one silver lining to come out of the horrible pandemic has just forced so many firms to shift to hybrid, often, if not fully remote.

But certainly it seems like it's a win-win, so, yeah, it's one of those things, it's an amazing moment that you think there must be other things in society that we've been doing wrong. We inherited this working practice, maybe going back to the industrial revolution, we never really got out of it, maybe five days a week.

I also think with my kids start huge summer holiday, we have this massive summer holiday because kids used to be released to go harvest the fields, but I don't know anyone's in Jordan harvesting anymore. Hey, so, Jack, again, I'm so thrilled to finally catch up, you're the legendary having been the person that was, it's very hard to find any area where someone is literally decades ahead.

And one final question is, are you at home? I guess you can tell I'm at home right now, are you at home as well?

>> Jack Niles: Yeah, I moved to home around 1970, no, 1989, and been there ever since. And I do not miss the commute to downtown Los Angeles, not one bit.

 

>> Nicholas Bloom: No, I'm a bit lucky, I'm actually gonna go in later this morning. I'm actually gonna go in and see Joanie, one of my students, present a fantastic paper she'll be giving in the conference. I could talk forever, in fact, your podcast I listened to was 45 minutes, it was fantastic, I actually re listened to parts of it.

So I wanted to say thanks so much for coming on, Jack, it's great to meet, virtually very appropriate both of us at home. And happy birthday, happy 50th birthday, I'm also 50, so I was born in the year this study first started. And it's amazing, it's hard to think of anything, even Nostradamus, I don't think you could predict that far ahead, so, thanks again.

 

>> Jack Niles: Thank you, and best wishes for your constant, I look forward to parts of it.

>> Nicholas Bloom: Yes.

>> Jack Niles: Okay.

>> Steven J. Davis: Thanks, Nick, thanks, Jack, that was great, there's three things I wanna remark on briefly. One, in surveys, when you ask people what they like about working from home, it's remarkable.

Avoiding the commute comes at or near the top routinely, so he was dead on the money on that. Second, and Nick indicated this, it's really striking how it took a massive shock to shift the equilibrium, and there were lots of forces resisting that shift, and those forces are still in play to some extent.

The third thing Thing. Despite what you just heard from Jack and Nick, I'm still really worried that when I'm not watching him, Nick's not working very hard. But I'm trying to get over that. Okay, so we're going to shift to the first session, which is titled managing remote work.

Here's the format. There's 30 minutes per paper, three papers. Speaker gets 20 minutes to present his and his coauthor's, or her and her coauthor's work. We'll limit things to clarifying questions in that period of time. And then we'll have ten minutes of kind of open Q and A and discussion.

So I think Raj Choudhury is going to present first paper. It's titled Office at Offsite: How Temporary Colocation Shapes Communication in Fully Remote Organizations.

>> Raj Choudhury: Raj, the floor is yours. My name is Raj Choudhury, and like Steve said, this paper is joint work with three amazing coauthors. So, Charles Ayoubi is a postdoc at HBS.

He's on the market right now, guys and he's amazing, okay? And then two faculty, co authors from INSEAD, Sujin Jang and Victoria Sevcenko. And this is the paper where we'd love to get your feedback either within the ten minutes or during the coffee break. So, feedback on everything.

So please. I'm looking forward to your comments and suggestions. So, the idea of the paper is motivated by the phenomena that now we have this new organizational form that has been called, in prior research, the all remote organization. So in a 2020 paper, we talked about this. So these organizations are typically a few hundred people.

The largest one I know of is GitLab, which is about 1500 people, and they do not have any offices. And in a recent paper, Jen Rhymer, who was a postdoc at Stanford, goes deep inside these organizations and she profiles them. And one of the things they do, for instance, as a management practice, is they are communicating asynchronously.

So they're using Slack, I know Brian is going to be here tomorrow, and similar asynchronous means of communicating a lot. And so I just made the slide to give you a sense that this is a really emerging phenomena. And in a separate paper, along with co authors, we are exploring the incentives of the founders of startups to adopt this model and a preview of that paper, among others.

A key incentive is that as a founder, if you don't have leasing costs on your balance sheet, you retain more equity at IPO. So it's a very strong financial incentive to adopt this organizational form. And our study is using data that one of these companies, Zapier, gave us.

So thank you, Zapier. But there are lots of challenges of these organizations, too. And one of them, of course, is communication. So one of the big benefits is that because you do not have offices and you're all remote, these firms can hire from wider labor markets. And I'm going to show you that's true in the case of Zapier.

But also because now you are hiring potentially demographically different and a diverse set of workers it's also a potential challenge, maybe. How do you make them communicate with each other? And there's a long literature in the distributed work literature in organizations, which has talked about how diversity among workers creates potential conflicts on communication.

And one of the ways this can be solved is by bringing these workers together in temporary colocation events. And I'm citing from my fellow organizers a paper where they essentially have data which shows that 45% of the workers who are fully remote, were meeting their coworkers in person at least monthly.

So it's not that these organizations that do not have offices, their workers never meet. They meet somewhere for short periods of time. But what we don't know, and correct me if I'm wrong, if I missed any paper, I would love to know that, that we do not know what effect these colocation events have on subsequent communication or performance by these workers.

So that's what we're trying to study in this paper. So this is our research question. Does temporary collocation among fully remote workers impact subsequent online communication, especially among workers who are spanning demographic boundaries? And the other paper I should cite very quickly is the recent nature human behavior paper, which showed that in fully remote settings, communication networks become more siloed.

So I'll preview our findings, since we have only 20 minutes. What we find is that these temporary colocation events, at least in our study, have a positive effect on subsequent online communication. However, these effects are more pronounced for workers, for dyads, which are demographically similar. And the only caveat to that second claim is that we find that small periods of constrained collocation, and I'll define how we measure that in our study, can help bridge demographic differences when people are meeting in person.

So personally, I'm excited about finding three a lot, because I think that's the unexpected finding in the paper. So this is our firm. It's a fully remote firm, no offices. They are a SaaS company, established in 2011 and have 281 employees at the time the study, we collected the data, which is end of 2019.

And this is the good news. They, of course, communicate a lot on asynchronous channels such as Slack. There's a lot of variation in terms of countries, time zones, and cities these workers come from. So for econometric purposes, that's great, because it gives us variation. But I think it also makes the more substantive point that all remote companies, even with 281 employees, they have 17 countries in their sample.

So they are indeed hiring from labor markets, which are pretty far away. This is the colocation event that we study. It's a retreat that happened in Orlando just before the pandemic. So this was in January 2020. So they brought the entire, or not the entire company. The retreat was open to the entire company, people self selected to join the retreat or not.

And I believe such retreats are commonplace among all remote organizations. They meet quarterly, in some cases bimonthly, some cases even monthly. So the data we have is, we have descriptive data for all 281 workers at that point in time, January 2020. We know all their communication on Slack for a period before and after the retreat.

So from October till May, we have entire access to the communication data. And the data we're going to use is from what they call the general Slack channel, so that we don't observe the communication that's happening within teams. We want to observe communication that's happening more between teams at an intracompany level.

And then we have the information on who attends the retreat in Orlando in January 2020. So our main dependent variable is for dyads, and I'm going to explain that unit of analysis in detail. We know whether they are communicating before and after the retreat. So since we have 281 employees, we have 78,000 plus dyads.

And the main independent variables are is The dyad, both of the members attending the retreat, their demographic differences. And I'm gonna talk more about this, whether or not they share cabs from the airport to the retreat location, because that's gonna be the way we measure collocation in a constrained space for a short period of time.

So this is our identification strategy, of course, the first thing I'll say is that attendance to the retreat is not random, people are self-selecting to attend the retreat, and we'll try to control for that in many ways. I'm gonna come to that, but the argument is that at least the cab rides from the airport to the retreat location are quasi exogenous.

And what we are actually doing is we are even taking a more constrained measurement of the cab rides. We're gonna only look at dyads which shared the same cab, but not the same flight, because if you are on the same flight and then you take the same cab, then potentially there are confounders about your location.

You might know each other, you may have spoken on the flight, so we're gonna exclude those dyads, we're only gonna focus on the dyads that are sharing a cab, but not the same flight. And I'm gonna show you the observable correlations between sharing a cab and everything we can observe.

And the argument that the company gave us was that these cab rides were composed by putting people together such that their flight arrival times were within 20 minutes. And that is the claim for why these cab rides are quasi exogenous, and then, of course, we have an extensive battery of controls which I don't have time to go through.

So these are some descriptive statistics, we see, of course, that the interactions post the retreat go up, and what is an interaction? It's one of three things, you either tag your colleague on slack, so I tag @Nick on slack, or I directly respond to a comment made by Nick, it's a reply, or I like something which Nick posted.

Because in slack, there are often multi party communication, we are parsing out the dyadic communication by looking at likes or tags or replies. And the main dependent variable is, do we communicate? It's a 1 or 0, I'll show you the level of communication later on as we go along.

So we find that the probability of communicating almost doubles, and then we have a lot of information about these dyads, how many times have they met at prior retreats which might be confounder what time zone they live in? The time zone difference, the difference in genders, ethnicity, nationality, and of course, whether they share a taxi, and the sharing cabs is only there for 3% of the dyads.

So I'll go very quickly through this, so these are the two quote unquote treatments whether you attended as a dyad, you both attended the retreat, that's column one, and whether you shared the cab. And I know the font is very small, so I'm gonna read it out for the cab sharing, the only correlation we find is whether you are from the same country or the time zone, and it's a negative correlation, of course.

And that is perfectly in line with the explanation that the cabs were organized within 20 minutes of the flight arrival times. So if you're living in faraway time zone countries, potentially your flight arrival times are different, you will not be in the same cab, hence the negative correlation.

So this is the baseline result, and I have four main results to show, and I'm gonna show the extensive margin results throughout the paper, we have the intensive margin in the appendix. So, as I said, as you would expect, attending, co attending the retreat, if both Steve and I attended the retreat, we are more likely to communicate after the retreat, and the probability increases about 100%.

So the probability doubles, then the two main tests we have is whether that interaction is likely to happen for demographically similar or demographically different dyads. And so this is the specification, and this is the table, I know you can't read it, so I'm gonna show you the main point estimates of the table that we care about.

So what we find is that if both the employees attended the retreat in the post period, so we are constructing this interaction both in the pre and the post period. In the post period, the probability of communicating goes to about 4%, the baseline was about 2%, so it doubles.

But as you can see here, this interaction effect with demographic differences is a negative 1.4%, so the effect is weaker if you are demographically dissimilar. How do we measure demographic differences? We are counting the number of differences each dyad can have, it can be 0 if you are absolutely demographically similar, or a maximum of 3, because we are counting demographic differences on nationality, ethnicity and gender.

And so if you take all 3 put together, this effect is about 35% of the effect, weakening from the average effect. So if a dyad has all 3 demographic indicators being different, the main effect gets washed out. So there's no advantage of the dyad attending the retreat, if they're completely demographically different, that's the main takeaway from this table and this shows up in the margins plot as well.

So, as you can see, the margins plot, this is for dyads of attendees on the left, you can see that the effect is much stronger, the blue line is no differences. The red line here is, if you have one difference on demographics, it could be gender or nationality or ethnicity.

And if you have all three differences, that's the orange line, then the effect weakens substantially. And these are the dyads on the right which did not attend the retreat, and of course, we don't observe, this is a counterfactual test, we observed no such effect of demographic differences, so that's the second result.

The baseline effect was communication goes up, but this shows that the effect is salient for dyads which are demographically more similar, but I think this is the most fun result, right? So, and this is the result that looks at dyads that share a cab versus diets that don't.

And we of course have the specification and the table, but I'm just gonna show you the margins plot. So what we find is that for the dyads that shared a cab, there is no negative effect, moderating effect, of the demographic differences. In other words, the dyads that share a cab, there is some connection that's built among people who are demographically dissimilar, and so this also shows up in the point estimate here.

So, I worked out today all the marginal effects on the back of this, so what we essentially find, let me summarize this table. For identical dyad that did not share the cab, but co attended the retreat, the increase in the probability of communication is about 4.9%. And for a dyad with one demographic differences, who attend the retreat and share the cab, the increase in communication is about 4.7%.

I have 5 minutes left, so, if I was a dyad with one difference, I get as much from the retreat if I shared a cab compared to a dyad that was completely similar and attended the retreat, so that's the. Sort of like the main finding from this table, and then I'll run through a couple of things and then just summarize.

So this shows the intensive margin, this is the number of interactions, and we find similar effects with intensive margin results. We find this is just looking at the performance of the workers and we find that ties to different genders, because my prior was, it's unclear whether having a diverse set of ties is good or bad for performance.

Turns out that it is, on average, good for performance to have a diverse set of social ties for your own performance, so two of them, the ties to different genders and the ties to different ethnicities, has a positive correlation on the. This is of course looking at the pre retreat performance, and we are still collecting the data on the post retreat performance.

So I'll summarize, I probably have a minute or two left, so the three empirical findings temporary collocation is associated with more online communication between pairs of workers. The main effect is driven by increase in online communication for workers who are demographically similar. However, these brief periods of constrained temporary collocation reverse this effect such that pairs are demographically dissimilar also benefit.

And then just since we have practitioners also, I thought I'll just draw a broad implications. I guess our research is pointing to these, for lack of a better word, we calling them liminal spaces, so if you have better word, please let me know. But we all go to conferences and we stand in the line, in the registration line, and we start talking to a stranger, or we go to the bar and we start talking to strangers.

So liminal spaces facilitate short, voluntary, informal in person interactions. And I would argue that these findings are also salient for large organizations that do hybrid work, because think of people coming to the office and then wearing speakers and just sitting in corners and working on spreadsheets. How do we facilitate ties between demographically dissimilar folks in the office, once again, and it'll be fun to test this out in an office setting.

And the final thing I'll say and stop is, we as a group have long observed, that the in person is happening in a heterogeneous set of venues. The downtown office is only one possible venue, folks are meeting in suburban offices, I see folks meeting in the gym that I attend.

And the retreats has been a phenomena for quite some time, and it'll be fascinating to compare these interactions and the effects thereof happening at these different venues. So I'll stop there, I look forward to your suggestions and thoughts, thank you.

>> Steven J. Davis: Okay, thanks, Rod, that was super interesting, I think we've got questions.

 

>> Jan K. Brueckner: Jan Bruckner from UC Irvine, I wonder if the firm pays the airfare to the retreat, cuz otherwise it's sort of like commuting cost, and that reduces the advantage of remote work.

>> Raj Choudhury: They do, the short answer is, and the argument they've given is we are saving so much money on not having offices, we can pay back a little bit in making sure people come to the retreats.

 

>> Callum Williams: Hi, Callum Williams from the Economist, just on your first conclusion, doesn't the research actually point to the importance of involuntary interactions rather than voluntary ones? And actually, the moral of the story of this research is when companies organize these retreats, you need to do more of these kind of cab rides, right?

So I'm just interested in your interpretation of the conclusion.

>> Raj Choudhury: No, absolutely Callum, and, the spirit was absolutely what you said, it's spontaneous things that people don't mind. And we are still trying to figure out what the cab ride represents and how we could generalize, but you're right.

And we've thought about the bar in Benihana, for instance, which is a Japanese restaurant where you sit down, have a couple of drinks, and talk to strangers. But the line, the registration line at conferences, so it is something that's very spontaneous, has a short period, so people are not gonna talk to the stranger for a long time.

And we're trying to find other examples that we can study as well, here.

>> Lindsay Relihan: I'm sort of interested in the construction of the dyad as the observation level, it's actually a very small company, just 281 people. And so you get a big sample size from the combinatorics of the state space.

So I'm really wondering, are we talking about a very small set of people who are doing very intense interactions? And so you're just getting this big effect that's coming from a very small set of pairs, or is something more widespread, and I'm Lindsay Rollingham, Purdue University?

>> Raj Choudhury: Yeah, so it's a great question.

So the reason our unit of analysis is the dyad is we're trying to observe connections people make through slack channel, but we are clustering for the dyad. So we use a two way dyad cluster, because without that, the standard errors are very, very small and everything is significant.

One thing we've thought about is multi employee communication, cuz right now we are super focused on the dyad, and three ways we are measuring the interaction is replies, tagging, and liking. But there's a lot of multi party communication also going on, which we are completely neglecting right now, so if you have ideas, please let me know.

 

>> Raj Choudhury: Steve.

>> Steven J. Davis: It seems like there's scope depending on what you know about the individuals for more along the lines of different demographics. So age is an obvious one, I don't know whether you have results along those lines. Do people also seek out others of their same age and they only talk to people or widely differing ages when they're kind of compelled to by circumstances.

Another one is marital status, and I really wonder whether the interactions between men and women and their impact, both their impact and the extent to which they're sought out differ between married and unmarried people.

>> Raj Choudhury: Both great suggestions, Steve, and I think we have those data in the survey we conducted for these employees, so we can look at that, yeah.

 

>> Nicholas Bloom: That was great, in fact, you don't often see research, it's also very relevant to the real world. So it's, well, in the short term, all research is very relevant eventually. It's like Keynes quote, I guess everyone's been thrilled to some long dead economist. One thing I wondered about is, what's the spillover effects?

A bit like Lindsey's thing, a little bit, which is for people that don't go. I realize you don't have much time, but is it the case that if Jose and you go to the conference, does it have any effective if I don't go on my communication? I could see it going either way, maybe there's crowding out, which means it's less obvious what the aggregate effect is.

Or maybe I say, hey, I've always wanted to talk to this guy Raj, I never met him, but Jose, I know you and you know him, can you connect to something?

>> Raj Choudhury: That's a great question, Nick, we can definitely look at that, yeah.

>> Speaker 9: Raj, thank you very much, super interesting.

So I was wondering more about the medium to long term implications for the firm in the sense that do you observe any sort of real outcomes? By real outcomes, what I mean is things like retention, job satisfaction, even maybe the productivity of workers?

>> Raj Choudhury: Yeah, so we are focused on the performance outcome, we have looked at the decay of the communication over time.

We don't observe a decay because that's a question around how frequently to organize the retreats. We can do more about thinking about the job satisfaction and other indicators.

>> Raj Choudhury: I think she had.

>> Alma Keshavarz: Hi, thank you so much. My name is Alma, I'm a student here at Stanford.

I'm really interested in how you foresee this study being expanded or kind of what would be the next step in terms of recreating it. Whether it would be interesting to look at the frequency of calls between coworkers that were at the same company. Or like General before was saying, about retention and possibly hiring patterns, whether interactions between hiring managers and workers of a certain background or demographic influences their hiring patterns in the future.

What do you think is the next step in this, with this specific question?

>> Raj Choudhury: Yeah, so a couple of thoughts come to mind. And, you know, if you have other ideas, please let me know. Even, you know, you can join the project if you have. So I guess the immediate question that we are thinking of is now the firm has grown since the study was conducted and whether we observe similar effects with a scale of close to 1,000 people.

So that's something that we are planning to. And this time we'll probably do randomized doughnuts at the next retreat. But I think the two questions I pointed at the end, especially for the PhD students, I would be very curious to see what differs and what's not different between the interactions in the office versus at other locations, whether the quality of those interactions are different and the effects of those interactions are different.

 

>> Steven J. Davis: Julia.

>> Julia Hobsbawm: Julia Hobsbawm from Bloomberg's, Working Assumptions and Work Ship. So are we talking about intimacy and trust being the real story? And therefore, are you going to measure for those that want people more back in the office why the temporariness is so relevant that you almost create stronger bonds, not just because you're in a taxi perhaps, but because you're only with someone for a short period of time.

 

>> Raj Choudhury: It's a great question. And, you know, there's been research on the interactions within an office, and, you know, I think the research dates to the seventies as well, which shows that the interactions are really spatially constrained and if there's a wall between the two employees, less probability.

So I feel we should definitely go back to the office and see what's happening there. And yeah, I agree with you. Do we have time for one last question? Yeah, so at the back.

>> Jeanette Garrity: Hi, Jeanette Garrity Siepere. Any place you want to go with this approach? To examine another issue on communication in this situation, specifically with management, between management employees you focused on between employees.

But there's a fundamental issue there. The other way.

>> Raj Choudhury: Yeah, so maybe it's a perfect opportunity to make a plug for another paper we're working on, and I think that paper was presented last time. So we also have an experiment of virtual water coolers where very senior mentors met new interns through Zoom.

And in that study also, we found that the effects were stronger if the mentor and the mentee were demographically similar. So I think demographics seems to be the common variable in these two studies. Sure.

>> Speaker 12: I don't know whether you saw it, but in the Wall Street Journal a couple weeks ago, there was a story about Hershey's Chocolate Corporation, which is located in a small town in Pennsylvania.

And their remote work arrangement is that people have to show up for, like, six weeks a year, and the rest of the year, they're free to live wherever they want.

>> Raj Choudhury: They have core weeks, what they're calling core weeks.

>> Speaker 12: Yeah, that's correct.

>> Steven J. Davis: Okay, thanks so much, Rod, that was a great way to kick off the conference following the Nick's interview with Jack.

So next up, I'm not sure who's presenting. Giorgio Presidente is presenting a paper with Carl Frey titled Remote Control aligning incentives in a world of remote work. Take it away.

>> Giorgio Presidente: Thank you, Steven. So this is joint project with Pia Andres atELSC and Carl Benedict Frey at the Oxford Internet Institute.

I don't need to spend much time with this audience about describing the rise of remote work and the fact that recent estimates suggest that about 15% of US jobs are gonna be a hybrid or fully remote. We start from this, from this observation, and in the paper, we argue that not being physically at the office might create information asymmetry.

In particular, the actions. The way the employee behaves is not directly observable by the employer. So their action or effort, in the language of the contract here, is hidden to the firm. And if you ask the principal agent model what you should do in these situations, it will tell you that you need to pay linked to performance.

So use performance pay that, in other words, you need to provide variable compensation that is attached to the output of the employee. So we ask, how often do remote jobs provide performance pay? So, in the paper, we study optimal contracts when people work from home in the principal agent model.

And we find that in a variety of setups, working remotely, working not at the office, generate moral hazard. And so the model will prescribe stronger reliance on performance pay. So, empirically, what we do, we use the near universe of US online job postings to characterize employment contracts. And to do so, we use dictionary methods, and we study whether companies increase or not reliance on performance pay, just as the principal agent model would prescribe.

We are gonna argue that our findings might have important managerial, but also macroeconomic implications. Not least because there is a huge in all the literature that emphasizes the insurance providing role of the firm. So basically, by paying employees a fixed base pay, at the same time the company's insuring against the income shocks.

And this is important because it's a kind of insurance that employees cannot buy on the market because there is moral hazard. So in a nutshell, what we do theoretically here, we look at two different ways of modeling remote work. The first one is pretty general. You can just assume that by working from home, you have a switch from a full framework to a partial information one.

Now, while this is general, it's a bit naive, because as you can imagine, even if you're at the office, your employer cannot come luckily into your head. So there is always some kind of moral hazardous to make it a bit tighter, a bit more specific. One way is to take the principal agent model, making a bunch of assumptions, specialize it a little bit, and model remote work as an increase and expansion in the set of competing non-business activities.

So the classic example here is that in many jobs, private calls are forbidden on the workplace. So if you're at the office, you can't do it, otherwise your boss will fire you or will notice it. Well, when you are at home, you can do many things that you are not supposed to do in the office.

And I mean, unless you have a very intrusive software monitoring you, you are gonna get away with that. In both cases, when you consider these different ways of modeling working from home, the model will tell you always the same thing. Link performance, link pay to performance so provide variable compensation.

So about the data. This study is based on online job postings from Lightcast. We consider a period between 2019 and 2022. There are two key elements of this data that we exploit here. The first is that we have the full text of the vacancy that we use to tag vacancies and find out if they provide a performance pay or not, and if the vacancy is on site or allows for remote work.

Plus, we have posting dates, employer name and location, and importantly, eight digit occupation codes. In the Lightcast, their own classification is very similar to the standard occupation classification. So this is the list of keywords that we use to identify performance pay vacancies. Now, the way we got this list is pretty simple.

So we analyze manually, we read a decent number of job postings and we selected the sentences the way vacancies tend to express to offer performance pay. And then once we do that, we use the list to tag each vacancy with a dummy variable equal to one if the vacancy mentions performance pay.

So this is a building block. We do the same thing for remote work, but here likely we have a bunch of papers that do that. And so we base our tagging procedure on this list of words. This is just an example of a vacancy that includes for which the dummies we use are both equal to one.

So this is a medical organization and the employee will have excellent commissions and bonuses and they must be able to work from their home office. Okay, there are issues. So first of all, false positives. There are some vacancies that provide sign on bonus, signing bonus or referral bonuses.

So we exclude this because these bonuses are supposed to attract the worker to sign the contract, but they do not really give incentives according to their performance. Then a similar issue here. Some vacancies use the word commission to describe an institution. For instance, American Joint Commission on Cancer.

We perform robustness checks by excluding the word commission from the list of keywords. Similar thing for remote work. So we have home office intended as a headquarter. We also untag vacancies that we identify as negating the keyword performance pay or remote work. I will say more on this later.

On. So once we harmonize all names in the data set from Lightcast, because sometimes they're inconsistent, creating imaginary employers, we harmonize that. We drop vacancies with missing information that don't allow us to tag the vacancies. We drop military occupations because it's not really clear what a bonus there is.

And we end up with over 100 million vacancies from 2,000,500 employers in 50 states plus Washington, DC, and around 1000 eight-digit occupations. So this is what we found over all the sample. The average share of remote vacancies, 8%. And the average share of performance pay vacancies is around 17.5%.

Now this chart presents the evolution over time of the two series. So the share of performance pay and remote job postings over time as a percentage of all vacancies in the data set. Now let's start from the black line that I'm sure is familiar to many of you.

So you have this shape you get when use the dictionary methods, basically very small, under 5% up to the pandemics. Then it increases and then it falls down around the end of the sample. Probably more relevant, more novel for this paper, is the share of performance paid jobs.

So we noticed that the share starts at the beginning of the sample, around 15%, then increases, but mildly until the last quarter of 2020, and then it accelerates after that. So for the narrative of the paper, we would have loved to see this increase a bit early, just coinciding with the increase of remote work.

That is not the case. Now, in the paper we discuss a couple of reasons why that is the case. We believe one interesting one could be the impact of uncertainty. Now, if you go back to the principal as a model, you will find that uncertainty about the economic environment mitigates the the incentive to give performance pay, because otherwise you need to insure the worker and give overall higher compensation.

And during the pandemic, there was a lot of uncertainty, especially at the beginning. So one way of rationalizing the lag of the increase is here to say that at the onset of the pandemic, you have so much uncertainty that firms will wait one second to revise the way they write down contracts, okay?

And they start to do so just when uncertainty about what is happening is going down. Now this chart shows the share of performance paid jobs by broad occupation category. Now what is encouraging is that, and perhaps not surprising, the occupation with the largest share is sales. That is intuitive, and the least is education on a library.

Now admittedly, whatever is in between is harder to rationalize, okay, because we don't have many priors, but at least the largest and the smallest make some kind of intuitive sense. We also perform another exercise. So we show the share of performance pay vacancies by occupation, but distinguishing between on site and remote.

And you can see that for all occupational groups, the share of performance pay and remote occupation is larger than one on site. So this is, it agrees with the principal agent model that remote jobs should be a stronger reliance on performance pay. Now, to make the analysis a little more formal, we need to run a bunch of regressions.

And so we need to define a unit of analysis. Here we decided to, I mean, we try to identify the decision center, so the entity that decides whether to give a remote job or not and whether to offer performance pay or not. So we define two entities here.

The first one, we call it a recruiter. That is a pair employer name and eight digit occupation. So, for instance, the Amazon hiring of engineers. And the second one, that is going to be our unit of analysis, is a pair of a recruiter and a state. This is because in our data, the same company and the same recruiter hires from different states.

That's in the majority of cases. And we are going to exploit that. Now, just to give you a sense of what our The units of analysis we are using here are a few examples. So Amazon Virginia hiring of computer engineers, same entity, so Amazon Virginia hiring of data scientists.

And you can do that for various firms when you fix the firm and the occupation and you change the location or you change everything, so just the same company in different states and different occupations. We experiment with two kind of identifications. The first is very simple and is at the vacancy level and is a linear probability model, essentially linking the dummy for performance pay with the dummy for remote work.

And if you look at it, it's a kind of a repeated cross section, this one. So for a given branch and posting date, the beta, our coefficient of interest is identified by the variation, the cross variation in remote work and performance pay. We interpret beta year as a probability, a conditional probability, that is the probability of offering performance pay given that the vacancy is remote.

And we classed the error here at the company level, but we have so many observations that everything tends to be significant in this exercise, so it doesn't really matter. So these are the results of the vacancy level identification, if you put number in perspective here, the model is saying that around 17% of vacancies offer performance pay.

So the probability of offering performance pay of on site vacancies is 17% and remote vacancies are 5.4 or 30% more likely to offer performance pay, okay? So this is in line with our expectations. But there are measurement errors, in particular, a recent paper documents that there are many, many false positives if you use dictionary method, I've already mentioned this before so I will not repeat it.

The problem for us is that that creates attenuation bias, so we underestimate beta if there are measurement errors. So there are two ways to go, either you design a fancy language model that has been done, or your more pedestrian way. You find an instrument that is correlated to remote work but is uncorrelated to the measurement errors.

So the way we approach this is as follows, so first of all, we aggregate things up at the recruiter level. So we are gonna work with the share of remote jobs and the share of performance paid jobs at the branch level and date, of course. And this result in a standard two-way fixed effect estimator and beta is still interpreted as a conditional probability.

Now we're gonna exploit the fact that the same recruiter has branched in different US states, plus the fact that different states mandated strict workplace closing rules at different times. So in our sample, 85 recruiters are in more than one state, 7% in on state, and on average each recruiter in 22 states, so we have a substantial amount of variation.

Now, what is important? 36 states impose workplace closing for all non-essential services on March 2020. Four States, followed by a lag between one and seven months, we call these states laggards, and 11 states did not mandate strict workplace rules, they only imposed milder restrictions. Now, you can see that this is the average difference in the share of remote job postings in mandating versus non-mandated states.

Now, you can see here that this is an index, so it's zero before the pandemic, you have this strong increase at the beginning of the restrictions. And then as the non-mandating states also started to issue remote vacancies, this thing goes to zero, converts to zero. Now so based on that, we define this treatment variable equal Adam, equal to one, if a state mandates a workplace closing rule and to clean up things, we drop laggard states.

So that our control group is states that never mandated strict workplace closing rules, while the treatment group is those that implemented it on March 2020. We cut the sample in the last quarter of 2020 here to avoid this part, this confounding effect of the catching up by the non-mandating states.

And we end up with about 10 million branches that are observed over time, on average, three times in our data set. And we run a standard two stage least square estimator here, this is the first stage, in the second stage, our parameter of interest is this phi that is still interpreted as a conditional probability.

Okay, these are the results in column one, there is the first stage is decently tight. So states in which strict workplace closing routes were implemented had, on average, 1.3 percentage point higher share of remote work, remember, that's around 80% on average in the sample. And the second stage, in column two, we find that the correlation between the share of remote jobs and the share of performance pay at the branch level is about 0.36.

So let me sum up and discuss the findings. So, as predicted by the principal agent model, we do find a positive and significant correlation between remote work and performance page of postings. At the vacancy level, we find that remote work vacancies are roughly 30% more likely to offer performance pay.

The within branch correlation between the shadow remote jobs and the shuttle performance pay is 0.36. So this is consistent with some attention buyers not much so slightly larger than 30%, and with these numbers, we can compute the average marginal effect of remote work that is about 0.03. That if we combine with the average share of performance pay vacancies, we find that remote work is associated with a 16.5% increase in the share of performance paid jobs.

Now, it's a pretty large number if you think about that, because these are jobs that could have been fixed base pay and now there are variable compensation. So this might have important implications for macroeconomic stability because now the workers have less access to insurance, that means that their behavior will change, so access to credit will change and even labor supply.

So our preliminary results suggest that one unintended consequences of remote work might be that one to increase macroeconomic volatility, and that's it.

>> Steven J. Davis: Thanks, Giorgio. Two questions at the outset. One a clarifying question, it wasn't clear to me how you handled ads that say you can work from anywhere.

They're in the data set, we know they're in the data set, we struggle with how to use them as well, I'm not sure how you're assigning the location of such jobs. Then on the substance of what you're doing, you've got a tremendous amount of data here which I think you can exploit even more fully than you did.

So your basic specification, where you are. Examining the responsiveness of the movement to performance pay as a function of whether you're adopting remote work. You can carry that out at the occupation level, and maybe not for all 1000, but certainly for some tightly defined set of such occupations.

And then you can tell us where this responsiveness of performance based pay. What are the kinds of occupations wherever that response coefficient is really big? And what are the kinds of occupations where it's small? And you could maybe tell us something about the character of the work, the nature of the educational requirements for the work.

I think you can paint a much richer picture here on what you've already got is a really, really interesting study.

>> Giorgio Presidente: So I just answer one by one so that I don't miss anything. So for the people that could work from anywhere, that's just remote work. What we care there is about the employer location and the employer.

We know it even if the people work from everywhere. I don't know actually how relevant it is in the timeframe that we consider, because this at the end is just at the beginning of the pandemic. So I'm not sure there were many vacancies, but you might know more than me in this case because it was just the beginning, so they weren't so flexible.

Point taken, absolutely. And then we will have to rely on the characterization of occupations, for example, from ONET that you have some kind of certain standard indicators, like average age of years of education and things like that. But what I think is also an important next step is that to select those occupations where this fact of not being at the office creating information asymmetry is really relevant.

Because, like managerial occupation is less relevant than, I don't know, a call center employee, for instance, at least a priori. Natalia.

>> Natalia Emanuel: Natalia from the New York Fed. I understand this is not in their current dataset right now, but if you extended it forward where we can see the difference between hybrid and remote, it would be interesting to see if you also see a difference there in performance pay.

Because hypothetically, if I'm seeing my employees one time a week, two times a week, then I need a little bit less of this sort of remote control, as you call it. So it'd be very cool to see if that sort of, there's a much larger effect when people are fully remote versus hybrid.

 

>> Giorgio Presidente: Absolutely, thanks. That is on the to-do list, provided we can extend enough just to see really many hybrid at least defined as hybrid posting, thanks.

>> Speaker 15: So two quick thoughts on the exclusion criteria. I was little concerned that the mandating states might be drawing both employers and employees, and that might be correlated.

So something to think about, and I'm not saying there's any perfect instrument. But the other idea, I loved to build on Steve's point, even within the sample of firms that are posting remote work jobs, I wonder if you can exploit the variation there. And seeing whether the firms which are going remote but are not monitoring or doing performance pay contracts are doing differently than the firms that are posting remote jobs but not doing performance.

And Revolio Labs is one data source which I've recently come across. They have amazing data on all kinds of things. So happy to chat.

>> Giorgio Presidente: Thank you, yeah, sure.

>> Speaker 16: So, really, really fantastic paper. I hate to repeat a theme. I'm also gonna make two points. So one was it links to an old paper, I'm not sure if you know it, which was by Tom Lemire, Daniel Parente, Bentley McLeod, which I think came out in the QJ in 2009.

I used to teach it a while ago, which shows they claim that performance pay is a factor increasing income inequality. So I can't remember what this papers a while back that adding that into the mix there. One takeaway from this is remote work, in the long run, is a driver of rising inequality.

Interesting, they have not, I think they claim, about a quarter of rising inequality is from increased performance pay. So that's one, literally. The other thought is that something I've seen, and it's great to kind of numerically estimate it, is that remote work is better suited for more performance payable jobs.

So when I did like the call center, it was very natural to look at call centers cuz it's perfect for performance pay. There's another take in your paper, which is, can you look at, if you did like a dingle and Niemann type analysis for how performance pay amenable each occupation is, can you use that to explain any of the increase in remote work?

You could say pre-pandemic was basically none. It's gone up. Is it the case? It's gone up a lot in occupations performance pay. That second bit is kind of relevant. There's a big debate right now, like the federal government or the education sector, how much remote work they should have.

Your work suggests a little bit, maybe less than you'd think otherwise, cuz, like, if I'm the federal government employee, you can't really have performance pay. So there's a lot of kind of really big picture, quite broad implications of this.

>> Giorgio Presidente: Well, so the first part of the comment.

So the other way around, essentially. Yeah, sure. Obviously we are not talking about causality on anything, it's just correlation. So we don't really care about, we don't see it as a problem. But I think you're right, that could be relevant also the other way around and about inequality, that would be an additional macroeconomic implication of this, right?

So thanks, one and two.

>> Jose: Yeah, Jose from I was going to suggest if you could think hard about measuring other features of the employment contract that might matter. Like whether contracts that are temporary and can become permanent or contracts that, or ads that specifically say that promotions are explicitly linked to performance.

Because, I mean, yeah, so monitoring is definitely, and principal agent issues are definitely something, but to look at in this context. But there's other features of the employment contract that employers could be adjusting, and it might be hard to train your model to identify.

>> Giorgio Presidente: Yeah, that empirically for sure.

But even in my admittedly limited understanding of the principal agent word, it starts to get complicated to include issues as how long is the contract? Unless you can reduce it to something more standard. I don't see an obvious way to model it directly. But empirically, yeah, absolutely. They might be using other levers.

And in fact, what I'm saying, you can model remote work as a set of expanding tasks. Actually, that literature at the beginning wasn't taking that as exogenous, was stating that set as endogenous. So how much freedom can you give the employee to ensure that they land? So, yeah, absolutely.

And then, Emma.

>> Andrii Parkhomenko: Andre Perhomanko, University of Southern California. So I have a question about this finding of a decline in remote postings in recent months. And I believe Jose, Nick, and Steve have a paper that uses the same data, and they have a similar finding. Is it possible that this is not really a decrease in postings that advertiser won't work?

Maybe it's just a new normal, right? So work from home is such a normality now that firms even don't talk about it. We've seen this in some occupations even before COVID, you wouldn't advertise remote work in an academic position, for instance?

>> Giorgio Presidente: Yeah, sure, I mean, if you ask them, they will tell you that it's just mistagging of the vacancies, that that's there's a large employer changing the way they are advertising stuff, and that's why you have that decline.

That is a likely explanation. We care only partially because our analysis, if you remember is stopped before for that decline for other reasons, but also for that. So we have a measurement, a mismeasurement problem, certainly because we are using somehow a rough way to tag data, but our time horizon is so limited that we care less about what happens after the last quarter of 2020.

 

>> Speaker 19: Great presentation. I was just curious if there were any way to get it like a back of the envelope or do some hypothetical choice experiment type thing where you got at people's willingness to pay for a contract that allowed them to be remote but have performance pay.

Because it does seem like this may be a way that it's like people think of remote work as an amenity, but if it's really always tied with performance pay, it'd be interesting to know how workers would feel about that bundle compared to an on site job.

>> Giorgio Presidente: Yeah, thanks.

Yeah, that would be a good way to go forward then so is performance pay an amenity?

>> Speaker 19: I assume it's a dis amenity. Yeah, but for the average worker, it should be a dis amenity since it's adding a lot of uncertainty to their contract.

>> Giorgio Presidente: Maybe we can do something in that direction, because we do observe the base salary for some occupation is a much smaller set of vacancies and pretty noisy, I have to say.

That's why we didn't use it yet but maybe there is a way to use it in the sense you're suggesting. Thanks.

>> Lindsay Relihan: Lindsey Bellingham, Purdue. My comment relates a lot to Nick's I'm wondering about the interpretation of the fact that you're finding is one of being like employers are recognizing this is an issue, and so providing more performance pay, but this is all happening in the context of an extremely competitive labor market that's varying by different kinds of occupations that are more or less remote workable.

So I'm sort of thinking about, like, all the other things that are changing for these kinds of workers, like you're now entering a national labor market rather than a local labor market when you have a remote job posting. So how do we think about alternative explanations for the results that you have and ways to exclude those other ones?

 

>> Giorgio Presidente: Sure. Thanks. Well, on the one hand something that has to do with the wage pressure could be related to the lag we observed. In particular, at the beginning of that phase there was positive wage pressure because there was scarcity of workers then I don't know after that what happens.

So that phenomenon could be used to explain the lag. I've tried to think about other channels linking the fact that you work remotely to performance pay other than purely information asymmetries and so incentives. I could not come up with one. If you have some, please let me know.

Thanks.

>> Steven J. Davis: Okay, thanks so much for another really stimulating talk. Our last paper who's presenting Davide? Davide Rigo gonna present a paper titled work-from-home and firm's resilience evidence from the Covid-19 pandemic and I will let you mention your co authors.

>> Davide Rigo: Yeah, just to try this, okay it works.

Okay, Hi everyone. First of all, let me start by thanking the organizers for including our paper in this program. This is joint work with Filippo Boeri and Ricardo Crescenzi, who are both my colleagues at LSE. This is pretty much preliminary work, so we are really looking forward to your feedback and comments that can guide us to develop further our analysis.

We are all aware that the Covid-19 pandemic has forced a large fraction of firms in modern economies to switch to remote work. However, little is still unknown about the impact of work from home on firm performance, especially in response to the crisis. The answer to this question is not straightforward.

We know that some firms were ready to take advantage of this shift to remote work, instead, some other firms they didn't have the complementary investment and the digital skills, and this may have disrupt their operations. The motivation behind this paper is that we observed that in the literature so far there has been an increasing number of work studying these questions.

However, we think that most of the work has been focused on one specific firm or context, which is great when you want to precisely identify a specific channel. But obviously it fails to account for the potential uneven effects of remote work across different types of firms in different places.

The second observation is that studies that try to enlarge the set of firms covered, most of the time they fail to account for unobservables/. For example, the literature on technological adoption shows that the better managed firms are the ones that are able to reap the benefits of digital investments.

So it could be that firms that have high quality of management are actually the one that are more efficiently adopting work from home and at the same time performing better during the pandemic. So we try to fill these research gaps by answering to both questions. How do we do that?

Well, we use unexplored administrative data for the universe of employees working from home in Italy, and with such a data we provide a comprehensive analysis of the relationship between work from home adoption during the Covid-19 outbreak and firm performance. We exploit a difference in difference framework by comparing firm sales before and after the pandemic between a group of firm that adopted work from home and group of firms that didn't do that, and we instrument work home adoption with the local availability of broadband Internet connection.

So far, we have found some preliminary evidence that actually work from home adoption has a negative impact on firm responses to the crisis, and also we find that which is kind of expected complementary investment in digital technologies such as, purchasing laptops and setting up servers before the pandemic has a crucial mitigating effect in this relationship.

 

>> Davide Rigo: So let me talk about the data. We have this incredible source of information provided by the Italian ministry, and why so? Is because Italy is a very bureaucratic country, and employers they have to declare when the workers. Workers work remotely because they need to guarantee their insurance coverage.

So every year, employers, they need to pay an insurance price to the state, which depends on the type of activities that the worker perform. As such if the worker is working from home, his risk of accidents or disease is lower, and so he has to pay a lower insurance premium.

So on top of being a legal duty, that's also a monetary incentive to make such declaration. And on top of this, we also know that since the start of the pandemic, the procedure has been simplified. And employers, they only have to fill out an online form to do that.

Oops, sorry.

>> Davide Rigo: Okay, we also think that Italy is a great setting for studying this question. We can see the number of workers working remotely that jump six time with the start of the pandemic, which coincides with the national lockdown imposed by the Italian government in early March.

And Italy was the first advanced economy that imposed these restrictions, if you remember. What is interesting from this is that the pre pandemic level is quite low. So, Italy was really lagging behind in terms of the so called high city revolution, and also in terms of work from home adoption.

 

>> Davide Rigo: Okay, we matched this data with firm level information from Orbitz, which is a quite popular database providing balance sheet information for the quasi universe of Italian firms. We have the population of medium and large firms, and we observe more than 80% of micro firms. We also web scrape information on local availability of broadband Internet connection from an Italian website created by the government to monitor the rollout of broadband in recent years.

And we have such information at a very granular level. So in Italy, there are more than 400,000 census areas, this is an example. On the left, you have four municipalities. The first one is my hometown, which is a little village of 8,000 people. And as you can see, it's completely covered by fiber in 2019.

And then we have four other municipalities which are partially covered by the fiber technology, so by high speed broadband connection. To give you a comparison, each polygon depicted with a gray line is a sensuous area. And we have such information about fiber technology at that level of granularity.

Here is the campus that we are today holding this conference, and we can see a buffer, an equivalent buffer for the mean census area. Actually, the median census area is even smaller than that buffer. So it's really granular, if you think about it.

>> Davide Rigo: The last piece of information is about ICT, we bought this data from a private provider, which is the best, to our knowledge, source of information of ICTs investment at the firm level.

This is our empirical strategy, we are looking at changes in sales, which is our measure of performance between 2020 and 2019. And we are looking at two measure of work from home adoption, One is extensive margin, what we call the extensive margin of work from home, which is.

Which equal one when the firm switch to remote working during the pandemic. And intensive margin is the share of work as remote work between 2020 and 2019. Then obviously we want to try to make these firms as comparable as possible. So first of all, we control for a set of firm characteristics, such as firm size, labor productivity, age, and average wage.

And on top of that, we control for industrial differences, also accounting for those industries that were allowed to operate during the pandemic, the so called essential industries. As well as for travel to work areas, which are based on commuting patterns, and there are roughly 500 travel to work areas in Italy.

 

>> Davide Rigo: So, let me talk a bit about the IV, our IV is the local availability of fiber technology. We explored a massive public investment by the Italian government started in 2015, with the aim of ensuring 100% of coverage by fiber technology by 2020. And Italy was a country that was really lagging behind in terms of broadband connection.

So before that there was the EDSL technology, which is lower, and we're gonna talk more in next slide about that. What I think is interesting about this deployment is that, in order to minimize public spending and in order to cover the whole Italian territory in such a short period of time, the rollout was progressively done in adjacent territory.

So it was mainly driven by distance rather than economic factors. Here we are looking at the share of firms served by fiber in the future in near t on the y-axis, and the x-axis is the distance to the closest area covered already by fiber technology. And we can see that the probability of being covered decays exponentially if you are in an area which is not covered.

If you are more than six kilometers away from an area that is already covered, the probability of being served by fiber in the future is almost zero.

>> Davide Rigo: Okay, another feature is that we are measuring supply rather than demand. So rather than actually consumption by firms. So we are bypassing the usual endogenarity concern with demand-driven measure of broadband Internet.

And we are lucky because the Italian government proved to be not so efficient, and they didn't complete the project by the year 2020. Even though the fiber technology coverage is quite high, 87% of firms are covered by the fiber, though the share is quite homogeneous across firm sizes, regions, and some industries.

We can see that non-financial services firms are more covered by the fiber, consistent with the idea that they're more clustered in urban areas. And also, we can see the differences between different areas, residential, industrial, and rural. And this will come back later, because we will try to control for that.

And if you're wondering how these firms differ with one another, I have a slides on that coming up. But before, I want to discuss briefly about the relevance of the instrument. We think the fiber is really more suitable than EDSL, which was the predecessor, and the previous technologies present in Italy for remote working.

Not just because of its superior speed, but because the fiber allows download and uploads. Capabilities, this was not possible before. Also allows scalability, it allows to increase the number of devices connected without losing significantly losing in terms of performance. And this was definitely not possible with ADSL and also is more reliable as a lower latency.

So it's really hard to work remotely with the ADSL technology. And instead of fiber really, I think we tried to convince you that really facilitated a remote work task. The first stage confirmed that,

>> Davide Rigo: Here we have the first stage. We are looking at the LS relationship between the adoption of fiber and extensive intensive margin of work from home.

We can see that the F Stat is also quite high, well above the conventional ten threshold. And now finally, talking about validity, we are raising some questions here. So armed firms covered by fiber different from firms not covered by fiber in some regards they are, so they are smaller and younger, and we try to control that in our aggression.

But they also don't have different sales and they don't have different productivity level, they don't pay different salaries, and also they have same level of digital intensity.

>> Davide Rigo: Another concern that you may have is that the rollout was targeted to areas that may actually be growing more in the future, that are expected to grow more in the future.

So what we do, we exploit the timing of the broadband rollout in an event study design. And we implemented the Callaway and Sant'anna estimator in this case, where in the contour group we only consider never treated firms. And we can see that there are no pre-trends in log sales, which is our measure of firm performance.

And interestingly also the effect of the broadband on firm performance seems to be null. And this is consistent with previous studies showing that actually the average effect is null and is driven mainly by some heterogeneity across firms.

>> Davide Rigo: The other concern is that actually the fiber maybe trigger an effect on firm performance that goes beyond the remote work.

So there may be some third factors that actually may play a role here. One potential concern that we had is that firms, during the pandemic, in order to face the lockdown restriction, they started to sell online. But actually we can see that e-commerce in Italy is not such a big thing.

Here we are looking at the share of enterprises with e-commerce on the left and the share of enterprises turnover from e-commerce sales on the right. And we can see that it's always most of the time for most industries below 20%, and also the change between 2019 and 2020 is lower than 5 percentage points.

Still, we are gonna show you a robustness in which we exclude industry that are using their intensive of e-commerce as a robustness for our result. These are the main result. We can see here the effect of the IV, the two stages least square on the change on log sales.

So we have a negative effect both at the extensive but also intensive margin. And here we are controlling for travel to work area fixed effects. One other thing, in order to try to get rid of as many confounders as possible is that we interact this travel to work area with the type of area.

So we are comparing firms within travel to work area and within residential area, or within rural areas or within industrial areas and the result is still there. We also run a bunch of robustness, this is the results for e-commerce. So we are looking at a subsample of the data for industries that are less affected by e-commerce.

 

>> Davide Rigo: Still we believe that this is not enough. We are taking very seriously the plausibility of our IV. And we are working on acquiring new data that allow us to test further channel that may actually have a role in this relationship. As a last piece of evidence, I want to show you the heterogeneity in this relationship.

We are interacting this work from home measure with the number of laptops purchased by the firm before the pandemic. And also the servers that they have set up at the company level. And servers are crucial because allow centralization, allow workers to coordinate among each others. And what is interesting is that column 1 and 2 show the extensive margin of workflow adoption.

So the switch to work from home and the coefficients are not significant. And column three and four look at intensive margin, so a higher number of workers in remote work. And as we can see, investment in ICT seems to matter in managing a higher number of remote workers.

So firms that invest more in that seem to be more efficiently able to handle a larger number of workers in remote work.

>> Davide Rigo: Okay, that's all. To conclude, we found a negative effect, even though it's preliminary, on firm performance during the Covid-19 outbreak. We found that investment dice matters and this is consistent with some previous work.

And as I said before, we are still working on acquiring new data and on improving on the robustness. So we are pretty much looking forward to for your comments on this, thank you very much.

>> Nicholas Bloom: Thanks Bunch, I had a question about the data.

>> Davide Rigo: Yeah.

>> Nicholas Bloom: One of the first slides you showed said there was a peak of 1.5 million Italian employees working from home during the pandemic.

 

>> Davide Rigo: Yeah.

>> Nicholas Bloom: I just went online to try to get a number for Italian employment and that which I think includes the public sector as well. So it may not be quite the right number. But my crude calculation, that implies a peak work from home rate of only 7%.

Which one- We have a paper on that is 12%, 13%. Okay, that still seems low, I mean, in the US it was over 50% at the peak of the pandemic.

>> Davide Rigo: Yeah, that's a good point. The exercise we do is that in Europe we have the labor for service, which is the annual quarterly service of the population.

And there is one question about remote working, and there are weights, and you can aggregate that at the national industrial level. And based on the labor for survey, the share is around 14%.

>> Nicholas Bloom: In 2020 or now?

>> Davide Rigo: In 2020.

>> Nicholas Bloom: Okay, it's really super low compared to other rich countries, I think, during the peak of the pandemic, but maybe that's right.

Italy is a very peculiar country, there are definitely cultural differences and definitely- LAUGH].

>> Davide Rigo: There is a literature, On that showing that the level of digital adoption is quite low and also level of digital skills and some preconceptions as well.

>> Nicholas Bloom: Okay.

>> Davide Rigo: And especially also we are characterized by small and medium enterprises, contrary to the US.

And also this play a role because as what we've seen here, the effect is heterogeneous. We are gonna work further on this, but it seems that also from other work we have been doing, that small firms are the ones that are really not able to adopt for home and adopt it efficiently because of all certain number of reasons.

And I think this is also showing in our results, but it's definitely showing in the adoption of work from home.

>> Davide Rigo: Yeah.

>> Speaker 21: Thank you very much, very interesting. So I have two points, I have more, but I will tell you later. The first one is I struggle to see the exclusion restriction because now, okay, you look at distance, but then you don't see anything about the initial rollout of the fiber, which is probably driven by economic conditions.

So then I think this is something to more carefully think about, that's something. And then the second point I would like to make, so the second bullet point, pre-pandemic investments in IST were a crucial factor or mediating factor. To me this basically tells us something about the firm size story in the sense that larger firms are handling this better rather than smaller firms, but you bundle them together and then that's what we get here in your main results.

 

>> So maybe you should look at this by firm size.

>> Davide Rigo: Yeah, that's a great point. Lemme go to it. We have this result by firm size and what we can see is that the fact is mainly driven by small firms, which yes, more firms include small and medium, but yeah, definitely that's the direction we are going.

And this is consistent with also our conjecture that work from home requires a certain degree of complexity that small firms don't have, as well as complementary investment. The other question about disocclusion restriction is so, so far we are using the cross-sectional variation in broadband. So definitely firms that were already had the fiber before are supposedly more in cities, in urban areas and that could definitely play a role.

What maybe we can try to do in the future, and we are thinking about is to look more at the discontinuity. So try to narrow down the analysis to compare firms that are even more similar in terms of geographical location. And I think that's what we're trying to do, but maybe we can work further on that, yeah.

 

>> Speaker 22: So in the United States there was lots of anecdotal evidence that said, during the pandemic, people were spending a lot of money on home goods like furniture, they were buying tons of furniture. But furniture is an industry where work from home is not gonna be very useful.

And so you may wonder, just looking at the regression, whether that's an issue for your regression. But now you have industry-fixed effects, right?

>> Davide Rigo: Mm-hm.

>> Speaker 22: Which evidently deal with this problem, but I wonder if you're confident that my little story here is not influencing your results.

>> Davide Rigo: No, I'm sorry I didn't get the key point of the question

 

>> Speaker 22: So in other words, you had a small increase in work from home in the furniture industry cuz it's not possible. But you had a big increase in sales. And so that would tend to generate a negative correlation between work from home and sales. But it's not about work from home per se, it's an industry effect, right?

But then you've got fixed effects for industry, those are your NACE things, right?

>> Davide Rigo: Yeah, so the fixed effect is accounting for any sort of demand and supply effect of the pandemic on different industries.

>> Speaker 22: But also the nature of the industry when it comes to feasibility work from home.

 

>> Davide Rigo: Yeah, also that, yeah.

>> Speaker 22: So you're not worried about my story then?

>> Davide Rigo: No, but probably I didn't get it. Well, so maybe we can talk about more later.

>> Speaker 22: Yeah.

>> Speaker 23: Hi, David Agwol. I guess I wanted to hear a little bit more about the e-commerce angle in particular because so in the US there was a 35% increase in e-commerce and there were some firms that were well equipped to do that and others that were not.

So there were really a lot of heterogeneity and kind switched sales from some firms that were kinda brick and mortar to others that were readily online. And I guess I'm a little bit worried also because I get this figure, but I guess I'm also a little bit worried because the instrument has been used in the e-commerce literature as well as a shock for e-commerce.

And so maybe could you talk about how does the effect then kinda differ across high-intensity and low-intensity e-commerce firms that were kind of ready and able to do that?

>> Davide Rigo: Yeah, that's a great point. So we are going with, well, one direction, I already show you to you.

Another direction is that there are representative surveys run by the Italian government on ICT, and we will know through that survey whether the firm use e-commerce, the intensive margins or the percentage of sales. And what we can do is replicate the first stage by looking at the changes in sales through e-commerce and to check whether the fiber technology is actually driving that.

That's, I think, a test that should rule out the channel. But with the data available to us, I think the best we can do is to try to exclude industries as much as we can. We can maybe run different set of battery robustness analysis looking at different set of industries.

So so far, we included only the chemical industry, rubber industry, construction, transportation, manufacturing, metals. So it's not perfect, but that's the approach we are using at the moment.

>> Simon Krause: Simon Krause from the University of Munich, thank you for the great presentation. I have one question about your instrument.

So you focus on fiber availability at the firm's location, but isn't also the place of residence of the workers where they were working from home quite relevant for the success of the firm? So how would you like us to think about that?

>> Davide Rigo: Yeah, so what we do here is to try to create a different measure of exposure to fiber technology by building a buffer around the firm.

And this should also account for potentially the labor market of the firm. And so the availability Of the fiber, not just for the firm, but also for the individual, is an alternative way in order to pick up this bidirectional relationship in remote working. Obviously, the power of the fiber is the fact that it allows scalability and it allows the uploads of big data and downloads of big data, which are mostly importantly for firms rather than for this specific worker.

But still, I think it's a good point and we are trying to address it by building these buffers around each firm rather than looking just at this sensuous area as a measure of local availability of fiber.

>> Raj Choudhury: This is Raj, a few quick thoughts. So one explanation that I was thinking of was that the firms in high broadband areas have more distributed supply chains because they have suppliers from far away.

And we all know supply chains were disrupted. And one way you can just address that is to look within the industry, just pick some industries where the supply chains are super local and see if the effect survives there. So that was one thought. I was also wondering, why not look at profits, why only sales?

And my third thought was there's a phenomena which has taken off in Italy that they're calling southern working. So people are migrating to the south of Italy and working from anywhere. And I wonder if you looked at the southern Italian effects, just on their own and whether they are similar to the effects of other regions or whether it's something different.

 

>> Davide Rigo: Okay, well, great comments. Thank you. We didn't do any heterogeneity across regions. What we know is that actually the fiber technologies really spread the rollout in the south of Italy. So definitely could be and a channel to look at, because there is the potential, the infrastructure to remote work there.

So, yeah.

>> Steven J. Davis: I think one last question.

>> Callum Williams: Callum Williams from the economist. This might be a bad question, but it's back to Stephen's question about one and a half million people working from home. To what extent is it possible that the product. So, yes, it's true that Italy had a small number of people working from home relative to America, but it's also true that Italy during 2020 had very high unemployment.

However, it was hard to measure. But it did have high unemployment because there were lots of people say, who weren't gonna go into the office, but also weren't able to work from home, right? So I guess my question is, to what extent is it possible that the following phenomenon is driving your results?

That people who couldn't work from home then weren't working at all? Those people were below average productivity, therefore, mechanically dragging up the productivity of the firms that weren't working from home and therefore driving your results.

>> Davide Rigo: Yeah, I think that's part of the story. I think that some firms were really forced to adopt this digital practice, then they were not ready for it.

So actually, there wasn't unemployment. There was a lot of hiring freeze. So there were a lot of people at home probably not being productive at all. And that's and that's definitely maybe part of the story that we have in mind. So it's definitely coherent with ourselves, and it's coherent with the story that there is a lot of the originality.

And it's important to account for that.

>> Steven J. Davis: Okay, thanks very much. That was a great first session. I wanna thank the presenters in the audience for being succinct, and the remarks were on time. We will resume promptly at 4PM.

Part 2:

>> Speaker 1: So the next session's gonna be a little different. It's titled Brilliant But Brief. These are student presentations. Each presenter will have ten minutes, then we'll have five minutes for Q and A, open discussion. Since time is compressed, when you do ask your questions or make your comments, please keep them brief.

We're gonna try to stay online. First presentation is hiring the ideal remote worker by Emma Williams Baron. Emma, it's all yours.

>> Emma Williams Baron: Thank you, this is collaborative work with Claire Daviss and Erin Macke. The three of us are graduate students in the department of sociology, just five minutes down the street here at Stanford.

 

>> Emma Williams Baron: So to quickly set the stage, the 20th century saw rapid progress towards gender equality in many domains, but progress has slowed or stalled in recent decades. In the world of work, gender bias and hiring is particularly persistent, and mothers in particular face many penalties in their workplace outcomes.

Scholars have noted we may need a serious shock to the system in order to restart meaningful progress to equity. Everyone in this audience probably recognizes that the pandemic has been an incredible shock in many regards, and our focus here is on remote work, which has been rapidly adopted and continues to remain quite prevalent.

So this raises the question for us. How does remote work impact employers gender bias towards workers? We know that in the hiring process, employers seek signals of workers competence and commitment to work, and the idea of the ideal worker is someone who's strongly committed and devoted to work and available to work.

But this conflicts with cultural perceptions of mothers, who are held to strong descriptive and prescriptive stereotypes that they be devoted to caregiving. We argue in this project that ideal worker norms are deeply tethered to specific job features, and that changing those features, such as remote work, might impact gender bias in hiring.

We know from prior literature about several job features that have been linked to employers gendered preferences, and some work has found that motherhood penalties in hiring are smaller when jobs emphasize flexibility. We also know that remote work is assumed to be more flexible. So we developed two hypotheses for women.

First, that they will be seen as less hireable than childless women when applying for in person jobs. So this is replicating prior research, but similarly hireable for remote or hybrid jobs compared to childless women. And we expect these patterns to likely differ for men, since research has documented fatherhood premiums in in person work, and also fathers face flexibility stigmas when they demonstrate commitments to caregiving.

So we hypothesize that fathers will be seen as similarly hireable to childless men when applying for in-person jobs, but they'll be seen as less hireable in remote or hybrid jobs. To evaluate these hypotheses, we conducted an online survey experiment with 1,500 people with hiring experience and at least a bachelor's degree in the United States.

To approximate the universe of people making hiring decisions for remotable white collar jobs. Each survey participant evaluated two job candidates for an open position. The position was described as either remote, hybrid or in-person, and participants saw either two women or two men, one of whom had signals of parental status and one of whom did not.

In the interest of time, I'm gonna focus mainly on two main hiring outcomes, which are, which candidate was chosen for hire by the participant. And which candidate the participant thought that the company would actually hire regardless of who they would hire themselves. And finally, about our design. We use these interactive digital resumes to display the job candidates information, which allowed us to collect digital trace data.

Digital trace data are records of activity and behavior as people interact with a digital system. So this can include clicks, mouse movements, time spent on the stimulus, etc. And we're developing a methods paper on incorporating digital trace data into online survey experiments, which I'm happy to talk about more in the break.

So here's our experimental stimuli for the job description, which signals job type in two places, and we developed the tasks and skill sections using O*NET data. And then here's what our interactive digital resumes look like. We signal the gender of the job candidates in the gendered silhouette, the name, and by using gendered pronouns throughout the survey.

And we signal parental status following prior literature in two ways. First, by describing a conversation with spouse and two kids versus spouse and pet, or volunteering on the PTA, or volunteering in a neighborhood association.

>> Emma Williams Baron: So now let's get to some results. I'll start by showing you the predicted probability of being chosen for hire by the participant across remote and hybrid and in-person conditions.

So if we didn't find any bias in these hiring choices, if parental status didn't matter for being chosen for hire. We would see that the bars for both childless women and mothers would be equal at the halfway mark, 0.5, indicating a 50% chance of hire for each of them.

But instead, we find a motherhood penalty emerging in the in person condition, such that childless women were picked by participants for hire 58% of the time, meaning mothers were picked 41 for 42% of the time. And then, very remarkably here we find no results for motherhood penalty in the hybrid and remote jobs.

But we might ask, is this really about moms or is this just parental status so let's move the moms or the women up and bring in men. Here, unlike for women, we find no parental status penalty or premium for men in the in person, remote or hybrid conditions.

And now let's turn to our second main outcome measure, which is which candidate the participants thought that the company would hire, regardless of which they thought to be hired. This is an interesting measure because it's measuring their third order beliefs, which are linked to actual bias that we see in the world, and it also helps us overcome social desirability bias.

So here we see again a substantial motherhood penalty for in person work, with childless women being preferred over mothers 66% of the time. And now we also see gaps emerge for hybrid and remote work, though they are noticeably smaller, with childless women preferred 60% of the time in hybrid and 62% of the time in remote jobs.

And then I'll move those folks up and we'll bring in the men. So interestingly, here we find again, no fatherhood premium or penalty for in person or hybrid conditions, but one actually emerges in the remote condition. So here participants thought the company would hire a father 57% of the time.

To help get at why these penalties might be emerging, we analyzed several measures of perceptions of the candidates. And the most notable one that I want to highlight for this presentation is perceptions of availability. So participants rated mothers or childless women as more available than mothers across all job types, in-person, hybrid, and remote.

But that gap is largest for in-person physicians.

>> Emma Williams Baron: And then we also analyze digital trace data collected through the resumes. Here I'll focus on time spent evaluating the resume. And I want to highlight that participants spent more time reading women's resumes in the remote condition, and the least amount of time in the in-person condition.

And then this pattern reverses for men. We argue that the time spent on the resume can get at how important parental status was for making the choice. So if participants think parental status is a useful signal for making the choice between candidates. Then it becomes an easier choice to make and it's quicker to make the decision because they have more information that they can use to make that choice.

Whereas if parental status is less relevant to the question of who we should hire. Then participants have less information to go on making that decision between two otherwise very similar candidates, so they need to spend more time evaluating the materials.

>> Emma Williams Baron: So to sum up my key findings, we document motherhood penalties in in person jobs, but they shrink or disappear in remote and hybrid jobs depending on the measure.

And I wanna again, highlight that null results here are particularly striking, given that prior literature finds very stubborn motherhood penalties across many different types of jobs. And basically, whatever you throw at it, motherhood penalties are often very, very stubborn. And we find that moms are seen as less available for work across job types, with the biggest gap for in person jobs compared to childless women.

And they are rated as also needing more surveillance in in-person jobs, but not remote or hybrid. Whereas for men we find reverses of these patterns. We find a fatherhood penalty emerging in remote jobs on one hiring measure, but no other differences in hiring outcomes. And we find that fathers are of seen as less available in remote and hybrid jobs, but not in in-person jobs compared to childless men.

And then finally, we find participants spend less time on women's resumes when they are applying to remote compared in person.

>> Emma Williams Baron: I said that backwards, they spend less time in the in-person positions compared to remote positions, and the opposite is true for men. So to sum up our main contributions, theoretically we argue that changing job features, in this case the in person, remote or hybrid size of the job, can have really important impacts for who qualifies as an ideal worker.

We think moms are less penalized in remote jobs in these data because employers and decision makers might think that work and family can be more easily combined that way. And empirically, we contribute a focus on how remote work affects employers gender bias towards job applicants, this sort of demand side focus.

Whereas most prior literature does reverse looking at supply side features about how remote options affect workers preferences by gender and parental status. And then methodologically, we demonstrate how online survey experiments and sociology can incorporate digital trace data, which offer some really exciting new ways to understand experimental results.

And we think there's enormous untapped potential there. I will end here, and thank you very much.

>> Speaker 3: Any heterogeneous effects based on whether the amount of majors, gender evaluators, employment status, whether they were in a remote job or a hybrid job or an in person job?

>> Emma Williams Baron: Okay, great, and so for any online listeners who may have missed the first part, you're asking about heterogeneous effects by gender unemployment status of the evaluator.

And we did look at this, and we actually found that women were harsher on mothers than men were, which is kind of hard to piece out. So we're still thinking about that one. And then employment status of in person remote. We also looked at that, and I don't remember anything exciting appearing, so I think we didn't see anything, otherwise I would remember that.

But I can look back and talk to you in the break about if there's anything there.

>> Raj: Question, did you ask these hiring decision makers to describe in their own words the explanation of the hiring decision they made?

>> Emma Williams Baron: We actually did, yeah, so we have open ended text data.

 

>> Raj: You're gonna tell us what you found?

>> Emma Williams Baron: And we're still analyzing coding, there's a lot of data to code through. So we have a research assistant who's been going through those data, and I have nothing to report at this time. Speaking just from having read through it myself, we find that people are very open with dissing moms sort of across conditions.

They can come up with lots of reasons why moms are not as committed to work or they're gonna be distracted. But in the remote condition, they seem to be talking more about positive things of motherhood, that, you're very flexible, and you can juggle lots of responsibilities. Which was pretty interesting to see, and we saw across the board lots of positive comments about dads.

He's settled down, he's responsible. He must be really committed to work because he has to provide for his family, things like that, but no statistical significance I can report to you at this point.

>> Speaker 5: This is related to Raj's question, can you see if the evaluators are themselves parents?

 

>> Emma Williams Baron: We did collect those data, and again, I don't remember anything interesting popping up when we analyzed that. So I'm going to say for now that we didn't find differences, but I will look back at our robust checks and report back to you.

>> Speaker 6: And this probably you don't have in your data, but it would be interesting to know if it's partly about feelings, about persistence, whether they think the mother will, take the job in person, but then quit quickly.

I don't know if that's something you can tease out of the data that you have.

>> Emma Williams Baron: We do have a measure of commitment, which maybe gets a little bit at that, but not quite exactly what you're asking for. Yeah, so we don't have a perfect measure of that.

 

>> Speaker 7: Were these supposed to be for nine to five jobs or sort of more totalizing jobs that might absorb your whole life, say, research?

>> Emma Williams Baron: So the job description didn't describe the hours expected of work. But it's a marketing manager position, and you might expect that's kind of a normal nine to five, I think most people would think that.

But obviously, these types of roles often swell to absorb your whole life if you let them. So it's sort of up into the eyes of the evaluator on that one.

>> Speaker 8: So I thought it was incredibly interesting, I was actually gonna say, one question is, when did you do it?

I wondered about if you had funding, just to repeat again. I'm guessing it was the US and it was maybe 21 or something 2020.

>> Emma Williams Baron: It was 2022, so May and June 2022 and running again would be a great idea. So if anyone out there wants to share something.

 

>> Speaker 9: Maybe someone listening, but also it'd be interesting to see in another country, because the US is I would have thought of these gender norms. I think it's not top, but it's pretty high. If you could find another country that was why not. I mean, it was a fantastic study, I'm surprised you got such good, you have one and a half thousand, which is, you can just, yeah, it's great.

 

>> Emma Williams Baron: Thank you so much.

>> Steve: Okay, we're gonna get us started off on the right track, thank you very much.

>> Steve: Okay, next we have Hira Farooqi, and I hope I don't have mispronounce your name, you can tell me. Who will talk about telework availability, women's labor market outcomes, and fertility choices.

 

>> Hira Farooqi: Hi, everyone, I'm Hira, I'm going to be talking about my paper on, as Steve said, telework availability, women's labor market outcomes, and fertility choices. So even though in recent decades there has been a substantial amount of progress made in terms of reducing the disparity between men and women's labor market outcomes, disparities, as we all know, still exist.

And in particular, current empirical literature agrees that the bulk of the remaining gap in men and women's labor market outcomes can now be explained in terms of the differential way in which parenthood affects men and women. This is even more concerning for the case of the United States, because United States lags behind other developed economies in terms of policies that can support families with young children.

And in addition to that, there are sizable socioeconomic costs associated with motherhood penalty as well. Well, in terms of lost economic output as well as its role in contributing to women's socioeconomic insecurity. And so one policy solution towards addressing this situation is the role of workplace flexibility. And that's what my paper is about.

In particular, in this paper I ask to what extent can greater access to remote work jobs improve women's labour market outcomes around the time that they become mothers? To address this research question, I formulate and estimate a structural dynamic lifecycle model of women's employment type and fertility choices.

I use estimated parameters of my model to conduct counterfactual analysis that looks at the impact of different policies that affect the availability of remote work jobs. My model estimates suggest that mothers of children under five years of age, so in particular, women with preschool age children rather prefer working in remote jobs compared to on site jobs.

However, because women's decisions, especially in terms of employment and fertility, are jointly related, the exact design of the policy instrument matters and I'll discuss this more in a little bit. So I estimate the sample that I use for this paper comes from National Longitudinal Survey of Youth 1997.

For those who don't know, NLSY is a rich panel data set on a cohort of men and women born in the US in the 1980s. It includes very detailed information on their employment histories, their marital histories, and their relationship histories, and their fertility histories. The sample that I use for this paper comprises of white women who are observed between the ages of 23 and 39 years of age during the years of 1997 and 2018.

My final sample includes observations on about 2200 women with 27,000 person year observations. Unfortunately, NLSY 97 does not allow me to distinguish whether a job that a woman is working in can be performed from home or not. However, I do observe the standardized occupational codes of jobs reported in NLSY.

I combine this information with information that is available in occupational requirements survey, which is an employer level survey conducted by Bureau of Labor Statistics. Which in addition to other employer characteristics, also asks whether an employer provides the ability to work from home to their employers. So I combine these two pieces of information together, and I'm actually able to assign each woman in NLSY 97 with whether she works in an onsite job or a remote work job.

I'll quickly describe the descriptive statistics. Specifically, when we come to the labor market statistics, we see that in this sample about 75% of women are employed. 57% or the larger majority, as we would expect, are employed in jobs that require them to be present on site, and about 19% of them are employed in jobs that require remote work.

When we look at earning differentials between on site and remote jobs, we see that remote jobs perform substantially better in terms of wages, both in terms of annual earnings and also when we look at hourly wages. So these are definitely better jobs. I'll briefly give a very high level description of the model.

So the model tracks women's decisions from the time that they leave school, which I assume to be around the age of 23, all the way till they retire at the age of 65. Each year during this period, a woman decides whether to work in an on site job, a remote job, or remain unemployed.

And if she's in the fertile period of her life cycle, she also decides whether to have a baby or nothing. Women's utility depends on consumption, so they enjoy consuming higher levels of consumption. Their preferences depend on having children and also working in the labor market. When women are making their decisions about which type of job to work in, they are aware that jobs in onsite and remote work different in three main ways.

So firstly, wages, as we saw in the descriptive statistics, are different in these two sectors. There are search frictions in the labor market. So women know that the probability with which they receive a job offer from each of these two sectors is also different and depends on their characteristics and their labor market statistics.

Women also know that the utility cost of working in the labor market with children and without children is also different depending on which sector they choose. Women's choices are constrained by a simple budget constraint in the model. On the income side, this budget constraint includes their earnings if they work, depending on which sector they work in.

And if they're married, it also includes their husband's earnings. On the cost side, I assume that if women work in the labour market, they have to incur a monetary cost of child care. And so overall, every period, women act optimally by choosing the alternative that maximizes the expected present value of their remaining lifetime utility.

I estimate the model in two steps. Firstly, I estimate some of the objects non structurally by using NLS 597 data. This includes information on a woman's spouse's earnings, marital transition probabilities. And child care costs are estimated using the survey of income and program participation. I also calibrate the discount factor to be about 0.95, which is standard in the literature on life cycle labor supply.

And then I estimate the remaining parameters of the model, which include women's parameters related to wage equations, wage offers, as well as utility parameters using simulated method of moments. Now just a snapshot of results. Overall, mothers dislike working in the labor market. And that's sort of in line with what we observe, that women's labor market participation takes a hit around the time that they have children.

According to my model estimates, working in the labor market costs mothers about $3,500 to $4,000 in terms of utility. I also find that mothers of young children dislike working in onsite jobs significantly more compared to working in remote jobs. Mothers of young children inside jobs, according to my estimates, would be willing to pay about $6,000 annually for the flexibility of working in remote jobs and that is about 17% of their annual earnings.

I also find, based on parameters of the job offer probability, that college educated women are much more likely to get job offers from remote work compared to women who don't have college degrees. And I also find that there's a high penalty of job switching specifically in remote sector compared to on site sector.

Now I'm gonna comment on some of the counterfactual policy that I evaluate. So the first policy that I evaluate is this situation where we basically allow the same flexibility in remote work to also mothers who work in onsite work. And I basically stimulate the effect of this policy by equalizing the disutility parameters for mothers of young children working in onsite jobs and remote jobs.

And this is kind of similar to Australian Fair Work Act which was passed in 2009, which basically, allowed parents of young children the right to ask for different types of work arrangements from their employers. According to my estimates, as a result of this policy, overall employment rates don't really change that much.

We do see an increase in the number of mothers who work, but there is no change in the proportion of non mothers who work. On the other side, fertility increases significantly so both in terms of birth rates, which increase from 6% to 9%, and also in the decrease in the proportion of non-mothers.

So basically, what ends up happening is such a policy is basically only inducing more women who would have been employed anyway to now also start having children because this policy is targeting mothers in particular. In contrast, I next look at a policy which basically just increases the availability of remote jobs for everyone in the labor market.

And this is inspired by the observation that, according to work by Dingle and Eamonn, 37% of jobs in the US can technically be performed from home. However, we only see about 20% of workers reporting any incidents working from home and this is for data from before COVID by the way.

So what would happen if you basically just increased the average probability with which women in my model would receive job offers from the remote work sector? And as a result of this policy, I see that we have widespread gains for women overall in the labor market. Overall, the labor force participation increases by about 7%.

At the same time, fertility rates remain stable. Increased employment, in this case, we see an increased employment in remote jobs compared to on site jobs. And we see increases in employment both for women with children and without children. So just to wrap up in this paper, I study the role of locational flexibility on motherhood penalty.

I find that locational flexibility is greatly valued by mothers of young children and my policy simulation suggests that an increased supply of remote work jobs can lead to widespread gains in women's employment, conditional on the exact structure of the policy that we're using. And that's it, thank you so much.

 

>> Speaker 12: Great, I don't see any hands up, so I will start two quick things. Be really as just a check on your model, you should reestimate it for men, okay? Obviously, you can decide how you want to deal with fertility, whether it's fertility of their spouse or just ignore the fertility entirely, but you sort of expect to get smaller effects for men.

Second, be nice to take your model and tell us how much welfare has increased for women as a result of the increase in remote work according to your model.

>> Hira Farooqi: Both are great points, and actually I've never heard anyone else suggest that I should do this for men, which I think is super interesting, I'm definitely going to look at that.

I am in the process of doing those calculations that would allow me to look at, get an idea of like how lifetime utility changes.

>> Speaker 13: Hi, thank you, I have a question about the interpretation of the willingness to pay estimate that you have. Whether we should interpret that purely as a willingness to pay for the flexibility, or whether it also reflects something that we've seen in the first presentation the kind of hiring discrimination that women face for in person jobs.

So I might take a lower paying remote job if the probability of getting a job is higher.

>> Hira Farooqi: Thank you, that's a great question. I think in my case, because this is a partial equilibrium model and it's only looking at the individual's decision problem. I wouldn't go as far to say anything about how, whether the individual is also thinking about how the employer is reacting to these choices.

 

>> Speaker 14: I think such a fascinating paper and the one that preceded it. This is a sort of non academic lay question about what the trend is here, because the paradox is very clear, isn't it? That there are penalties and premiums that women are experiencing as a result of post pandemic trend, but the structure of work remains, if you like, masculine or based on an old premise that men would go into a fixed place of work for a fixed period of time.

So I guess I want to ask you, are you optimistic that the actual structures of work might change as a result of the kind of data you're producing?

>> Hira Farooqi: Thank you, that's such a lovely question. I love laymande interpretations of all this econ talk that we do here.

I am optimistic, because I think if we think about this as a series of steps, the first step is ensuring that the attachment to labor market remains like that is the first thing that's preserved around the time that they have children. And then maybe the next step could be thinking more carefully about the exact structures and what that means.

From my perspective, I think I'm optimistic because at least this particular dimension of flexibility means that they get to stay associated with the labor market even around the time that they have children.

>> Speaker 14: Post pandemic.

>> Hira Farooqi: Right.

>> Speaker 14: There's a different impetus, isn't there?

>> Hira Farooqi: Right.

>> Speaker 14: Do you feel that?

 

>> Hira Farooqi: So I'm not sure I understand this because so one obviously, like this research agenda is inspired and heavily with the work that professor Golden has done. But the exact measure of flexibility that she looks at is a little different than remote work. And I can only speak to the one that I'm focusing on here.

 

>> Speaker 15: Hira, thank you very much, super interesting. So telework availability can affect the number of children, that's for sure. But also it can affect the quality of children, right? In the sense that, if you have access to time you survey, whether you could basically look into whether those who can telework spend more time with their kids and then help with their homework and so on.

That colliding would be a very nice addition to your paper.

>> Hira Farooqi: That sounds amazing, and that's probably the next paper I'll write, I've been looking for follow up questions to this, and I would love to talk more about this later on.

>> Steve: All right, thanks so much, okay, next up for the plate, Hoyoung Yoo I hope I said good, I didn't reason really well that time.

The welfare consequences, of incoming remote workers on local residents.

>> Hoyoung Yoo: Okay, so thank you for including this paper in this exciting conference, my name is Hoyoung Yoo candidate at the University of Wisconsin Madison. Today I am very delighted to present my work, on the welfare consequences of incoming remote workers on local residents.

So as people in this room all know and are very excited, the number of remote workers markedly increased in recent years in the United States. These remote workers, have the geographic mobility and can move around and can potentially affect local residents if they move to new destinations. When we think about the relationship between the local residents and remote workers, there is an interesting economic aspect arises.

Which is that high skilled, high income remote workers will consume goods in the local economy, for example, going to restaurants, bars, coffee shops. But, they will not compete with local residents for local jobs because they have jobs in hand when they move in. So this is an interesting question to understand how these kind of, remote workers having these aspects will affect the local residents.

And simultaneously, a number of cities in the United States have implemented so called, remote worker relocation programs that subsidized remote workers moving in. So in this paper, I ask, whether the welfare effects of incoming remote workers on local residents? Even though the policymakers have some intention to implement these programs to benefit local residents, but we don't know much about how these incoming women affect local residents.

Furthermore, I asked the counterfactual question, what if these women workers are subsidized by local residents through tax? So this is what this paper about. So let me give you the overview. So this paper begins with collecting a census of details of all the remote worker relocation programs in the United States, adding to some Rogers work.

And then the empirical setting of this paper is Tulsa remote, which is the first and the largest program in the United States. So given these two interesting institutional effects, I wanted to better understand the program. So visited Tulsa in person, and I acquired the program data, and I learned that the first cohort of women workers actually all kind of concentrated in downtown area.

So by using the spatial distribution of remote workers and I conduct the event study analysis by having the local employment by industry sector as the outcome variable. Especially, I classify industry sector into two sectors, one local service sector and two, all the other remaining tradable sector. The motivation of this classification comes from whether the remote workers affect local residents directly or not, which again comes from non-tradability of goods.

Then I find the labor reallocation between these two industry sectors. Then I turn to a structural model to quantify the welfare analysis incorporating the equilibrium effects, and run a counterfactual experiment. So it would be great, we can kind of do all the counterfactual experiments in the real world, but we cannot do that for many reasons.

So setting up the model allows me to run some counterfactual experiments that I am curious about to answer in the model. So this map shows all the remote worker relocation programs in the United States. And as long as remote workers can verify the full time employment, they will get into the program and get some incentive packages, mostly cash grants.

And funding usually comes from local government and sometimes even non-profit organization. And Tulsa remote case, the funding coming from actually the non-profit organization, billionaires donation. So again, I use the remote workers downtown concentration as empirical strategy variation, and I find this is the event study analysis results. I indeed find in downtown area, the population increases and they show high income, but on the local resident side, the number of local service jobs increase, however, the number of warehouse jobs declines.

 

>> Hoyoung Yoo: Then I turn to a structure model, and there are all mathematical equations behind, but not to bore you, I wanted to deliver the model intuition by using the graphs. So let's think about five markets without the program before remote workers coming in. So, as I mentioned earlier, local service goods market the equilibrium is determined within the local economy.

They are all consumed where they are produced. However, on the bottom graph, the tradable goods market, this economy is small enough to take the price of tradable goods as given. And this is set by the outside of the economy. So that's why we see the horizontal line of the price.

Then for service labor market, trade or labor market, the equilibrium objects are determined within the city and also the land market. So now let's go through how these remote workers change the local economy equilibrium. First of all, they will consume some local goods, go to restaurants and they need to eat something, so the demand curve shift to the right.

Then now acknowledging the potential positive profits, more local service firms want to enter and they need more labor to produce goods, so labor demand curve also shift to the right. Then this increased employment in the local service sector comes from the tradable sector. The reason is because previously tradable sector working workers are aware of more relatively lucrative jobs in local service sector, so they switch their industry sector and go to the local service sector.

Interestingly, the remaining trade of sector workers, though also experienced the wage gain, because now the labor force in the trade of a sector becomes scarce. Also the price of land goes up too. But tradable goods demand also goes up, but the price doesn't change. So now let me show you the estimation results.

So I used indirect inference. The key parameter of the model was elasticity of labor supply substitution of workers, which is identified by the demand shock driven by this influx of remote workers. And I crucially used the event study estimate the local service sector employment increase to discipline this elasticity.

So from the estimation results, summing up, the local residents commonly experienced the price increase of local service goods and land, but there is the variety gain. For example, there were only like Italian food, American food, but now remote workers all sharing the fixed cost of local service farms to enter.

So now there can be Japanese food, Thai food, which can benefit local residents too. On the other side, the local residents had heterogeneously experienced the effects of remote workers because of the different wages by industry sector and they have different consumption share and the land ownership. So all in all, there is the increase in nominal income on average, but this was stronger for low skilled workers due to their higher preferences attached to local service sector.

However, on the welfare side, this was stronger for high skilled workers, because high skilled workers have greater benefit from the variety gain who are more closely aligned with remote workers taste preference. So the last exercise I conducted, as I hinted earlier, was about the counterfactual experiment. What if remote workers are subsidized by local residents because the Tulsa remote case was the billionaires donation exogenous funding.

So this can be meaningful informative experiment to other cities which are considering adopting this program. So one more government budget balance constraint kicks in. So now the tax is collected from the taxpayers and that is distributed to remote workers. And the key feature is that the number of remote workers here is the function of the subsidy amount.

And I showed this is still beneficial up to the certain threshold, but the benefit size becomes smaller. Economic efficiency gain here is the consumption externality that can be shared with more population. So let me conclude, so this paper sheds light on a new place based policy, remote worker relocation programs.

And the program can benefit local residents on average, mainly through the higher wages and the variety gain. Generally, policy evaluations of remote worker relocation programs should consider, equilibrium effects such as labor reallocation, distributional effects, skill heterogeneity between high-skilled workers. Low-skilled workers, local economic condition, the elasticity of housing supply and the industry composition, as well as some public finance perspective if they are subsidizing remote workers by tax.

Lastly, I want to comment that this paper does not address the welfare in the other cities where remote workers live. But I would say this opens a lot of new, more exciting research questions to be answered, thank you so much.

>> Andrei Parhamenko: Andrei Parhamenko, very nice paper. I know you didn't have a lot of time to talk about the model, but I'm curious what happens in the housing markets because these incoming migrants, they may increase rents, but they also increase house prices.

So renters would lose, homeowners would benefit. And you probably have enough data about the housing, unless you are doing it already, you probably have enough data to do some kind of back of the envelope calculation.

>> Hoyoung Yoo: So I do included housing market in the model. Sorry, I didn't show the event study analysis about the housing prices, but indeed actually, I didn't find that much statistically significant effect on housing price increase.

And that reason was because the housing supply listing in Tulsa is very high, 3.35, which was relatively from Sizes paper, 2010. So because of that I disciplined how much price increase by using this housing supply elasticity, borrowing from the literature, but indeed the model calculate the welfare analysis effects from the landowner side and the renter side as well.

Yeah, sorry that I didn't show, yeah, too much, but I can talk more later, too.

>> Jesse Matheson: Jesse Matheson, University of Sheffield, so first of all, very cool paper. I just wanted to ask, can you say anything about what would be an optimal number of migrants into a city that we would see?

Cuz eventually, I would think that the strain on land rents is going to become such that you just get nothing but negative welfare effects by adding more remote workers. And also, I mean, thinking about what happens to the welfare in the cities that leave, I think that's a really interesting direction to go with this.

 

>> Hoyoung Yoo: Yeah, that's a very good question. Hard to answer within one sentence because there is a lot of depending kind of factors that we need to consider. But in Tulsa case, what I can show was this hump shape graph. At the peak, this was the optimal number of women workers, but there was indeed some group who are hurt by the program.

But this is only say about the aggregated impact of the program.

>> Speaker 19: Hi, super quick question, beautiful paper. I was curious if you had any predictive models of what might happen if and when remote workers might leave these places. I know that we saw a shift from tradable jobs, manufacturing jobs, to service jobs, and there's the concern that remote workers might not always be there to be on the receiving end.

So what happens to the economies when they leave?

>> Hoyoung Yoo: Yeah, that's also a great point. So from the empirical finding, the Tulsa remote program administrators say 87 percentage of remote workers stay even after the one year of the committed period. So it can be kind of empirical question, in the long run, what will happen?

What if the retention rate becomes so low? But yeah, that's a good point.

>> David Agwell: Hi, David Agwell, thanks, it's a really interesting paper. So in the model, you have a counterfactual where you're thinking about the remote workers being subsidized, but it's also possible that the reverse could be true.

That remote workers could actually be subsidizing residents, which would be the case if taxes were, say, based in the state of employment and location of employment. So it might be interesting to think through, are there any other outcomes, different outcomes that would arise in that case, and then that maybe raises the issue of the job is mobile, but the person might not be.

 

>> Hoyoung Yoo: I see, yes, sorry, this was my bad because I included the taxpayer remote workers as well. So incoming remote workers and local residents now pay together to subsidize remote workers. And another aspect of remote workers is that even though they are, let's say, being paid by $1, but what they can bring is $1 plus their income.

So this can be a little bit different from just focusing on local residents, too. We can talk more later.

>> Speaker 21: Yep, I really like this, by the way, very, very nice work.

>> Hoyoung Yoo: Thank you.

>> Speaker 21: The question I had of some of these different jurisdiction relocation programs that you talk about.

Some of them have different success rates or stickiness rates in terms of how long people actually how many come and how many stay, and seems to be tied to local acceptance. And that seems to be tied to, this is kind of a question here. How the local community funds and then helps integrate in and help assimilate in outsiders who are coming in?

Versus some communities where they're, you're just here in an Airbnb, and then we don't want to talk to you, you're creating traffic, and there's an us versus them split. Do you see any of that matching up with what you're doing here?

>> Hoyoung Yoo: Yeah, that's an interesting question. So in Tulsa, remote case, they try a lot of effort to kind of make these remote workers be more engaged with local community.

By holding monthly dinner event, and distributed this $10,000 cash grants over one course of the year, not just as the one time pay thing. Yeah, they are paying a lot of effort to that, but it's still open question whether how the outcome will be, yeah.

>> Steve: Okay, thank you so much.

 

>> Hoyoung Yoo: Thank you.

>> Steve: Last presentation of the day, who is the digital nomad by Alma Friedman.

>> Alma Andino Frydman: Thank you so much, hi everyone. Very impressed by your stamina today. It's been a long day, but a lot of really amazing work has been presented. My name is Alma Andino Frydman and I'm studying economics and sociology here at Stanford.

So today I wanted to share some data from my SIEPR paper on digital nomadism. I'm gonna give a really high level overview and I'm very happy to talk more after the session. So to start, you might be wondering, oops, sorry, okay, so to start, you might be wondering what a digital nomad actually is.

So the US Chamber of Commerce states that digital nomads are remote workers, including self employed individuals, freelancers and employees, and they travel domestically or overseas, and the Internet keeps them connected to jobs, coworkers and clients. The term was first coined by Makimoto and Manners in 1997, and while they vary in age, profession and duration in different travel destinations, I define them as location independent workers who rely on ICT's to work remotely while they travel.

They make residential decisions based on interests and passions rather than proximity to a fixed workplace and with no fixed itinerary or plans to return home. They are a very compelling case study into how people make choices about their lives independent from the demands of a geographically constrained job.

So as my own undergraduate journey was interrupted by the pandemic, I wanted to understand how remote arrangements would change my own future and wanted to look at digital nomadism. As there has been very little research into this lifestyle within academia, despite the pandemic massively accelerating this work arrangement.

So I obtained funding to travel to seven different sites in Mexico and interview 50 subjects about their professional choices, their work patterns, their identities, and the advantages and sacrifices of this work arrangement. I was most fascinated by how instability and flux of a travel lifestyle impacted their identity, their productivity, and the role of work in their lives.

So, to start off, there has been a drastic increase in the number of individuals that identified as digital nomads. As well as the availability of jobs to sustain this lifestyle and the number of countries offering visas and paths to citizenship to remote workers. So 66 countries now have visa programs, up from 54 when I first submitted my paper last year for review.

So this is growing quite quickly. Of course, the migration of high earning individuals working from their laptops to poorer countries has sparked debate around currency arbitrage and gentrification. And while the number of nomads themselves isn't that overwhelming, their purchasing power is quickly changing the world and academic literature on the subject is quite limited.

So from my interviews I did ethnographic research. Subjects tended to be young, male, and highly educated. They worked jobs in software engineering, in marketing, in design data, and earned an average annual income of about $93,000 a year, compared to the average salary of $20,000 a year in Mexico.

The majority were from the US and Canada, Argentina, Mexico and Colombia, which made sense with time zone restrictions. So throughout the study itself, I was surveying individuals who were rotating between digital nomad hotspots. Which are places with the infrastructure to sustain remote work such as coworking offices and their corresponding social circles.

And I was particularly curious as to how constant change impacted their professional identities and the quality of their work, as well as their adherence to different workplace norms and their self concept as professionals. Now of course, all of this data is self reported, but it was still striking to see how many subjects reported increases to their self assessed productivity and motivation to work.

70% of subjects reported some increase to their productivity and 88% reported they were more motivated to work. They attributed this mostly to lower levels of stress, a desire to prove themselves to a potentially skeptical employer. And a motivation to efficiently complete their work so that they could go and take advantage of leisure time in whatever travel destination they were in.

Furthermore, the self accountable nature of their work necessitated them to be extremely disciplined and focused in completing their work. Many cited the advantages of shared cowork spaces as commitment devices to stay on task, and they were generally used for silent and individual work rather than the collaborative nature of a traditional office setting.

And in each new cowork that they worked from, digital nomads established some sort of impermanent office dynamics that lacked the strict workplace roles associated with traditional offices. Which I found particularly interesting as many companies such as Meta or Chevron are actively selling office real estate. And so the traditional office is undergoing this transformation, its need and its function is being reevaluated.

My second central finding was around the choices, or rather the sacrifices that nomads made in order to sustain their location independence. A critical yet unsurprising conclusion was how subjects were overwhelmingly logistically and relationally detached. So as you can see, the vast majority were single, childless, and over half had sold their belongings to go and pursue this lifestyle.

It was fascinating to see how many had even ended relationships in order to travel full time. It heavily signaled their personal commitment to this lifestyle and the concrete decisions that they had made in order to work in this arrangement. But they didn't just make personal decisions to sustain this lifestyle, they also made really striking professional ones.

When I asked them how much of a raise they would need in order to return to a hybrid arrangement that would geographically restrict them, 70% of subjects said that they would just not consider it, no matter the raise. And for this majority, location independence was worth more than any raise, which presents a different view to how economists have traditionally modeled what motivates a worker's professional pursuits.

So the ability to work and travel is central to their identities. And thus workers sacrifice promotions and pay raises and were even willing to search for new jobs to maintain their nomadic lifestyle. So their identity was extremely centered on location independence. And while their work was important, it was not the central thing that defined them.

They overwhelmingly defined their identities according to location, independence, adventure and learning, and were strikingly detached from traditional identity definers such as political affiliation or nationality or religion. What defined them was constantly in flux, and their personal identities developed often alongside their social and their professional networks. So this brings me to my last central point, which is the subject of my most recent paper, which is that digital nomads are personally, socially and professionally defined by constant flux.

So this is the figure of my core finding. Essentially, when digital nomads travel, they change their social networks as well as their worker and professional identities. When their personal identities change as they pursue new interests in new places. Their social circles change again, which caused them to gain new professional skills or grow entrepreneurial ambitions and their professional networks.

So you can move clockwise here, starting from worker identity, and see how the prioritization of the remote status allows for personal development in new places. And then in these new places, they develop their sort of socio personal self identification, which brings them to a new situation, a physical situation.

Changing their social circles, again, which inspire new professional networks and ambitions. So I argue that this transformation occurs at a much faster pace than you would find in a traditional work arrangement or even a work from home arrangement, and it impacts the quality and the pace of their work.

Now my conclusions support, my overall findings support the conclusion that they maintain their lifestyle above many other personal and professional considerations, and they make active decisions to support this objective. They detach themselves from belongings, from relationships, and inherited ideologies to travel freely. Digital nomads value their location independence not as an extended vacation, but rather as a key element of their identity.

So there is obviously much more work to be done. And I've been so energized hearing all of your empirical studies presented today. There, as of right now, has been no empirical study conducted on digital nomads, which is something I'm hoping to change. And I am very open to any feedback, any possible research questions and directions coming from this initial ethnographic study.

So thank you all so much for your patience and your time. And yeah, I'm very grateful to Stanford for the support and the funding that I've received. And to my advisor, Nick Bloom, for all of the help throughout the last year, thank you guys so much.

>> Jesse Matheson: Hi, Jesse Matheson, that was really interesting.

Do these nomads, do they see themselves as this being something they're going to do for a fixed period of time, or is this a lifelong decision that they see themselves as making?

>> Alma Andino Frydman: It's a really good question, I think the lifetime value of someone like a digital nomad is still very undefined, because before the pandemic, this was not a very popular lifestyle.

From the majority of the interviews I've conducted, people were, like, planning to do this basically until they found a place to settle and have kids. That was the main end to this lifestyle. But there were many nomads that I spoke to that were well into their forties and fifties.

That saw themselves doing this throughout their life until they no longer physically could.

>> Callum Williams: Callum Williams just two very quick points of clarification, I guess, and forgive me if I've misunderstood. So the first one is, on the self reported productivity what's the counterfactual? And the second is, the 70% figure on the wage increase, does that account for cost of living differences?

Cuz I guess people tend to go to places that are cheaper to live. So once you control for that, what's the sort of true figure?

>> Alma Andino Frydman: Yeah, absolutely, so I think that that is something that I'm still working out on with the data. I think on both of those points, because this was all points that the data that I coded out from ethnographic interviews.

So I was not able to necessarily control for all of the factors specifically, on the second point, that was a central motivator as to why people wouldn't return home, because most people were either American or Canadian, where the cost of living is extravagantly high. And people were saying that in living in Mexico and earning 75% of what they would have earned in San Francisco, they had a quality of life that was beyond anything that they could have afforded before.

 

>> Speaker 24: Thanks, really interesting, I like the ethnographic vein. And because this is such a new phenomenon, at least at scale, you can't get a whole life cycle history. But there are other groups that share some similarities with this and three come to mind, and it'd be interesting to take the same ethnographic approach to them.

One is, military officers in the US military, in the army. Basically, you have to be prepared to move your entire family for two years, and there's people who've been doing that for decades. So if you go and interview colonels, major generals and so on, you can get a life cycle perspective on this.

Second, are people who want to be upwardly mobile in the managerial ranks of large international hotels. Same thing, you're moving sometimes across continents, but often across large distances with a family in many cases, frequently. The third are the diplomatic services, at least in many countries. So it's not exactly the same group, but it shares some of the similarities of you're tied to a place, but only for a particular period of time, then you gotta move on.

You have less control over where you move on. So anyway, I think it'd be interesting to compliment what you've done by looking at those groups as well.

>> Alma Andino Frydman: I really like that, I think I'll definitely consider that in my literature review for my next paper. But I think essential differentiator between all of those groups is that digital nomads don't have to be tied anywhere, that's the whole point, right?

So they don't want to have anything restricting them. Often that includes any kind of financial or relational commitment, which was very evident in my ethnographic findings, but thank you for that.

>> Speaker 19: I guess this probably doesn't go totally with the digital nomad idea, but I'd be curious whether firms are also arbitraging this cost of living difference.

Anecdotally, one of my friends now is a software engineer. He used to work in the Bay area, now works in Mexico at the same firm. So I don't know if it'd be possible to get data on that, but it'd be interesting to see if this is changing firms' locational decisions.

 

>> Alma Andino Frydman: Yeah, absolutely.

>> Raj: So thanks for your presentation, this is Raj, of course, this is a topic I deeply, deeply care about. So one quick thought is there is also a class of workers who are experiencing digital nomadism for short periods of time. So they would go for the summer, and some companies are using that as a way to structure flexibility.

And the reason I bring this up is, and maybe those workers have different motivations and incentives compared to the permanent nomads. But given what you said in the beginning, the 65 countries which have now introduced these visas, now, no country tracks whether you're in the country for the entire year or for two months.

And if this takes off, this would be a huge shock to immigration policies.

>> Alma Andino Frydman: Absolutely.

>> Raj: Because what prevents me now to go to Canada for two months on a digital nomad visa, work there, and then move to a different. So something to think about in terms of the duration of their stay as well.

 

>> Alma Andino Frydman: Absolutely, I'm really interested also in how work, travel, like, combined, even short term digital nomadism is going to completely transform the tourism industry. You can already see infrastructural changes in how hotels are organizing their business centers and the different amenities popping up. For example, in Mexico City, where a local bodega has been turned into a co working office specifically for this group of people.

And one other note on the digital nomad visas. I was speaking at a conference in Bulgaria of all digital nomads, where one of my collaborators asked, how many of you who are all digital nomads have a digital nomad visa? And only one person out of like, 300 raised their hand.

So clearly, these visa programs aren't actually being utilized because they don't want to be tied to anything at all. That includes a tax code, that includes some sort of immigration status. So what's really challenging about conducting an empirical study on this, and I'm very open to suggestions, is how do you even track how many digital nomads are in a country if the country itself doesn't keep track of it?

A firm doesn't want to admit that their workers are overseas. And it's very difficult to survey thousands of remote workers to self admit that they're working potentially illegally in another country.

>> David Agwell: David Agwell, so one thing that's usually linked to a place of residence is government services. And obviously, if they're not in a place for a certain number of days, then they might not be able to get access to those potentially health insurance, health care, unemployment insurance, education, right?

And so maybe the visas are covering this, but even in the US at the state and local level, this would be an issue. Are they just giving up on this? Are they somehow falsely declaring a place that's their primary domicile to get access to those? Or, what are they doing to get those services if and when they might need them?

 

>> Alma Andino Frydman: That's a fantastic question and actually something that I have been working on for an internship this summer also, so really would be so happy to talk with you more afterwards. The short answer is that they are getting private global health insurance and travel insurance in terms of planning for permanent disability or retirement.

This is just not something that's in their, like, foresight at all. They tend to be very in the moment and so that, yeah, that's just not something that they're really planning for in terms of how nations are dealing with it. There are specific residency requirements in Europe where you have to be in a country for 163 days in order to access those social safety net benefits.

And some people are actually even skirting the requirements so that they don't have to pay high enough taxes in one country. But then they are kind of falling through the cracks of these social systems, so happy to talk more about that afterwards.

>> Steve: Okay.

>> Alma Andino Frydman: Thank you guys so much.

 

>> Steve: Clearly your next CIPHER grant should be to pay for you to go to the top ten resort destinations in the world so you can search out all those digital nomads who aren't actually registering in the visa system. Thank you, that was great, thanks a lot, thanks a lot to everybody.

Part 3:

>> Speaker 1: Welcome back to the second day of the remote work conference. We are delighted to have you. We're going to continue with the same format as we had before, of presentations that are about 20 minutes long and then questions and answers after that. We are delighted to start off our first session, which is on gender, with has the rise of work from home reduced the motherhood penalty in the labor market, with one of our organizers, Emma Harrington.

>> Emma Harrington: Okay, great. We're so excited to have this on the program. I'm Emma Harrington. I'm at the University of Virginia, and this is joint work with Matt Kahn, who's at USC. So we start with the observation that when women become mothers, they often leave the labor force. This is one of the biggest drivers of differences in labor force participation by gender.

Women and men without children are pretty equally likely to be working, but mothers are less likely to be working than fathers. And this is a particularly important driver of gender gaps in highly paid careers. Traditionally, when people have thought about what would make an occupation family friendly, they thought about flexibility over when to work and to some extent, stability over when to work, that people could work during normal business hours.

We're asking whether flexibility over where people could work could also reduce motherhood penalties in the labor market, especially in highly paid careers that might have more trouble offering people stability and an assurance that they could really only work during normal business hours. And this is particularly relevant when we think about changes in technology that have made remote work particularly more feasible in certain computer oriented occupations.

So here, memory has gotten cheaper during this period. And so now we can lug around these small PCs that are as powerful as mainframe computers used to be. And then also, Internet's gotten a lot faster. And so you'd think that that would also make working at home much more of a good substitute for working in the office.

And so in this paper, we're going to be asking whether work from home reduces motherhood penalties in the labor market, and we're going to be using changes in technology before the pandemic that led to faster increases in work from home for people with certain college degrees relative to those with other college degrees.

We're gonna find that changes in work from home and college degrees in the decade before the pandemic were strongly predictive of changes in mothers' employment relative to that of other women. And so given that we see that it seems that remote work is particularly valuable for mothers, this raises the question of why there weren't more mothers working from home before the pandemic.

We developed this signaling theory of mothers' choices to work from home, where we argue that before the pandemic, choosing to work from home may have signaled low labor market attachment and reduced firms investment in women. But that may have changed by this post Covid equilibrium that may make that less of a signal.

And so we see some suggestive evidence of faster employment growth for mothers with small children after the pandemic, which is consistent with that theory. Although we think that there's gonna be more time for the sort of dust to settle for other things with the pandemic. So I won't talk quite as much about that.

So we think of this paper as contributing to a couple strands of related literature. So, one is the debate about the effects of remote work on mothers. And there, Claudia Goldin and others have argued theoretically that the rise of remote work should be a really big boon to mothers.

But during the COVID period itself, there were really unequal burdens of childcare responsibilities. And so that may make it hard to see the impact of remote work on mothers' attachment to the labor market using the COVID experiment. And so we're gonna try to study the changes in remote work in a more stable childcare period.

We're also contributing to this existing puzzle of why remote work wasn't more prevalent before the pandemic. Despite the fact that it's reasonably productive in lots of settings and workers seem to have a high willingness to pay for it. We make the argument that those who valued it the most may have also had the most to lose from asking for remote work if it signaled lower attachment to the labor market.

Okay, so from here, I'll talk about the data, then go through the empirics. So we're primarily gonna be using census data, which gives us information on whether people are primarily working from home. And so this is a bit of a course measure. It's telling us whether they're working from home most of the time, as opposed to more people use it sort of more sporadically.

Importantly, it's also gonna give us college degree information. And that's really useful for us when thinking about employment, because it gives us information on people's skills and how suitable they are to remote work, even when they're not working. Whereas opposed to, say something like occupation, where we only see it when people are employed.

We're also gonna see labor market outcomes and standard demographics. And we're gonna be particularly focused on mothers, defined as women with children, where the eldest child is under 15. But we can also focus on women with smaller children. And the results are pretty similar. We're also gonna supplement this with American time use survey data, which is gonna be important in giving us a more continuous measure of remote work.

We see a time diary on a random day, and so we can see people who maybe are using remote work more sporadically as opposed to every day. And we can also use information on the intensive margin of remote work. There's lots of people who do some remote work in the evenings, even if they spend most of their day in the office.

And finally, we're going to use the current population survey to try to get at what's happened in this post Covid world, because that goes up to 2023, whereas the census that's publicly available at the micro level only goes up to 2021 currently. Finally, we're gonna use a little bit of information on occupations from O*Net.

And particularly we're gonna use information on occupations that are computer intensive in two senses. One, that you're working with computers to write programs and do some other sort of technical things. And then we're also gonna use a measure of using computers to communicate and do less technical work.

Okay, so now I'm getting into the technological changes in work from home before the pandemic. So here we're looking at the x axis of the year. The y axis is a share of people who are working from home most of the time from the ACS. And this is the quantile of computer importance within the occupation.

And so we're seeing, consistent with the story at the beginning of the talk, that we're seeing rises in this primarily working from home in these computer intensive occupations in the blue here. And that was much stronger than these less intensive computer occupations. And this particularly what mattered for mothers.

So, strikingly, you can see that mothers in these computer intensive occupations, even in 2019, almost 15% of them were working from home most of the time. And this is even more dramatic if you look in the american time use survey at working from home some of the time.

We also see a relatively similar pattern when we look at a very different set of occupations, of ones that are intensive in communication and media. So using computers in very different ways. So that suggests that there's going to be some college degrees during this period where there's rises in remote work and others that are relatively less impacted.

So that's how we're gonna try to characterize women's skill sets and how big a change they saw in remote work. And so first, to make things concrete, we can think of a couple of examples. So people with education degrees were relatively less likely to work from home, and that didn't also change very much.

Over this period, they were kind of tied to the schoolhouse or what have you. Marketing, by contrast, is a pretty computer intensive occupation. You have lots of people who are using computers for both communication and, to some extent, analysis. And there you see a much larger increase in work from home of about 0.3 percentage points each year.

And so we're gonna characterize every college degree according to this measure. And so that's gonna give us this variation in how big a change there was in remote work across all college degrees. And so these are just the biggest college degrees. We do it for all of them.

But it sort of intuitively makes sense. A lot of things that have a big in person component had small changes in work from home before the pandemic, whereas things that were computer intensive in one way or another are the ones that are seeing the biggest increase. And a thing that's worth noting here is here to the left, we are categorizing those degrees sort of mechanically as those that have the biggest versus the smallest changes in work from home.

And so that's allowing us to see that there's meaningful changes in primarily working from home, but there's much bigger changes in sort of these measures of hybrid work. On any given day, the likelihood that you were working from home increased a lot in the occupations associated with these increasingly remotable degrees, relative to those where they weren't so suitable to remote work.

And even this we should think of as being an understatement of how big this trend was, because this is just capturing the people who happen to be working from home on the day that they were interviewed. It's going to be even bigger when you think about the people who have the option value of working from home.

This is an even more intensive measure of remote work, where here we're looking at the share of people who are working from home during work hours, and we categorize it separately by people working from home during business hours versus outside of business hours. And so here you're seeing that, in general, before the pandemic, people were doing a lot of their late night work, for example, at home.

And that was particularly true in these remotable degrees. And you might think that that would be particularly important for mothers who might find it particularly costly to be in the office during the evening. So now I can turn to how this relates to changes in mothers' employment. And so here we can go back to our example of education versus marketing and look at the changes in mothers' employment during this period.

So mothers were increasingly likely to be employed in both types of degrees, but the rate of growth was much larger for marketing. You might be worried that there's just other things happening with marketing. Employment in general could be rising. So we contrast what we see here with what we see for women without children, and there we see more parallel changes, anything a little bit stronger for education.

And so what we're most focused on is this difference, the change in mothers' employment relative to that of other women. And so for that we see much bigger impacts for marketing than for education. And so then we can do that for every single degree, the same method and correlate the changes in mothers' employment versus non-mothers' with the changes in work from home.

And that produces this picture of looking at for every different college degree, what was the change in work from home during this decade, before the pandemic? And then what was the change in mother's employment relative to that of other women? And we see this really strong positive correlation where things like marketing have big changes in work from home, big changes in mothers' employment relative to non-mothers.

And then educations on sort of the other end of the spectrum. Obviously, I'm the one who chose those examples, so you might think that they were selected. Here that sort of illustrates that that is part of the more general trend, that degrees like marketing that solve big changes in work from home are also the ones that saw changes in mothers' employment.

And those that have more of an in-person component, so much less of a change during this period. This just puts the same thing in table form. We're seeing that a one percentage point change in work-from-home is associated with even more than a one percentage point change in mothers' employment relative to non-mothers.

In some ways, that magnitude, on its face, seems way too big. I think it's much more reasonable when you think of this as being a very small measure of a much bigger change in locational flexibility. This is people who are primarily working from home. There's similar trends for people who have some degree of hybrid work that is probably at least twice as big, probably more like, I would guess more like four times as big as this.

So that makes this magnitude seem a little bit more plausible. We can also see that it's not being driven by other correlated changes in the nature of work within these degrees. You might be worried about lots of other things. So you might be worried that the change in mothers' attachment to the labor market is actually what's driving the change in work-from-home.

That as mothers become more attached. They advocate for work from home. And so we have reverse causality. You can define the change in work from home using just the sample of men and you get really similar results. We can also do a triple difference with fathers where we look at the change in mothers' employment vis-à-vis women without children and compare it to fathers' employment vis-à-vis men without children.

And the results for that are almost identical. And so this suggests that it is something about mothers specifically that they benefit maybe more from this vocational flexibility. We can also look at other measures of women's labor market outcomes. So we can look at hours, for example. And there we're seeing increases in mothers' hours relative to that for women without children in these more remotable or increasingly remotable degrees.

That largely goes away when we condition on being employed. So most of that's coming from the extensive margin. But this is suggesting that it's not just that we're pushing women from not being employed into working really short part time jobs. It seems that the whole distribution is moving up because of locational flexibility.

We're also seeing increases in log income in these degrees, even when we condition unemployment. Suggesting maybe being able to work those evenings is helping women advance in their career, even over and above being employed. Okay, so now we're gonna think about, we've shown you some evidence here that remote work seems to be particularly valuable for mothers and their ability to work.

Why weren't more of them working from home before the pandemic? Why was it just 15% in these computer-oriented occupations and not 80%? And so to think about that question, we're gonna develop this signaling theory of the choice to work from home, which I'm just gonna sketch out briefly.

In that model, firms are going to be able to invest in workers skills. That's gonna pay off only if workers persist in the labor market. Mothers are gonna have private information about their attachment to the labor market and their likelihood then of paying off for the firm to invest in them.

Work from home can mitigate the cost of working, and that's going to be particularly valuable for those women who find it particularly costly to work and who are likely to drop out of. There's also gonna be a fixed cost initially of adopting remote work. Because work from home is more valuable for mothers mother With high costs of working, asking to work from home is going to signal low attachment.

We're going to argue that before the pandemic, we may have been in this separating equilibrium, where, because it sent a signal of low attachment to ask to work from home, firms were leery of investing in mothers who made that request. And so, as a result, high attachment mothers are going to work on site, because they want the firm to invest in that, whereas low attachment mothers are going to work from home or drop out.

We argue that because the pandemic overcame these fixed costs of adopting remote work, it's plausible that now we're more likely to be in a pooling equilibrium, where all mothers choose to work from home, at least to some extent, and so it sends much less of a signal about your attachment to the workforce.

Firms then invest in mothers regardless of their locational choice. And that opens up the possibility that women can use remote work much more fully than they did before the pandemic. We think that the dust is still sort of settling to be able to test that prediction. But we do see some suggestive evidence when we look in the CPS and we look at employment of college-educated women and differentiate them by the age of their eldest kid.

Here in black, there's women without children, and then the other lines are women with children, and particularly the lower lines are women with young children. And we do see some evidence that there's faster employment growth for these women with young children compared to other women in the wake of the pandemic, when they've been able to use this locational flexibility in a much bigger way.

Okay, so to conclude, we just talked about how technological changes led to changes in remote work for some degrees versus others before the pandemic. And mothers were more likely to use that locational flexibility, and it seems to be strongly related to changes in their employment patterns. And so then we argued that it may not have been used more before the pandemic, in part because we were in this equilibrium in which asking for remote work signaled lower attachment.

Thanks.

>> Speaker 3: Hi. A great paper. So I have a suggestion, which is, I think you can compute the rates of return on different majors, first under the pre pandemic assumption of typical working periods, and then compared to the post pandemic differences in working periods. Right? So majors are going to have a higher payoff now that you'll spend more time in the workforce.

So there's a paper by Keith Chen and Judith Chabalier, who argues that women actually have low rates of return in many professions, like nursing, in particular. Exactly. Because women are spending so much time out of the workforce now that that's changed, I suspect many majors actually have much higher rates of return for women, and I also wonder how much that's leading to endogenous changes in enrollment.

So at NYU Stern, we actually have now more female applications for business school than men. So there's a huge change in that dimension as well.

>> Emma Harrington: Yeah, I think that's a great point. And I mean, I think something that we didn't talk about at all, but I think it's exciting.

It's like the most, the ones that have seen the biggest changes in remote work are really remunerative occupations and also ones where women are really underrepresented. So I do think it's going to potentially change the distribution of women across majors. That's a great point.

>> Speaker 4: It'd be interesting to see whether the effects you're measuring are stronger in metro areas with high traffic congestion.

So in other words, work from home is more attractive in LA than it is in Kansas City, where congestion is less.

>> Emma Harrington: Yeah, that's a great point. I can tell that you're also a Matt Kahn co author. Yeah, he's been pushing me in that direction. So I think that's a great, it's definitely on the to-do list.

 

>> Speaker 5: So my recollection from working with the work from home data from the ACS is that the rates are much higher for the self employed. And in terms of the models, something that you may want to think about is that the choice of working from home or not often means, at least before the pandemic, it often meant choosing to be self employed versus working for a firm, which also involves more labor market risk.

 

>> Emma Harrington: Yeah, that's a great point. So in general, a sort of interesting thing is that pre pandemic mothers were much more likely to, who had small children were much more likely to be self employed than women who were similar but didn't have children. And that was like all because that was the way that you could do remote work.

These trends, I think, somewhat surprisingly to me, women aren't more likely to become self-employed in these increasingly remotable degrees. Maybe because, you know, that is traditionally the way to be remote, but as it becomes a better substitute, then maybe they are able to get remote jobs even within firms.

That's a great point.

>> Speaker 6: Super cool, you know, two thoughts. So first, I was thinking whether some of these trends might change even more because some of the professions where you had low share of working from home earlier, like nursing and education, it's become more acceptable now with telemedicine to do nursing from afar and secondary education, you can do classes on Zoom.

So I'd be really curious how this plays out because we have changed many professions. The other thing I was thinking was, in terms of another outcome, you could look at career continuity because at least anecdotally, lots of women in dual career situations are now able to not give up their jobs because the spouse has moved, because they can work remotely.

 

>> Emma Harrington: Yeah, those are great points.

>> Speaker 7: I'd add on to what he said over here about the entrepreneur piece. So you kind of answer that part. The other one, in terms of pre- and post-data, you can get looking at a specific employer and seeing where they used to have very low work-from-home rates, and then it's increased rapidly in the same organizations, often with the same people, federal government, there's a lot of open data on that we can talk about if you want.

 

>> Emma Harrington: That's a great idea. Yeah, I think that would be really cool.

>> Emma Williams Baron: Hi, this is Emma Williams Baron. Thank you for this very compelling paper. This is fantastic. And one thing I'm concerned about is potentially, if mothers are working from home much more, this could more deeply entrench unequal divisions of household labor.

And I wonder if you could speculate about how that would fit into your signaling theory here and how employers might respond to that.

>> Emma Harrington: Yeah, that's a great point. So, I mean, a striking thing in the American time use survey is that when people are working from home, it's really frequent that they have secondary childcare as one of their activities.

So I do think that that's an issue. And there's a nice paper by Abby Adams Prassell saying that a lot of times women are less productive at home because they potentially are juggling multiple things. And so I do think that that's a consideration. I mean, I think that may be part of the reason that pre pandemic, it was used in this sparing way that, like, you could work from home at these late hours when the kids were already in bed, and that was like a potentially very productive time.

So I think that that's definitely an interesting question that we'll think more about.

>> Speaker 9: I just wanted to ask about the data sets that you're drawing on and whether you think they're going to change and evolve to ask deeper questions. Because my sense is that quite a lot of the data that's been collated about women is a bit patchy.

 

>> Emma Harrington: Yeah, I think that that's a great question. I mean, I think it would be great to have more detailed measures of flexibility in these data sets. I mean, something that's been a struggle to some extent in this paper is thinking about just how many people had some degree of hybrid work before the pandemic.

I think that our sense is that that was something that came out of COVID but I think particularly for mothers, that was something that they were using beforehand. And that's something that's sometimes missed in these questions, maybe because they're not as much geared towards thinking about the specific work concerns of mothers.

 

>> Speaker 10: Emma, I don't know how feasible this is, but I'm wondering if it's possible with any of the subsets of data to look at US versus other countries. One of the things that we actually saw during the course of the pandemic was the issue of work from home with mothers was most acute in the US, followed by the UK and then continental Europe was not quite as bad because of more infrastructure, child leave, other implications that US just hasn't had.

 

>> Emma Harrington: Yeah, that's an interesting question. Yeah, I've been mainly focused on looking in the US currently, but I think that could be a very interesting direction.

>> Speaker 11: Hi Emma, great paper. Thank you. I was wondering if you're concerned about the role of childcare provision stability in these comparisons?

So before the pandemic, probably mothers were not working remotely as frequently was because childcare availability was a lot more stable compared to the period right after COVID when we saw a lot of interruptions in the childcare provisions. So just wondering what your thoughts are on that.

>> Emma Harrington: Yeah, so I mean, I think that that's part of the reason we are currently pretty cautious in how we're interpreting the COVID period.

So, I mean, if you projected what we saw in the pre-COVID trends to think about the massive increase in remote work, that would suggest that the motherhood penalty and employment should really decrease, particularly in remotable degrees. And we don't really see that. We see sort of a continuation of this trend as opposed to a huge level shift up.

And I think we're optimistic that maybe once child care normalizes, we'll be able to see a little bit more of that. But I think that is part of the picture. And I think something else we could do in more of the geographic vein is think about like do we want to maybe control for like what the child care availability is like and how that may interact with this picture.

 

>> Speaker 12: That's fantastic. One question on the quantification was I often think it'd be nice to know how much work from home increases GDP? And you could do some numbers from this actually. There's at least two people from the Fed. This thing even feeds into things like interest rate policy.

So I think you could estimate impact of this on labor force participation rates, which gives, think of a classic solar model that's just gonna push up airline. In fact, there's also skewed airlifts, like high-quality labor that's increasingly working as well. So I wouldn't be surprised if the numbers you get out of this were tens or hundreds of millions of dollars by really large numbers.

Because if you're pushing up female labor force participation by 2, 3% and it's highly educated women as well, actually I think the numbers would be absolutely enormous.

>> Emma Harrington: Yeah, that's an interesting point.

>> Speaker 14: You mentioned Abi Adams Prassl's paper where she finds that it's parents of kids under five.

And I know that in your robustness, you've looked at mothers with kids under five versus those with 15. But building on the two points from over here, it seems as though you could leverage policy changes both in paid leave that sort of New York state has implemented, but also in childcare provision to make sure that it actually is about having your kid at home.

So introducing childcare that is public and paid for that starts at age three or four rather than at age five might give you some leverage to look at the impact there.

>> Emma Harrington: Yeah, I think that's an interesting point. And thinking about how it interacts with these trends, I think could also be interesting.

I think some of my interpretation of this is partly that it is in a more stable child care period. And so it's partly this is so valuable because it allows you to not work from home most days, but be, like, ready when your kid is sick to go pick them up and not have that be a huge disruption and be able to continue working that day and be available during weekends or in the evening to be able to take calls.

So it's maybe not, the mental model I had in mind was not so much that people were doing this all the time, but doing it sort of sporadically. But I should think more if there are ways to more directly say that that is the underlying mechanism.

>> Speaker 1: And now we're going to go on to what works for her.

How work from home digital jobs affect female labor force participation by Suhani Jalota.

>> Suhani Jalota: Hi, everyone. I'm Suhani Jalota, and I'm a PhD student here at Stanford, and I am gonna be presenting my job market paper, What Works For Her. How work-from-home digital jobs affect female labor force participation.

This is joint work along with another PhD student at MIT, Lisa Ho. And as part of this paper, we offer jobs to more than 3000 women for over 60 days. And set up 35 community centers to understand what might be constraining female labour force participation in India and try to shed some light on social norms as an influence in constraining these women as well.

So married women's labor force participation rates are quite low in many emerging countries around the world. India is an outlier even to that trend, where despite its dramatic economic growth, we've seen that labor force participation rates remain even lower than 25% for women. And this tends to be even lower for urban women and for women that are in fact more educated.

So the middle-educated women seem to have the lowest labor force participation rates, despite increasing education doesn't really translate into employment. Most of these statistics are true for jobs that are located outside the house, which are most jobs today. But as this audience knows, there are many other opportunities.

But most of the jobs available to these women tend to be outside the house. And married women particularly seem to have about one-fifth the labor force participation rate as compared to unmarried women. And so, as part of this study, we really are focused on married women only. And there are many constraints as this audience is very familiar with jobs that are located outside the house.

There's just practical barriers to do with safety, travel-related concerns where women can't leave their home. Because of these concerns, there could also be just other inability to multitask along with childcare or housework concerns where paid unpaid work seems to be really high. There's also information related barriers where women might just not know what types of credible, trusted information about jobs or what jobs are available outside the home.

They could also face various social norm-related barriers where they're just not allowed by their families to step out of their house and work. There's just a lot of pressure. There are these gender norms that expect women to just stay home. And there's been various research by economists and social scientists done to explore many of these different barriers, particularly the safety, the practical and the information related barriers have been more explored because they tend to be less complex in terms of actually studying than the social norm related barriers.

In this study particularly, we're trying to understand which of these constraints seem to be more important in determining middle-educated married women's labor force participation. And for us to study this question, we need to be able to separate out these different practical information and social norm-related constraints. Particularly controlling for constraints that we can more easily control for, like the practical and information-related barriers to be able to better understand the influence of social norms in constraining women.

We also want to understand that if by reducing these constraints, these practical information constraints, some social norm constraints as well. Can we actually increase labor force participation of these women? And if so, then really by how much? And so now to study these questions, we need to find a job intervention that can help us disentangle the practical information.

Information and social norm barriers that tend to be quite intertwined with jobs that are located outside the house. And so to be able to reduce and control for some of these practical barriers, we need a job that is at home or near home that controls for many of the safety and travel related concerns.

We need a job that's part time or flexible, that allows women to choose the timing of her work. Maybe the work itself is only 2 hours a day at a time of her choice that allows her to multitask along with the unpaid housework and child care burden that she has.

The job intervention would also need to be child-friendly so that she can manage along with childcare. And also women-only, where in this context, we know that male coworkers tend to pose potentially a safety risk on women and might decrease labor force participation that way. We also want to control for some of these information related barriers that women might experience.

And so to reduce and minimize those, we should provide information about the job to women multiple times and at home, where we can provide it to them in a language and a way that is really accessible by them, and is also provided to them on their own devices that they're already quite familiar with, such as their own smartphones and in person.

We also wanna provide women with a job that itself, by its nature, is quite accessible and easily understandable by a population that's middle-educated or middle-skilled. So as we tried to find a job that met all of this criteria and would help us study our research questions, and we couldn't find any exactly fitting this bill, we developed our own digital gig work platform for women from home and from centers.

This is a women-only gig-work platform called Roni Work. You can see a snapshot of this here, where we've taken already existing tasks that are being performed in large business process outsourcing BPO companies, large call centers, and converted them into micro tasks that can be performed by women over their own smartphones through this mobile application.

And these tasks are things like image annotations, speech recordings, speech transcriptions, which are then used to train various AI models by various companies. These women can also perform these tasks from 35 local child friendly centers that we set up. And all of these centers are trying to mimic a home environment as closely possible as we can.

So they're located in women's own communities, maybe just a two to five minute walk for most women in this setting. So a lot of the safety travel, the practical barriers, and the information-related barriers are also quite minimized. While the tasks that I presented to you here are quite tailored for this particular application, there are many companies that are trying to provide gig work already to people around the world that are being used to train various AI models.

We've just taken this work that is normally not accessible by this population and converted it to a more accessible format for them. So first, I'll provide you with an overview as to just where we're going. It's a short talk, so very briefly, then talk about the experiment design and then our results.

So to be able to understand our research questions and understand which of these constraints between practical information social norm barriers might really be constraining women, we run a randomized control trial where we randomly assign women to either home based treatment or a center based treatment, both of which are minimizing or trying to minimize the practical and information related barriers that women experience, but only home controls.

For some of these additional social norm related barriers that might be expecting women to stay home. We also randomly assign women to different wage levels as we're interested in also understanding whether we can financially compensate women to be able to overcome some of these costs and take up maybe work from other locations.

So the main finding of the study is that even when we are offering women jobs from these local flexible centers in their own communities, where they literally just need to step out and walk to the center, the job take up from home is double that from these centers, and that no matter the wage.

So even when we are offering women five times higher wages, that is equivalent to a monthly household income, it does not increase supply at these centers. We also then try to understand, well, what is causing this large difference between home and centers? And we run follow up experiment where we then try to disentangle some of these effects, and we find that the actual housework burden and not actually childcare as much might be explaining some of this difference, and that some of this difference is also really explained by some results that we have that point towards social norms that might be keeping women home.

They also run a mini experiment with husbands where we offer them the exact same wage levels and the exact locations. And we find that this difference is actually quite gendered in that husbands are pretty wage elastic, but not actually sensitive to location, and that only women seem to be displaying this sort of behavior.

We also then ultimately talk about, and we don't get into these results as much today, but as relevant for this audience is that we find that the marginal productivity from centers seems to be slightly higher and there might also be other benefits to working from centers. So what do we really do in the experiment?

So we run a household survey, a baseline survey with our 3300 women and households. We then randomly assign them to either receiving a home based treatment or a center based treatment. We then cross randomize them into three different wage a low wage at about dollar 60 a month, a medium wage at $150 a month and a high wage at about $300 a month.

So the high wage is five times that of the low wage and is about the average husband income. This gives us about six treatments here. And then we have an additional control arm, which doesn't receive any of the job offers, but just receives a baseline and line survey.

We also randomly assign households to either receiving a husband's survey or not. And this is where the survey, the survey team is also providing the job details to the husbands. And we look at the take-up rates by husbands for these surveys as a proxy for husband job refusal.

And we don't get into those results here, but I just wanted to highlight that. And we also did run an endline survey and then do some follow-up experiments that I'll just talk about. So we run this field experiment in Mumbai slum resettlement communities, where the average age of the women in this sample is about 32 years.

We have about 50% of the sample in an open caste or general caste category. We have 50% of the women about who have young children below the age of eight. Almost all the women, majority of the women have children in the sample. About 50% are Hindu, and they're middle-educated.

So about 11th grade pass. And about 40% of them have their in laws living with them as well. So we also have some survey measures around whether or not women think they're allowed to work by their husbands. And explicitly, women state that 30% of them are not allowed to work in a job by their husbands.

This is how the treatment looked like for work from home, where women are just working on their own smartphones, very similar in centers. Women are working on their smartphones but are also able to bring their kids along with them to these centers as well that we set up.

And the centers are all located in the same building setups where their homes are. So the layout is pretty similar to the work environment as well. So now that we have assigned women to home and center and the different wage levels as part of this experiment, what happens to their labor supply?

Now, at baseline, all of the women were not employed. So our eligibility criteria required that women were not employed in any full-time employment or anything more than 20 hours a week. So now given that we find that reducing these Practical information constraints and maybe some of these social norm constraints, like in the home environment, we're able to increase labor force participation here at the extensive margin, at least by 56%.

And that's pretty large, given that the willingness to work in this context is actually also very high, but does not normally translate to employment, because maybe of a lot of these different barriers that have persisted. But if we can control for some of these barriers, we do actually see that the labor supply could increase.

And now when we look at the same sort of a job, exact same job, but from a center setting, which is right outside, we see that the take up is about half. So there does seem to be this large premium on work from home, even when the center is next door.

And this is when we're averaging this out for the three different wage levels. Now when we separate this out by wage, we see that generally, conditional on location, women are wage inelastic. So if even when we offer women $300 a month from home, it doesn't increase take up from centers as well.

There's a slight increase between the low and the medium wage from centers, but that's likely because there's a slightly higher reservation wage from centers, but generally wage inelastic. And that by paying women more to join centers, not necessarily going to be an effective solution at the expensive margin.

We don't see this trend. We see the opposite result, kind of for husbands, where husband's labor supply seems to be quite responsive to wages, particularly by the high wage, and that it is not generally quite responsive at all by location, at any of the wage levels. So this response seems to be pretty gendered.

So now we wanna understand, well, why is it and which of these constraints is more important for women. Given that these home-based jobs are already quite flexible, very similar to the home environment, and already child friendly, and providing childcare. Which seems to be one of the largest constraints for women's work.

So maybe there's still, even for the two hours a day that women are working at a time of their choice, maybe there's still some persisting household response responsibility. Some other childcare, housework that they're not able to control for. They could also be maybe persisting safety, travel concerns, maybe some other information barriers that we're not able to control for, or some of these social norms, the norms of domesticity that are expecting women to just stay home, even if they don't have actual burden of housework and childcare.

So through our various survey work, we find that the safety travel concerns and the information related barriers in this context, given the design of the experiment, don't really seem to be at play. So women are not necessarily rejecting jobs because of these reasons, but it could be due to housework responsibilities and social norms.

So we try to then explore these mechanisms a little bit more. First off, we find that childcare might not actually be driving a key difference between this gap between home and center, given that women from centers were already able to bring their kids to the center. And so we see that women that have young children and don't have young children don't differentially necessarily take up the job from centers.

So women with young children are not taking it up any less. And so then we try to understand the housework burden versus the social norms through this follow-up experiment. Where at N line, we then randomly reassign women into either receiving a home-based or a center-based treatment, like we had in the primary experiment.

In addition to that, we add three other sort of complications in work from home that are making them even closer to a home based environment, so making them even more costly. So we add an E2 arm, which is no multitasking. So women are continuing to work from home, but they can't multitask, along with other housework or childcare responsibilities.

How do we enforce this? We introduce, it's a digital platform, so we can introduce notifications that women have to immediately respond to, otherwise the tasks expire and they lose money. So this is our test for the actual housework burden. We then also add treatment arms e three and e four, which is where women are working from home, but they need to step out of their house to check in at a local center.

That's, again, a two to five minute walk from their house. They're just getting a code that unlocks their application, and they come home and they work from home again, but they have to be visible, so they have to be seen stepping out of the house and working. And in the ID arm, we added an additional social signal that a woman actually has to wear a work ID while walking to the center and back, making it even more obvious why she's leaving home.

So these two, we're trying to use to test the observed norms of domesticity. But there's still a lot about social norms that we're not able to capture in these experiments. And if also these arms were perfectly executed the way we designed them, then I think that the results that we would have seen might have also been a little bit more drastic.

But we do see slightly underestimate of these results, which I'll just show you. So, first off, we see similar results in this follow-up experiment, as we saw in the primary experiment for take up from centers being proportionately 66% lower than the take up from home. Now when we add the no multitasking arm, we see that this take up reduces proportionately by about 19%.

So that means that the housework burden might be explaining some of this difference, but not the largest chunk of this difference. And that this check in at the centers and the check in an ID might also be explaining some more of this difference because it proportionately decreases the take up by about 30%.

We still have some other unobserved differences that we then talk a little bit more about in the paper and attribute some more results that we find where we talk about why this might be driven by some more social norms as well. So this audience is particularly quite knowledgeable about this.

And we have not done very detailed analysis into the productivity or the intensive margin results. But I'll show you a preview into what we have a little bit here. So now that we have 56% of the women actually entering in home based work, about 27% in center based work, what happens to them now?

And because this type of job is actually piece rate performance pay, we can actually observe productivity by task, and we can do this for over two months that the women we're working on the platform. So first off, we see that the marginal productivity from centers is somewhat higher, so about 14% higher across the distribution.

And this is just to remind you, in this particular treatment, we did not provide any training to the women from home or center. There was no additional monitoring done from home as well. So potentially, and this is a combination of selection effect and a treatment effect. Through some other work that we've done, we find that this is a split by about a third, two thirds.

So a third is a selection effect, and about two-thirds seems to be the treatment effect in this case. And that generally also when we look at retention rates and completion of the work over the two months, that worker retention seems to be somewhat high. Also from centers, conditional on working.

So this is conditional on working that more women are staying and continuing to do the work from centers. The women that seem to drop out from home also seem to be women that seem to realize over time that they're not allowed to work by their husbands and then tend to drop out.

So we have some evidence to show that as well. But from centers, this sort of triaging into whether or not you're allowed to work tends to happen before. So it tends to happen at the extensive margin when choosing whether or not to work. But at home, they seem to realize this over time.

Maybe women have hidden the fact that they're working from their husbands, it's a pretty unobservable job. They can just be seen as playing on YouTube or something, but they're actually working in this job. So they find out over time and then they become less allowed. We also see that retention rates tend to flatten out at about this medium wage, $150 a month, but it tends to be quite low at low wage.

So this is where wages do really seem to matter for women, which is at the completion rate. And the market wage for such jobs is about 120 to $150 a month. This is in line with what we see for the retention. So we'd expect that about 25% of women from home would likely be sustained on such a work platform, given the market wage, and about 15% or so from centers would do the same.

So finally, when we control for some of these practical and information related barriers, we see that female labor force participation can increase, even from these center based jobs, these hubs within women's own communities, by about 25% at the extensive margin. That we can't really increase the statistic even more by more financial compensation as higher wages in this context, doesn't really seem to be an effective policy tool.

That's likely because women are constrained by additional things that are not able to be overcome by wages, maybe like social norm constraints, but we could provide women with home based jobs to overcome some of these, to increase the take up of these jobs by a lot. At the same time, you want to design these home based jobs in a way that is also a little bit more productive and is allowing women to be able to also benefit from some of these additional gains that women can otherwise have from centers, potentially with some sort of a hybrid model or stepping out and doing some sort of work from home, but then also meeting others from time to time.

We're doing further research now to look at the effects of work from home and working from centers on women's agency mental health, further employment, permanent employment outcomes as well. And we're just, we don't have the results right now to present to you on that yet, but thank you so much.

This was our paper on what works for her, and I'd be open to your questions.

>> Speaker 16: Amazing paper. Thank you so much for talking about it. I was wondering, you mentioned that there was about half of the targeted participants who were living with in-laws, and it's interesting to see, because there's evidence on both sides.

People who live with in-laws might also face stricter regulation regarding social norms. Ankriti's work talks about how women who live with their in laws have a harder time forming social networks and a harder time participating in the labor force. But on the other hand, they might also contribute to some of these domestic chores that, the burden of which falls on women.

So I was wondering if you were able to do some of the heterogeneity analysis and see how this impact varies.

>> Suhani Jalota: Yeah, exactly, that's a great point. We actually collected quite extensive data on in-laws involvement. So whether or not they were living with them, whether they were living close by, whether they were living in the same, like, city, even.

We don't find any heterogeneity by in-laws. And that seems to be because of this dual. Like, it's pushing up and down. That for some women, it's really helping with childcare, housework, burden, and for other women, it might be more constraining. So overall, the effect is not different by the in laws.

There's a separate paper, though, on the mother in law effect, showing that ultimately it seems to be a positive thing, ironically, for women to be able to actually work because the housework burden is just so large and it needs to be shared.

>> Speaker 17: This is excellent work. Thank you.

Do you see any difference based, conditional on household income or the husband's employment status? That's number one. And then also, do these women tend to have their own bank accounts as it is simply the fact of being paid, revealing then to the rest of the household that they are working?

 

>> Suhani Jalota: Okay, great questions. On the first one, by husband's occupation, almost all the husbands are actually employed in this context. And so I think it's about 85 or 90% of them are employed in full time sort of work. We're not able to necessarily see any heterogeneity by that.

With husband income, we do so generally higher. Richer households seem to have lower labor force participation rates across the board. That could be the income effect. That could be an additional signaling effect where you're trying to signal that I'm well off and so I don't want my wife to work.

We're not able to separate those two effects out. We would have really liked to, and we tried really hard trying to do that, but we often get pushback that it could just be the income effect, and it really could be. And then on the second question on could you remind me again?

It was. On the financial inclusion bit, this is all digital work. So it's digital payments as well. And only about 40% of our sample had their own bank accounts when we started. So everyone, including the control group, was offered to open a bank account if they did not have one under their own name.

And most payments went into their own bank accounts under their own name, so they could hide it if they wanted from their husbands. But if women explicitly said, no, we want it to be in our husband's account, we did not stop that either. It was on the woman, but we opened bank accounts.

So we had about more than 300 bank accounts that we had to open as part of this.

>> Speaker 18: Thanks. I love this study, partly because it confirms my prior that work from home offers tremendous opportunities to bring disenfranchised groups into the labor market. In your setting, it's women, but other settings.

But in terms of what I, the stuff about mechanisms is great, but the headline result is so large, the doubling of the potential to double the labor supply of these groups of women. There's a sense in which the mechanisms are secondary. And I want to elaborate on that point in two respects.

First, I believe it probably the domesticity norms are in fact a very powerful factor that's repressing labor supply outside the household. So I'm totally sold on that. But the accumulation of earnings over time has the potential to erode that social norm, because my guess is it's probably partly sustained by the fact that there's an extreme imbalance of financial resources within the household between the men and the woman, or the woman and her parents, if she's living with the in laws or whatever.

In some longer term sense, it's not clear that you're actually getting to the full effect of the unsocial norms and so on. But the other thing is, suppose I'm an employer, I want to train AI programs and I want to tap into this labor pool again. I don't care what the mechanism is.

If these people can do the work, especially if I can pay them piece rate that's 10% lower than what I get in alternative markets, there's a profit motive to scale this up tremendously. It's great to keep doing the experiments on the mechanisms, but I was also trying to figure out a way to cooperate with a party that has a commercial interest in scaling up what you've done, because that's what's going to really drive the social transformations.

 

>> Suhani Jalota: Completely agree on that. On the first point on social norms, there is this additional breadbinner norm, which is also what we were trying to study here, that if you give women a wage that's higher than their husbands, it might actually, in this context it was not as concerning because it was a short term job, but actually for longer term permanent jobs that could be more of a concern.

So we had women hide, especially in the high wage arm, they were hiding more from their husbands and they could do that because they said it was project based. So right now it's okay. But if this is a permanent salary that I would keep getting for a long time, it would threaten him so much that it might cause her to exit entirely or work in a different kind of lower paying job.

So there is some thing to do with this longer term social norms that we haven't been able to capture also here and on the private company bit. Absolutely. So, Ronnie, work itself actually is now working with various companies and is continuing work for women on the platform and in these communities.

And the idea is to definitely take this to the government even and where there's a national urban livelihood mission. Normally in agriculture, in rural areas, we're able to get people unemployed into the workforce, even just for 100 days of part time employment. There's nothing equivalent to that in the urban sector, in the urban areas.

And so we're trying to propose this as a solution for that. And being like, this can be great part-time work that even the government can provide and propose.

>> Speaker 19: Hi, yeah, I wondered if you asked what they spent the money on or what they intended to spend the money on.

Kind of to speak to this hidden income or to speak to kind of income pooling more widely.

>> Suhani Jalota: Right, most of them. So we have these details for the women, and most of them tend to spend it on their children. So it's primarily they're saving for their children or spending for their children, and across all the wages.

But particularly in the high wage arm, they're saving more, but again, for their children's education and their future. So it's, I think, only about 10% on anything to do with personal expenses. And when it's personal expenses, it tends to be own health bills. So something to do with their health care or sometimes even just nice clothes.

And so that's what we saw came up often for women, but otherwise, it's always children.

>> Speaker 20: This is awesome. I was just curious whether within your study it was possible to differentiate between religious groups in the US, remote work among mothers was really high before the pandemic in Utah.

That seems probably not random. So I was wondering in your study if it's possible to do something similar.

>> Suhani Jalota: Yeah, and we do see that. So for Muslim women in our sample versus non-Muslim women, that Muslim women tend to be a lot less allowed to work. So the completion rates, both from home and from center, are much lower, almost 30% lower for Muslim women.

And they seem to be learning whether or not they're allowed to work through this process. So in the follow-up experiment, they're even less likely to even start working. And this confirms some of our beliefs that generally women in this context are not even aware whether or not they'd be fully allowed because they've never had this conversation before.

So they learn as part of this experiment that whether or not they're actually allowed. And so in Enline, we see that more women are not allowed to work, but that's only because they seem to have realized this after talking.

>> Speaker 21: So how high does the household income and education have to get before these social norms weaken sufficiently to allow women outside the house?

 

>> Suhani Jalota: I wish I had an answer to that.

>> Speaker 21: What is the answer?

>> Suhani Jalota: Honestly, I mean, we've seen as incomes increase, that norms are getting worse. Higher caste women in India tend to have worse norms because you kinda wanna signal that you're doing better off. So it's not clear if it actually gets better with education or with income.

And so we don't really know if there's a tipping point or where it goes back, I mean.

>> Speaker 21: So if you were living in India and you had PhD from Stanford, you would be at home instead of-

>> Suhani Jalota: So for women with tertiary education levels, we do see that it increases slightly.

So essentially, about 56% of women that are illiterate or uneducated are in the labor force. This is about 20% for middle educated and 25% for tertiary educated. So if you have a PhD or a master's degree, so it increases slightly. But it's still quite likely that a woman with a PhD in India is maybe not employed, if she's married, because it just seems that nobody's sharing the housework burden.

And India is ranked last in the world when it comes to this share of housework burden along with husbands, because husbands really don't contribute at all.

>> Speaker 22: So two thoughts, Vanessa. Loved your study and completely echo what Steve said, these effects are just like one really, really eye opening.

So two thoughts, I was wondering if there's variation. You could exploit the variation around the age of the woman, and my prior would be the older woman will probably care less about the opinions of the husband or the in-laws. And I'm curious if you've looked at that.

>> Suhani Jalota: Yeah.

 

>> Speaker 22: But also on the productivity effect, since you're exploring that now. I was wondering if you could consider maybe in a future intervention the hybrid model, because in your center thing, they were sitting together. And I'm wondering if there are learning effects happening there that are not happening at home.

And one way to test that would be, for the home people, to have them periodically. Maybe once a week, come and give them feedback on their work and see if their productivity then closes the gap. So beyond teasing out the selection and treatment effects, if you can also induce some learning effects for the folks working from home.

 

>> Suhani Jalota: Absolutely, I think those are great ideas. By age, we do find that older women care a little bit less, so they are more likely to work, but not that much more. And yeah, absolutely, I think a hybrid is where we wanna study next as well.

>> Speaker 23: So some of the motivating facts that you began with were differences by labor force participation rate across different countries.

And it was striking how low India was relative to other countries, the same GDP per capita. And so I'm wondering whether you thought about sort of returning to some of those facts at the end. And trying to relate some of your work on mechanisms to ask whether they can explain those differences.

And sort of related to that, from what I understand, the differences in women's labor force participation across different indian states are maybe as large as the cross-country differences that you showed there. And it seems natural to ask, is that because of those norms? Or I mean, it could be that the cross-regional differences have some other reason that aren't necessarily the ones that you've explored before so far.

 

>> Suhani Jalota: Yeah, all great points. So I think there's been debate about why Saudi Arabia's higher than India when there's just so many more restrictive, similar restrictive norms there, too. Also, why is Bangladesh so much higher than India when we're neighbors and have similar norms? And I think that this is less studied and there's still a lot more puzzling things to uncover here.

And I have some thoughts, but I don't wanna, none of this is really scientifically valid or proven right now. But I think generally, on your point of just whether within country differences, we do another study in West Bengal. So this study was done in Mumbai. My co-author and I have another study in West Bengal, where we do a similar experiment.

But there we're trying to understand flexibility in terms of timings for preference for work from home and these local offices, as well as childcare facilities versus not. And we find very similar take up results and things that kind of really validates and reinforces many of these results here.

And we find similar influence of norms even in that context. So they're quite similar in terms of overall rates.

>> Speaker 1: Great, then we are done with this session, and we will rejoin at 11 o'clock on the dot.

Part 4:

>> Natalia Emanuel: Thanks so much for having me. This is joint work with Emma Harrington, who's here, as well as with Mandy Palace. And the standard disclaimer applies that the views do not represent the firm we are studying or the New York Fed or the Federal Reserve system. So, as we all know, very few people were working from home before the pandemic.

But even now, as we're coming out of the pandemic, firms tend to be fairly divided about how they are viewing the office. On the one hand, we have some folks like Mark Zuckerberg saying, actually a lot of people are more productive now that they're able to work from home.

And on the other side, you have some folks who are embracing the office. So here you have James Gorman of Morgan Stanley who said, the office is where we teach people, it's where our interns learn, and that's how we develop people. So this paper takes seriously the idea that these are not necessarily in conflict.

It's totally possible that you could have remote work boosting your productivity, particularly because we are doing this at the expense of developing people. And so we're going to ask, does sitting together increase mentorship, potentially at the cost of immediate output? We're going to test this theory in the context of a Fortune 500 company where we can look at software engineers.

Software engineers are an interesting group to look at, particularly because we can see data on mentorship in the context of code reviews. And I'll talk about what that looks like in a minute. And then we can also see their output in terms of the number of programs that are produced.

We're going to leverage the fact that at this particular company there was variation in proximity. So before the pandemic, they had two main buildings on their campus, and some teams were able to sit together and those ones would have daily stand up meetings in person. Then there were other teams that were separated across the two buildings.

And we think of these as multi building teams where a lot of them would meet online line. And because of that, they sort of proxied for what it might look like if you were working remotely anyway. And this was all due based on desk availability. So we can think of the fact that when these offices closed that we had differential loss in proximity.

And so we'll use a difference in difference design to try to understand the impact of sitting next to your coworkers. So we'll go briefly through a model that sort of disciplines how we're going to think about it and then talk a little bit more about our findings. So in our model of mentorship each worker is going to live for two periods, first as a junior, then as a senior, and they're paired together.

They either sit together or apart, and seniors have to give some feedback to their juniors and that is observable by the firm. But there's this additional depth of feedback which we're going to call mentorship, and that juniors can ask for and seniors can either choose to give or not give, and that part is very hard to observe.

So I think in the academic context, you can imagine we have to have certain number of grad students that we're mentoring, but the depth with which you're thinking about their work or how frequently you meet with them is much harder to observe. And so we're going to hope that mentorship actually makes juniors more productive as seniors.

But there's an opportunity cost both for juniors and for seniors when they're giving and receiving feedback. So this reduces their output. And because we're nothing rewarding the mentorship that we can't see, it leads to decreased pay as well. Additionally, we say that asking for mentorship is costly, so it is difficult to ask somebody to exert more time on you.

And so these juniors are going to ask for in person only, where it's potentially less imposing. And likewise, it's socially awkward to reject people when they're asking for mentorship. And so seniors are going to agree to give this mentorship, but only when asked. And so this gives us several testable predictions.

It's going to say that sitting together increases mentorship by increasing asks for it. It decreases programming output and it reduces pay. And then having sat together also increases human capital. So conditional on sitting together is going to increase pay later and increase quits for a better job. So we're going to be able to test these in our setting.

Just to speak a little bit about what this code review process looks like and what the job of a software engineer is. We are looking at 1055 software engineers at a Fortune 500 company and we can see their online mentorship. So they give peer comments to one another and we have about 175,000 of those on 30,000 pieces of code.

We can also see personnel data which will tell us about their proximity to their collaborators, their pay raises and quits. Our sample has been at the firm for about a year, or 16 months. They are pretty young, they're only about 29 years old, and they're about 81% male, which is perhaps not surprising.

Nationally, software engineers are about 75% male, and only 16% of these engineers are parents. So the code review process, because they don't necessarily want you to introduce a really big bug into the code. When you're going to change code, you will create a branch where you will then edit.

And before you deploy that code, there is a mandatory review process where a peer will go through your code, potentially flag issues. You might iterate back and forth and then it'll be merged back into the code base so it'll be deployed at that point. And giving this mentorship is an essential element of an engineer's job.

So here I have just a sample of what a job looks like when they're advertising at Microsoft, where it says two of the responsibilities are about providing high quality code review feedback, and coaching and mentoring other people. We also find that these comments are actually constructive. So we did a principal components analysis to try to generate those sort of themes in these content, and we're finding that they're about interacting with databases, they're about who owns the code.

It's about function output and code testing. So all of these are actually highly technical. We additionally had software engineers outside this firm review a random subset of the comments to try to understand are these helpful comments? And we find that the vast majority of them are in fact helpful.

That 70% are likely to cause change in the code and about 60% are explaining their reasoning. So they're very much constructive. The comments. Okay, so what do we find? Here, we had predicted that sitting together is going to mean more mentorship for people who are in one building teams.

And that after the offices close, we're going to see a bigger decrease in mentorship for one building teams. And that's exactly what we find. So in blue, I'm showing you the engineers who are on one building teams, and they're receiving about 22% more comments on their programs than the people who are in multi building teams.

And if you thought that this is just happening because people on one building teams are chattier or they actually need more assistance, then that would persist even after offices close. But when you can see the gray line is when offices were closing. Bless you. What we actually see is that a decrease going on there.

So that gap almost entirely closes, which tells us that this is something about actually sitting in close proximity to other engineers. That was just the raw thing. We also have included time varying engineers, time varying tenure, as well as program scope, because you could imagine you might get more comments if you are writing a longer or more in depth program.

And additionally, we'll include controls for an engineer fixed effect, time controls of Age and other demographics. Throughout, we're gonna be clustering by engineering team. And so you can see, even when we conclude our preferred set of controls, there is again this statistically significant decrease. I'm going to show you one table which shows you as we are including each and every one of these controls.

And in that top row, it's showing you that our difference indifference actually remains remarkably steady even as we add our additional controls. And that's notable because in the bottom row, you can see that the amount of variation that we're explaining, goes from about 2% to about 50%. In the light blue column, that's our preferred specification, and that's what future annotated estimates are going to show you.

So we also have a placebo check here, which is that if you're sitting near teammates, that should impact only feedback from other teammates and not from non-teammates. That sitting next to your teammates should not impact what happens with people who are not your teammates. And about a third of comments are from non teammates.

And that is exactly what we see. So in the blue, you can see that we have the statistically significant decrease. But in the gray, it shows that comments from non teammates have no impact whatsoever about whether or not you were sitting with your teammate. So this is robust to various different measures of feedback.

The total length of your comment. It also shows up in the speed of the feedback. It shows up in terms of each one of those principal component analyses, that you can see the same pattern for each one of those. And then it also, we see, suggestive complements in terms of other references to slack or other means of communication.

And then we find that this is robust in a number of different ways. That in the interest of time, I'm going to skip. We do find evidence that this is mentorship. So one of the things that we look at is that this is driven by feedback to junior engineers.

So on the left hand side here, I'm showing you people who have been there at the firm less than the 16 month median amount of time. And on the right hand side, it's comments to people who have been there for longer. You can see that first of all, in levels, the people who have been there a short amount of time are just getting more feedback.

They need more feedback. But that the also the difference about being on the one building team is coming from the people who are more junior. This also shows up if you're thinking about this by age with an independent effect. It's also driven then by feedback from senior engineers.

So again, you can see that the senior engineers are providing more of this commentary, and that it is particularly potent that the senior engineers who are on one-building teams are the ones who are giving this additional feedback. So, so far we've shown you that there is this piece going on and that it's from seniors to juniors.

So here we're also gonna show you one more piece about mechanisms. So we're looking in terms of externalities of having a distant teammate. So we're looking at, if I am on a multi building team, what happens to the feedback of people who are sitting together anyway? And we find that teammates who are sitting in the same building, if they have one teammate who is elsewhere, their conversation is still decreased.

So it's simply the fact that we are now having our meetings remotely, even though I'm still sitting next to you, I still see that cost. So we have about a 13% less feedback when one teammate is in another building. You can also do the same thought experiment, actually before COVID.

So there, we can see that four people who were on certain teams where they added somebody to their team and that person happened to be in a different building. So they shifted from being a one building team to a multi building team. Those are the people in orange.

You see a dramatic decrease in the amount of feedback that's going on, whereas the people who had somebody who are added to their team, but it doesn't change whether or not they were a one building team, then there's no difference in the amount of feedback happening. So both of these, it suggests that even one member who is elsewhere will negatively impact the people who are sitting together.

So the takeaways from this section are we're finding that proximity does increase online feedback for junior engineers, specifically, that's coming from senior engineers. And that proximity is increasing contact with these potential mentors. I will say that basically we are looking at online feedback. So you should think of this as a lower bound, because if you think that sitting together, it also means that it's increasing the amount that we're schmoozing and we're chatting, we're not capturing that at all.

This is purely the digital measure that we're seeing. So then on the next side, we're wondering, what is the impact on programming output? So here we had predicted that when the offices are open, there's going to be fewer programs that are written, because actually receiving and giving that mentorship takes some time, but that once offices close, there's going to be a relative increase in programs for the one building team because they are no longer going to be giving as much of that feedback.

And so that's exactly what we see, that the people who are on one multi building teams were actually producing, again about 23% more programs, and that we see this decrease over time. And so there's this increase in the people who are on one building teams when they're no longer spending as much time on the mentorship.

Again, this is driven by the people who are actually giving the feedback, the senior workers, you see that those ones in particular were the ones who were, going to be affected most potently by this. So again, this is robust in terms of various different measures of programming output.

And we think that it's unlikely to be due to shocks to specific engineering groups because it shows up again in different types. So, so far, we've basically said sitting together increases mentorship, but it decreases output. And together, for a firm, this then creates a now versus later tradeoff, that on the one hand, I could have some additional productivity now, but it might come at the cost of a less skilled workforce later on.

We do look at whether this is differential by gender. As I mentioned, 81% of our engineers are male, but we see a bigger impact on mentorship for women. So in female programmers, if they're on a one-building team, they're receiving more feedback than male programmers on one-building teams. And female engineers on multi-building teams are receiving less feedback than male engineers on multi-building teams.

So this effect is true for both types of engineers, but it's just much more potent for female engineers. And we can see that the impact on mentorship done by women. So this is the commenters, no longer people who are receiving, again, is more potent for women. And that this also shows up in the mentorship, that because women are providing more mentorship, there's also a greater output cost for women.

So female engineers are then producing fewer programs when they're on one building teams. So taken together, we can see that for female engineers, this tradeoff is even more potent. It's more essential for junior engineers training, but it's more costly for senior engineers output if they're female. We then think about what happens in terms of career outcomes.

And as I mentioned, you could imagine that if we are then giving more training to people who are on one building teams, then they are more skilled. And so what we can see here is that if the offices are open, juniors who are on one building teams might be paid less because they are spending more time on mentorship and they are producing less.

And indeed that's exactly what we see. And so here I'm showing you the dot is the point estimate on a difference-indifference. And that one, offices are open, we see that the difference Differences in mentorship do mean that people, female or, sorry, people on one building teams are paid less.

However, once offices are closed, there's no additional mentorship. So now we're just looking at the impact of having received that mentorship beforehand. And so the people who are on one building teams are now more skilled. And so you should see that they are paid more relative to the people who are on multi-building teams.

We also see this in terms of quits. So after Covid-19 came and offices closed, you can imagine that you can switch to more productive firms much more easily because you don't need to move. All of this is remote. And so those who are particularly more skilled. Right, those who have benefited from receiving this mentorship, those are the people who are going to be able to move more.

You see, overall, just an uptick in ability to leave or people quitting. But it's particularly potent for the people who then had been trained before. So for about two-thirds of the sample, we can see where they went and what job they took. And so we can use Glassdoor data to compare with their job beforehand and their job afterward.

And we can see that about 78% of the quits are to higher paying jobs. There is no differential before and after whether or not you were going to a higher paying job, but it does suggest that this massive increase is coming from people going to higher paying jobs.

So then we can think about what would the firm be doing if they were sort of cognizant about this? And so that sort of gets to the firm policy. And we can think if the firm was really cognizant of the benefits of mentorship and that they may exceed the costs, then we would imagine that when the offices are open, they would keep junior engineers and mentors on site.

But once proximity is difficult, so once offices are closed, they might hire more experienced engineers. So we'll test that as well with our firm. And so here, what I'm showing you along the x axis is engineer job level. So the most junior people are the L1s and the most senior are the L4s and above.

And you can see that the most junior people, as well as the most senior people, the people who would be mentors, those people are the ones who are in the office before Covid-19 and then additionally, the people in green, those are the managers. And you can see that most of them are also there.

And so this is suggestive that the firm is really trying to make sure that those who need mentorship and those who can give mentorship, those are the people who are really especially in the office. And this shows that similarly, when you have controls, and also if you're looking by parenthood status, as I mentioned, only 16% of our sample are parents.

But likewise, when you look at the office nationally here we're looking at data from the household pulse survey, you can see that it's the most junior people, the youngest people, and the people who are a little bit more senior who are in the office. And again, we can't really say anything about whether that's preference or firm policy, but it does seem to be showing beyond just our firm.

And then second was the question of once we are remote, who is the firm choosing to hire? So in the gray I'm showing you the percent of people who are hired by this firm. So before the office closed, they were doing most of their hiring, was at this very junior l one level, and they did hire a little bit at l two, l three, l four, but is much less frequent.

However, after the office closed, that's this. In the blue triangles, you can see that they shifted away from hiring junior people who they would have to train, and shifted toward people who were already trained. And so this is suggestive that they were aware of this sort of build versus buy tradeoff, that if it's harder then to be mentoring people, they're more likely to be hiring people that they don't need to do as much mentorship for.

So I know it's a bit of a drinking from a fire hose, but what have we shown you is that proximity increases online feedback for juniors, but mentorship has this opportunity cost, especially for seniors, and we see this decrease in programmatic output and that this trade off is actually more acute for women.

We see that this tradeoff then changes the pay path as well as influencing quits. And that firm policy suggests that the firm itself believes that proximity's benefits exceed the cost. We think that this also provides one answer to why remote work might have been rare before. COVID that if firms are aware that mentorship is extremely valuable, then the fact that workers might have been willing to pay 8% of their wages for the additional flexibility of working from home, and that you saw higher productivity that still might be outweighed by these mentorship costs.

So thank you so much. We would welcome feedback both online or in person.

>> Audience 1: Hi, thank you. Can you say something about team size that you didn't mention in particular? I was intrigued by the result that is enough that one team member leads to produce the effects you show.

If it is about coordination among team members, then we would probably see that effect not to be really relevant if just one person goes away. So can you say something about that?

>> Natalia Emanuel: Yeah, so teams are about five people. And so one person is a relatively large portion of that.

But I think the other piece that's going on there is that software engineers, it's fairly standard. Though they'll use the agile management system, which means that they will have daily meetings to say, this is what happened yesterday, this is what's gonna happen today. And these are my barriers.

But. And those happen in real time, no matter if you're remote, but once you're then separated across buildings, those are going to happen, online rather than in person. And it seems that one of the things we think about is that just the schmoozing before and after allows people to sort of say, I'm working on this thing.

Can you give me more feedback on it? We see a similar thing. I didn't show this, but in these two buildings, one of them has 74% of the engineers and the other has the remainder. And you can see that if you're in the bigger building, which potentially allows you to sort of bump into more people, you see a similar thing, that they get more feedback from non teammates.

So you could imagine just the ability to sort of talk in person, then means that you're able to get more of that online feedback.

>> Audience 2: Would the company that you are working with allow you to run programs that measure code complexity and parse for comments? So I work in an organization that does this kind of code review, and a lot of the time I find myself commenting on junior staff's work to say, hey, please describe this better.

Or, hey, you can simplify this and make it less buggy. And so the cycle complexity of the code, which you can pick up from a program, would be like a potential measure of quality.

>> Natalia Emanuel: Yeah, so, unfortunately, we don't actually have the code content, we only have the comment content.

And so we can't see if code quality is increasing or if bugs are increasing because they also don't peg bugs directly to the engineer. So I would love that, but no, I don't think that we're going to be able to get that data.

>> Audience 3: One thought is, this is actually great.

I've read the study before. It's really been insightful thinking about if there's interventions that companies can put in place to actually prompt teams that are distributed to actually do this on a more regular basis. The real challenge here for most every firm, which is you can't unscramble an egg.

Like, if you look at Microsoft's data, they've gone from 61% of teams being co located pre pandemic to 66%. Being spread across cities, not just buildings. And I worked at Slack and pre pandemic. I actually ran a project engineering team. We debated remote teams for three years running and never pulled the trigger because we'd never done it before.

 

>> Natalia Emanuel: Right.

>> Audience 3: But the difference for us was we could hire people in more locations, right? It just unlocks the opportunity. And we knew that there were trade offs to be made, but those trade offs are not going backwards because it's so hard to get engineering talent in the Bay Area, New York, London.

So I think part of it may be experimenting with things like, what can you do to drive intervention on this that gets managers to do it? May be worth thinking about there's experiments that can be run there.

>> Natalia Emanuel: Totally, yeah. I would say that this firm has tried to have a strong return to the office among their engineering group, and there seems to have been slight rebellion.

So in terms of.

>> Natalia Emanuel: Totally. Yeah, yeah. I totally agree that sort of harnessing nudges seems like a wonderful thing to be able to test. Yeah.

>> Audience 4: Two part comment. One is about being able to hire more senior people to what Brian said. Yes, if you say that we are routinely hiring people anywhere, suddenly the available pool of candidates is order of magnitude.

That's plus one to what Brian said. Two, in terms of companies. The other example, I've worked at open source companies that are doing the same code review process, and they're physically distributed. So it's not just no office building one and building two. It's like no offices and it's different time zones.

The question here is, could you take that as a comparison against what you have here and see if there's a difference because they have been doing that since long before COVID and continue today.

>> Natalia Emanuel: So I'll tell you what we've done within the company, and we can see that the teams that are not just distributed across building, but are distributed across campuses and so are potentially in these different time zones.

 

>> Audience 4: Yeah.

>> Natalia Emanuel: We see almost identical trends. And so it does seem as though it's not necessarily the time zone or the asynchronous piece that's going on, but I think the sort of different context of actually there is no office, and so we're all doing this, and there might be a different culture around that seems potentially interesting.

Yeah.

>> Audience 4: Yeah, it sort of enforces the sort of normalizes out for the casual conversation. You're talking about the bigger building and the smaller building. Other more casual conversations in the bigger building. This kind of zeros that out because everyone's in a different city.

>> Natalia Emanuel: Right. Right. But but the question is, do they get less good mentorship because they're all in different cities?

 

>> Audience 4: Right? You're right. Exactly.

>> Audience 5: Super interesting one. Like Brian, I've read this a lot. One thing that was new that was, I think, talks to something that's a very topical debate. I just want to ask you about was when you add an extra team member, whether in the same building or the different building, because really linked to this whole debate with hybrid as to whether you get people in on the same day and you have like eight people and one person doesn't come in.

 

>> Natalia Emanuel: Right.

>> Audience 5: Does it for seven onto Zoom. Is that what's going on there? It's really interesting that somebody new added that's in a different location damages all the current team members. I wonder if you knew-

>> Natalia Emanuel: That's exactly what we think is going on, that just having one person in a different building means that now we're on Zoom rather than meeting in person.

And so all the benefits of that schmoozing seem to go out the window. So you could imagine if there is that coordination, then you still might wanna meet in person and just have that one person zooming in rather than everyone in their separate siloed offices zooming into that meeting.

 

>> Audience 5: So I think we have time for one more very short question.

>> Audience 6: So I've noticed something during COVID and afterwards, and it's related to meeting with graduate students. I think that meeting with graduate students is better online when they can share their screen than when they're shoving their computer across the desk or showing you a piece of paper.

So what do you think about that?

>> Natalia Emanuel: Given I am not an academic institution or mentoring any graduate students, I can't give any personal insight, but I do think that there is an established norm in computer science that they will look at the same screen regardless of whether it's screen sharing or we're looking on.

Yeah, and even sometimes coding will happen with somebody's watching as you code in real time. That's like a normal way to train somebody. And so it's not obvious to me that that's at play here, but potentially in the academic context, it makes sense. Thank you so much.

>> Host: So next we have Anthony Dierks, who's gonna tell us about research output before and during the pandemic.

 

>> Anthony Dierks: Thank you. And thank you for inviting me. Title is research output before and during the pandemic. My name is Anthony Dierks. I'm a DC area economist, and I'd let you know that these views here are my own. They do not reflect the views of my employer. Okay, so this study is about outcomes associated with fully remote work, not I read.

So what is the consensus view on fully remote work? Well, look no further than the very well done paper by Barrero, Bloom and Davis where they note that fully remote work was associated with about 10% lower productivity. And one potential explanation is that workers kinda maybe have this lack of motivation and self control.

But the studies that are frequently cited to make this kind of claim is that kinda focus on call center and data entry workers. And there's really no example in which fully remote work was more productive. And so all the positive outcomes are kind of attributed to hybrid. So in terms of the big picture, should the negative productivity findings based on call center and data entry workers be generalized and extended to the entire knowledge workforce?

Instead, what are the productivity outcomes of fully remote work for individuals? I tend to be highly motivated with advanced degrees. So my goal today is to convince Barrero, Bloom and Davis to tweak the narrative about fully.

>> Anthony Dierks: Fair enough. Okay. So ask yourself, our economists, what is the likely effect of fully remote work on productivity?

And what I'll tell you is that the number one response I get is that it depends. It depends on the individual. The effect on productivity of fully remote work is heterogeneous. So for some folks the effect may be negative, or for others, the effects may be positive. So recall center data entry workers, maybe it's negative.

Or highly motivated individuals with advanced degrees, well, let's stay tuned. One other thing I'd like to point out that there's a literature about fully remote work that has documented, objective, not self assessed, positive outcomes for decades. So these are very difficult to find. I'll be honest with you with that.

But there's a paper by Geisler from 1978 for Blue Cross Blue Shield, South Carolina. There was written up by Phelps from Mountain Bell in Denver, 1980. Newman looked at Travelers Insurance company back in 1989. Dubrin looked at the data entry workers, the MPD group in New York, Boy et al.

Looked at call center workers at Kentucky American Water Company. And then Collins looked at some insurance techs at Lloyd's insurance in the UK. So these are all examples in which they provided fully remote work, but had positive outcomes for productivity. Okay. So again, the goal is to convince you that the effects of fully remote work on proactivity are potentially heterogeneous.

There's two other studies that speak to this potential heterogeneity. There's Dutcher, 2012, which provides experimental evidence that remote work for simple and repetitive tasks was associated about 10% lower productivity. Remote work for tasks requiring critical thinking and creativity was associated with 20% higher productivity. And then there's a longitudinal study by Monteiro, Straume and Valente, they look at Portuguese firms.

They find that remote work had negative productivity associated with the firm's primary links point, low-skilled workers. In contrast, remote work had significantly positive productivity effects for firms that undertook research development activities. So again, do the findings associated with fully remote call center and data entry workers were broadly applied to all knowledge workers?

What about researchers or academics? Well, this question's already kind of been received some considerable interest during the pandemic. There's a paper in the Journal of Finance where they documented a self reported decline in research productivity during the pandemic. In contrast to that, there's a review of financial studies paper that finds that among the top 50 schools, there was a 35% increase in productivity as measured by SSRN paper postings.

But what they also noted is that the largest gains accrued to the top 10 schools. There's been a subsequent paper written, looked the top 1000 schools, and they found an overall decline in productivity with increased inequality. And they said that maybe the extra time spent on teaching had an important negative effect.

But most importantly, out of all this is that all these studies excluded federal reserve system economists, and they did not face these teaching costs. Okay, so during the pandemic, the Federal Reserve System adopted some fully remote policies, much like everyone else. So the question is, were these policies associated with any effect on research output of its economists?

Did males/females, respond differently? Were certain age demographics affected more than others? What was the impact on collaboration? Was there any effect on inequality among economists? Can a general equilibrium model rationalize any of the findings. And then what would be the macroeconomic effects? So, we know that's important because increases in productivity tend to reduce inflation.

All right, okay, so here's what we did, we examined working paper output from the 12 Federal Reserve System regional banks, as well as a richer measure of output for Board authors. The working papers are a consistent and relatively parsimonious measure, which persists across all regional banks. Kind of avoided some issues associated with lag times related to the publication process.

So we use quarterly output per author as our measure of interest, and then we construct a time series for each author. So of course, there's caveats to using working papers. Working papers are obviously just one dimension of output. For the 12 regional banks, we did not include publications, revisions, book chapters, notes, and other research contributions.

And then output related to policy work is not included yet, also important. And some economists use SSRN to release new papers, which we are not tracking. Of course, there's also caveats to using the pandemic, the pandemic was a unique time period. And so one perspective on results coming from the pandemic is you had this land-grab of papers on COVID.

You had more demand for papers on COVID, you had the short-run, long-run trade-off, less time on conferences, less lunch with colleagues, more just pumping out papers. So as robustness checks, we're gonna try to exclude Covid papers, we're gonna exclude 2020, vaccines were more widely available in 2021, seemed to be more in-person activities at that point.

And then we're also gonna check what happened around the financial crisis, should be a similar land-grab events.

>> Anthony Dierks: Okay, so we created a database for each regional bank, which contained an entry for every working paper written by a Federal Reserve author. We documented demographic information for each author, gender, job title, PhD year, starting and ending dates, mostly by consulting CVs, files and LinkedIn profiles, this is all public data.

We constructed a time series for each author at each regional bank, and then the Board containing quarterly output totals from 2018Q1 to 2021Q4. We excluded economists who did not produce working papers over this time period and were not present at the start of the pandemic. So now we coded the pandemic starting in 2020Q2 through 2021Q4.

And so what we have is we had 437 authors with 1500 working papers from the regional banks. And then we had 500 authors with about 2400 research pieces from the Board.

>> Anthony Dierks: Okay, so this is one of the first pictures, this is the summary statistics. And what I'm showing you here is the percent change in output per author per quarter for pre versus during COVID.

So left, you're gonna see the regional banks, in the middle, you see the board, then you see all is combined. And we see the regional banks, there's about a 20% increase for the board, you see about a 30% increase, and if you combine them, you get about a 25% increase.

The takeaway here is that the entire system had about a 25% increase in output for this time period. Now, that's just a summary statistic. We can also put in some regressions to make things more formal. And given that we're using count data, which is zero-bounded and right-skewed we can use Poisson regressions in addition to linear regressions.

Cohn Liu and Wardlaw that show Poisson regressions are more appropriate for this type of data. And we also can control for author fixed effects.

>> Anthony Dierks: All right, so now here's the regressions, I've tried to color code things, make it look a little bit nicer, but it's not showing perfectly on the screen.

Nonetheless, there's three panels here, there's a Federal Reserve bank, the regional banks, the board of governors, the combined, I'm just gonna focus on the combined. What you see here is the constant in yellow, it's about 0.25, that's papers per quarter, so you multiply that by 4. You get about one working paper per year per economist in the pre-Covid time period.

And then look in the Poisson, kinda shows you the percentage changes. So for the regional banks, you had about a 17% increase, the board had about 28%, and the combined was about 23.9% increase. So the takeaway here is that there were large significant gains across the entire system during the pandemic.

So the next question is, all right, let's start to split up the data to see who's driving this. The first thing we do was we split the sample up into the top and bottom half based on pre-Covid production. And so here again for the pre-Covid time period, you can see the constant of about 0.094, you multiply by that by 4, that's about 0.5 papers per year for the bottom half of the distribution.

The top half, you see a coefficient of 0.3, yeah, that's the constant, multiply it by 4 you get about 1.5 papers per year for the top half of the distribution.

>> Anthony Dierks: Then we can look at during Covid, what you're gonna see here is that for the Covid coefficients, about 0.13.

So the bottom half of the distribution nearly doubled its output over this time period. You then look at the COVID time, the top half interaction term. You see that it's negative, but you gotta remember that you add that negative term to the Covid term and you get basically a zero coefficient.

So the top half remained productive but unchanged. So the takeaway here is that the gains in output were really driven by the bottom half of the distribution, and inequality declined, which is very different from what we saw with the university professors.

>> Anthony Dierks: All right, we can look at males versus females.

Again, the constant here is about 0.25, which suggests that there's about one paper per year for males pre-Covid. If you look at the female co pitch, it's about negative 0.009. So that's an insignificant difference for pre-Covid between males and females, there's very small difference. During Covid you see, again, that males increased their output by 24%.

There's a negative coefficient for the females, but again, you add that to the Covid coefficient. See, the females increase their output by 21%, that's an insignificant difference. So the takeaway here is that there was no significant difference between males and females pre and during Covid. And again, that contrasts with the evidence, I mean, from the university professors, which found statistical significance.

 

>> Anthony Dierks: Okay, we can also do this by age. And so we split this up into under 8 years, 8 to 22 years, snd over 22 years, and for under 8 years, these are the younger folks. And what you'll see is that they had about a 16% more relative output than everyone else during COVID So you take that coefficient and you add it to the other one, and you see that they had 36% more output compared to pre Covid.

So the takeaway here is that the youngest cohort actually gained the most during the pandemic. And that goes against this view that kind of younges do relatively poorly under fully remote. Now, if you move to the eight to 22 years, this is more of the age where people have small children.

What you see here is you see a negative coefficient of about 17% less relative output than everyone else during COVID, but again, you have to sum up the two coefficients. So you take 0.34- 0.17, they still were 16% more productive. It's just the takeaway here is that although the gains were not as large as the other age groups, they were still significantly positive over this time period.

Lastly, you can look at the oldest age group. There you see an 8% relative gain compared to everyone else. You add that onto the COVID coefficient, and you say that the oldest cohort, they were about 29% more productive. And so here, the takeaway is that the gains relative to other cohorts were positive, but it was insignificant.

Okay, so that's the regressions. Now let's think about some of the robustness checks. So these gains could be driven by just the presence of new research ideas and not necessarily do with fully remote. And this view rests on the assumption that the constraining factor on researchers prior to Covid was not time, but a limited number of research ideas.

So the global financial crisis is another time period in which there was a new shock to do research on if there were no changes in work from home policies over this time period. So this can possibly serve as a good period for comparison. Okay, so what we did here is we do this for the financial crisis.

So we compare the number of working papers from 2005 to 2008 versus 2009 to 2010. And then on the vertical axis is the percent change in research output. And so what you'll see is in blue is the pandemic. And that's exactly the results I showed you before, they were relatively positive across all the regional banks and the board combined.

But then when you look at the situation for the board here, you see that during the financial crisis, the board had 18% decline in the papers that were produced. So system wide, if you include the board, the papers actually declined about 1%. So the takeaway here is that we didn't see a large increase in research output around the financial crisis, in contrast to the pandemic.

Another thing we can do is we can exclude 2020. If you exclude 2020, you see a relatively smaller effect, but you still see a positive and economically meaningful increase of about 11%. You can also exclude papers about the COVID or the pandemic. Once you do that, again, you see a smaller effect, as you might expect, but it's still a positive and significant effect of around 10%.

All right, what about collaboration? We can also check collaboration based on the authors per paper, and that's something that we did. If you look here, the constant here suggests that before COVID the average number of authors per paper was about three. And then when we compare that to the during COVID time period, you actually saw a 10% rise in the authors per paper.

So the takeaway here again is that there's a significant increase, potentially, in collaboration, collaboration across the system, even when controlling the time trend. So some reasons why, you know, work from home may increase output. The average US employee saves about 70 minutes a day by avoiding having to commute, prepare for work, which is split both in additional work and leisure.

Home working is often better suited for individual focused activities, like coding or writing, as it's usually quieter, allows for greater time flexibility. There's also this idea that you can control the ambient workspaces, clothing layout, ventilation. Then there's also the view from Raj that, it's potentially the case that these workers experience greater satisfaction and utility and will exert greater productivity-enhancing effort in appreciation of the nonpecuniary benefit.

So that's kind of the empirical results, I can now move you to kind of the theoretical model. We're just gonna use a standard New-Keynesian model with endogenous growth. You want to use endogenous growth because it allows potential effects on productivity. So given that Bloom et al, found that there was about a 70 minutes a day in savings, we can simulate just a 1% increase, exogenous increase in the time endowment.

And so given that a typical model has households working a third of the time, the total endowment is typically 120 hours a week, 40 of which are devoted to labor. So 1% increase is roughly an additional hour per week. And that's consistent with the findings of Aksoy etc al, 2023.

So what happens when we simulate that? Well, there's a lot of figures here but I'll go through them nice and slowly. So in the lower right, you're going to see that the time endowment persistently rises by about 1%. And what happens once that happens? Well, you see an increase in the labor supply of about 0.35%, and then you also see an increase in leisure.

But as you increase the labor supply, what's going to happen is that it's going to translate into higher output growth, which is also going to translate into higher consumption. And as you have more labor, it's gonna increase the marginal product of research and developments. So R&D investment is gonna go up, and that's gonna also lead to higher idea accumulation.

And as you have greater idea accumulation, that's gonna spill over into higher productivity growth. Likewise, in terms of the macro outcomes, you're going to see that as you have a higher labor supply, it's going to be associated with lower real wages and lower inflation. And then that lower inflation is going to translate into potentially lower risk free rates.

So yeah, you increase the labor supply, it's gonna increase output. As you have more labor, it's gonna increase the marginal product of research and development. That's gonna spill over into greater idea accumulation and greater productivity in these models. All right, and then here's just one last thing I want to show on this, and this is just very kind of relatively superficial.

There's a lot of issues with this, but this is just showing the level of productivity along with its trend. So you can see right here, and orange is the trend and blue is the actual observed productivity. And here I can then show you the growth in that productivity, how it jumped up during the pandemic and then declined quite a bit.

And on the right side, what I'm actually plotting is the number of mentions of RTO, return to office. And so what I'm trying to show you here is that the rise in productivity and recent decline are consistent with model implications based on an increase in the time endowment and then a reduction in the time endowment.

Now, there's a lot of other composition issues, so I'm not gonna try to stress. I'm just saying that what you see here is actually consistent with what the model is predicting. So in conclusion, we find that research output significantly increased during fully remote. The bottom half of the distribution was responsible for the large gains.

In addition, there was an increase driven by the younger and relatively older cohorts. There was no significant difference between males and females. Collaboration, as measured by authors per paper, significantly increase. Findings can be rationalized in a generally good model with exogenous increase in time endowment. Of course, the caveat here is this is just a study about productivity, there's a lot of other factors that go into companies' and firms' decisions Decisions about what they do with the remote work.

And there's just one final thought. I don't know how much time I have, if I'm going too fast or too slow. Okay, okay. Just one final thought real quick. This is just my final thought because I have this venue. There exists a scenario out there that combines positive aspects and is fully remote and hybrid.

And to me, it's like kind of this idea of cluster hybrid, where you kinda get everyone together for these core weeks, every six or so weeks, and you get many of the benefits of fully remote work where you have these cost savings. Nationwide talent search. People can live where they want, but then you also have this in person dynamic that's also very important.

So I just wanna put that out there, so. Thank you.

>> Audience 7: Hi, really interesting. My question is colored by the fact that I was a board Ra during the great financial crisis in the real estate section. And the thing that people are doing a lot when they're in the middle of these kinds of crises is doing a lot of policy work.

That's what some sections end up getting consumed by. And so what I was really looking for in some of these figures was a comparison of outcomes for the sections that were more or less connected to the policy work associated with that particular crisis, because that could be a lot of time that's spent on for some sections.

 

>> Anthony Dierks: That's a very good suggestion. That's a good suggestion.

>> Audience  8: So there's a lot of interesting variation across the different feds and how remote they are right now. So the fed board is, my understanding is something like six or eight days in person a month or something like that.

And I think the St Louis fed is relatively in person. And my understanding is it's fairly idiosyncratic which feds are choosing which path. So I feel like just by calling up a few people, you could figure out the policies and then do kind of a cross sectional test across the different federal branches.

 

>> Anthony Dierks: So that's a very good suggestion, and I do not disagree with you at all. So that would be good to look at. Definitely, thanks.

>> Audience  9: I wanna try to make sense of some of the numbers at a basic level. So you told us that you get an extra hour a week of labor supply from the time savings.

That sounds right. Just with the evidence that I've seen, that's about a 2.5% increase in labor supply on a 40 hour work week. Okay, but so somehow we got from 2.5% increase in labor supply to a 25% increase in output.

>> Anthony Dierks: You mean for the-.

>> Audience  9: For the empirical results.

And you're telling us the model has rationalized that, there's a big gap between 2.5% and 25%. I'm wondering where it comes from according to the model, according to Review, one possibility is that we're just more efficient per unit time when we're working remotely. Is that what's happening in your model, or is it spillover effects across.

 

>> Anthony Dierks: So this model is just a representative agent, just general equilibrium macro model.

>> Audience  9: Okay, you told us that the model rationalizes-.

>> Anthony Dierks: As I said, some of the findings. I was asking, does the model rationalize any of the findings? And what I found was that higher productivity was associated with increased time.

That's all I was conveying.

>> Audience  9: Okay, but does it quantitatively rationalize the results or just directionally?

>> Anthony Dierks: I was thinking more along the qualitative dimension.

>> Audience  9: Okay.

>> Anthony Dierks: So that's fair. That's a fair comment.

>> Audience  10: Hi, very interesting paper. I have two comments. The first is inspired by Lindsey's comments since I was also a board RA during the global financial crisis.

During something like a financial crisis, that seems really important for economists at the board and at the banks. And so obviously this isn't something you can address empirically here, but if we're thinking about potential policies, having staff be there when they really need them seems much more important than, or it seems qualitatively different than the effects you might expect on research output or sort of normal times, but you might actually have.

I don't know if there's any way for you to address that empirically. It would be interesting, but I don't know if it's possible.

>> Anthony Dierks: I agree.

>> Audience  10: My second comment was thinking about people who start in the post remote work era. That's sort of related to some of the previous papers we've seen.

I started a new job during the pandemic in the remote work era, and my research, what happened to my research collaborations was that I basically continued collaborating effortlessly with people at my old institution. But had much more difficulty starting new collaborations at my new institution.

>> Anthony Dierks: Yeah, that's very fair comment.

Very fair.

>> Callum Williams: I'm Callum Williams from the Economist. I just got a quick question about this COVID exclusion that you applied to the data to sort of essentially control for the demand shock that you get to research during the pandemic. I guess I wonder whether that's sufficient control, because certainly my experience as a journalist was that there was a huge amount of demand for news or huge increase in demand for all sorts of news during 2020 and 2021, even things that were only tangentially or not even at all related to Covid per se, the role of the state, the role of the welfare state, the role of fiscal policy, the role of monetary policy, chinese american relations, and so on and so forth.

And so certainly The Economist, one thing that drove productivity for us was just that there was a lot of demand for news as a whole. So I wonder, is there a way around this? Are you able to focus maybe on a division of the Fed that has absolutely nothing to do with the news at all and see what happens to productivity in that department?

And that would seem like a clean way of isolating the effects you're trying to describe.

>> Anthony Dierks: That's a good suggestion. So I'm not sure that I can do that. That's a good approach, potentially, to address that.

>> Audience  11: So, very interesting paper. So I just want to build on what Steve said about how 4% endowment of extra time leads to 25% productivity.

So, drawing on my own experience, I was thinking whether the collaboration production function changes here, because the search cost of looking for distant collaborators is now changing. So, personally for me, I started working. Actually, contrary to the point made earlier, I found more distant collaborators during this period.

It's something you can empirically test. Are they going beyond the DC area to collaborate with economists elsewhere?

>> Anthony Dierks: That's very good. Yep, good question.

>> Audience  12: I also had a related comment, and following up on Raj's comment, I was wondering whether the COVID actually kind of democratized access to new research networks in the sense that now you have access to MBR meetings but now it's online.

You reach out to people. Someone from Harvard and someone from fed. Maybe, probably you guys had access to this type of networks, but now it's easier to reach out to these people. I think with the data you have, you can understand whether quality of quarters also increase.

>> Natalia Emanuel: That's very good.

Yeah, definitely.

>> Host: All right, I think we're good for the final. Great. So the final paper in the session will be presented by Ben Etheridge, again, about worker productivity during COVID and in particular looking at adaptations.

>> Ben Etheridge: Great, thank you. It's great to be here, of course. So this is joint work with Ashley Burdett and Yikai Wang, who are colleagues, Essex, and Li Tang who used to Essex and is now at Middlesex.

Okay, so I'll be brief with background and motivation. Okay, so the kinda bigger picture question that we're kind of thinking about with this study is that the rise of working from home was, of course, one of the most salient features of the COVID pandemic. And then post pandemic, I think surprisingly quickly, most of socioeconomic life has returned to the old normal, as if the pandemic was kind of just a bad dream.

But working from home as kind of the new normal has persisted. And so an important question is why? And several papers in this conference are about that. Well, the obvious answer is because people like it. But the more subtle question is, how do we get here? Okay, so what we do is explore individuals' experiences of working from home during the pandemic in the UK setting.

And when I talk about experiences, that's broadly, I mean, I'm gonna focus on work productivity, which as economists we think of as probably the most important thing. And also gonna think about the role of their circumstances. So our aims are to provide new evidence on what causes effective working from home and how this drove working from home adherence as the pandemic progressed.

 

>> Ben Etheridge: Okay, so what are we actually gonna do in this paper? So we're gonna briefly go over the data. So we use representative panel data from the UK Household Longitudinal Survey, which we don't know is a kind of a standard panel data like the PSID, although it's much larger and much newer.

So it's kind of more representative than the PSID, which has particularly rich, and I'd say uniquely rich data collected throughout the pandemic, including information on the self reported productivity. And just to be concrete, I'm going to talk about change relative to a baseline, so people report plus 10%, negative 10% or whatever.

And then work location, whether they're in the home or in the office. And actually that was collected before the pandemic as well. Okay, so if I had more time, I would sell the data, spend half an hour talking about the data, but I don't. So I'll just say a few bullet points to advocate their strength cuz I think they're a really useful resource actually going forward as well, because there's data before the pandemic, during the pandemic, and there soon to be data kind of post-pandemic.

So it's kind of unusual around the world that we see the same individuals quite high frequency across the whole pandemic. And the question on self-reported productivity is also kind of unusually precise in that it asks about the speed of task completion. So it has quite a precise interpretation.

And the paper, we do quite a lot of validation of these data, both externally, kind of predictive validity in the language of psychologists, and internally. Because we actually kind of have more than one measure within each wave, and then we have all these measures across waves. Of course, I don't have time to talk about those, but you can ask about them afterwards If you like.

 

>> Ben Etheridge: Okay, so to set the scene with some basic patterns to kind of talk about our contribution. Productivity during COVID-19 was on average similar to before, but there's systematic heterogeneity in several dimensions, so over time, going in and out of lockdowns. But today, I'm gonna focus on a couple of aspects which are gender, famil,y and job characteristics.

So, I mean, these things, I guess, have been known through various studies by people here, but I'm gonna kinda speak to them more in a moment. So, for example, mothers performed worse during lockdowns, although we see some catch-up between mothers compared to fathers as the pandemic progressed. So the second lockdown, mothers come more closely to fathers than the first lockdown.

And as we all know, kind of high earners performed better. And there are multiple reasons for that. Their jobs are more remotable, that kind of thing. Okay, so kind of more novelly, we wanna make a couple of kinda points in this paper which kind of speak to adaptation to working from home as the pandemic progressed.

Okay, so we can track individuals, as I said, over time and where they were located, and we find that there was systematic sorting based on idiosyncratic productivity experiences. So I've just spoken about kind of observable characteristics, but now, I'm talking about condition of those characteristics. We see that people who report being more productive when they're home, they're much more likely to be at home later in their pandemic.

But actually we see that the marginal group shifted over the course of the pandemic as we kinda went into and out of lockdowns. And so I'll kind of invite, I think, illuminating ways and I'll come back to that when I talk about the results in detail. So one implication of this is this kind of novel kind of micro-evidence on micro-behavior that kind of speaks to why in the second lockdown, for example, the UK economy performed much better than in the first lockdown.

And again, there's multiple reasons for that, but this is one contributing factor.

>> Ben Etheridge: Okay, so secondly, what we do is we wanna think about which factors affect productivity across locations. So I've just shown that location is endogenous. So we do that in the paper by addressing or kind of sorting using pre-pandemic commuting patterns about which there's kind of rich information in the survey before the pandemic.

So people report how they were traveling to work. Okay, so building on what I said earlier about focusing on gender, family, and job characteristics, so we find broadly that kind of these differences were kinda most seen in the home location, okay? So in office, when we think about the production function in the office, people look kinda fairly similar to before the pandemic.

And these kinda differences emerge in the kind of home environment production function. So parents performed worse at home only. And then going into more detail about this high earners, we look at some kind of components of being a high earner. So we see that managers perform better at home only compared to non-managers.

In the office, we see no difference compared to pre-pandemic. And then something that you don't get from kind of studies of individual firms, we can have information about the size of the firm that people are working at. And obviously, people working at large firms tend to be kind of higher earners.

And we see that employees at large firms perform better compared to employees at smaller firms at home only. We don't see a difference in the office. So again, that speaks to kind of evidence that's out there that larger firms were more able to kind of invest in technology which help promote work.

And then if I have time, we have many other results on kind of housing conditions, well, a few results in housing conditions. And then a nice feature of the survey is, before the pandemic, we have this rich information on personality traits and cognitive skills. And so we see some kind of interesting results on that level as well.

Okay, so to set the scene for the collection of the data, so you have in mind, also the pandemic in the UK and when the data were collected. So we have 2020 and 2021, and the orange line is hospitalization rates. And then these dark gray bars are lockdowns.

And these light gray bars are these kind of semi-lockdown periods. And obviously, the white periods are when we had more or less kind of freedom of movement with some caveats. So the top blue dots are kind of regular survey, well, survey collection, and then we have the COVID survey, which was done very high frequency.

And then the questions on reported productivity With the green dots. And there's some kind of subtle differences in nature of question, which I won't go into. But importantly, we have two collections in kind of lockdown periods in June 2020 and in January 2021. Which was kind of heaviest lockdown in the UK or the heaviest kind of incidents of COVID.

And then we have these two from kind of lighter, when there's more kind of freedom and people are kind of free to return back to the office. Okay, so I'm gonna kind of talk about our work on how people's productivity experiences drove their location choices. Okay, so to do this, we were on ordered logic models of working from home on a full set of individual and job characteristics.

And a full set of lag terms of where you were in all observed previous periods with the previous last working from home status interacted with your productivity, reported productivity changes before the pandemic. Okay and so the groups we have, I mean, yeah, so in the survey, there is a slight finer segmentation.

We've grouping people into not working at, homeschooling the office and then full time working at home. And then we have kind of hybrid workers or part time at home in the middle, okay. So we see working from home doesn't just depend on occupational characteristics we knew before, but also in these individual realizations of productivity.

So the marginal group changes. So we've got three transitions going from the first lockdown to the first kind of eased restrictions. And we see that you're much more likely to be working from home or more like to be working from home if you were working from home in the first lockdown full time and reported higher productivity.

So these are people all thrown back home. Obviously, the ones who found they were doing well, they're more likely to stay at home people who didn't like it or weren't doing well, they were brought back into the office. Obviously there's a joint decision between individual and firm, or maybe it's decision or by the firm.

There's some bargaining, I suppose. Okay, in terms of quantification, so we find ten percentage point higher productivity for a threshold worker. These two, three percentage points more likely to be working from home in the next period. So, I mean, that estimate is likely highly attenuated. I mean, so our measure of productivity is obviously measured with noise, as I said, in principle, we do have more than one measure productivity in each wave.

So we could try and kind of do a measurement instrument for that. But it's very non-classical, so required, it's probably kind of brilliant, feasible. And I'll come back to this later. So unfortunately, there's no data for those who are never at home in this first lockdown. Okay, so I've got these three transitions.

So I'm talking about the first transition on the left hand side. I said the full time, we see that they're more likely. So I've got these kind of marginal effect of lagged change in productivity from the previous period. And so we see this dark blue is the significant result.

Those who are working full time at home, they're more likely to go to stay at home in the first easing. Okay, so going from then the first easing in September to the second lockdown, we have that, of course. Well, those who are full time working at home were now whether or not they reported good productivity, it's irrelevant because everyone's kind of coming back home.

So it turns out to the people who are working part time at home that they're kinda more responsive to their productivity. Okay, and we see this in the middle picture. So on the right hand side of the middle picture, that's people who are full time at home. Of course, the standard errors are now large because there's no movement there.

They're all kind of people take some misclassification. They're all staying at home in the lockdown, but the people who are part time at home, they're still responsive. Okay, so then going from the second lockdown to the end of the survey when kind of COVID is kind of just in the background and there's vaccinations around and people are kind of feeling more easy about it.

So now, interestingly, I think we find that this marginal effect is strongest for those who were full time in the office. So now those who are more productive than pre Covid are much less likely to return home. So the interpretation is, okay, so these people, they've had some experience of working from home.

In a second lockdown, they're back in the office. And now if they find that they're productive in the office, they're gonna stay there. These are people who clearly don't like being at home. I go, or vice versa. They go back into the office. They weren't productive there. So maybe they return home.

Okay, so this is on the right hand side. The kind of heavy blue line is the strongest quantitative results. Unfortunately, we don't have the same guys or people who were working in the office in the first lockdown, as I said, for data collection reasons. Okay, so we've shown that your reported productivity drives location as the pandemic evolves.

Of course, the first sort of question is thinking about how does location affect your productivity in the first place? Okay, and that's obviously question in the QJQ paper is what is the treatment effect of working from home? And coming at this from a kind of Rory framework model, Rory model framework, I think of the treatment effects.

And then you also think about what's then production function across kind of treatments here, what drive, what are the marginal effects of different characteristics at home versus in the office? Okay, so we're gonna focus on number two, because our kind of data are not powerful enough to kind of estimate treatment effects of working from home.

And of course, the strength of our setting is we have kind of representative data across occupations, whereas it doesn't really make sense to talk about treatment effects compared to combined treatment effects across hospitality with IT coders. Okay, so we're gonna think about what's the effect of being a parent versus non parent in these two locations.

Okay, so, as I said, we're gonna use kind of Roy model sample selection, model instrumenting now a kind of binary for working from home. So we combine the part time with the full time into any working from home with people who are full time in the office with, we're gonna instrument with pre pandemic commuting mode.

As I said, that's collected. That's not just retrospective questions. That's collected before the pandemic interacted with distance and difficulty of travel. Okay, and yeah, I'll move on. Okay, the instruments are fairly strong, so distance and difficulty commuting important for both car and public transport users. So I guess there are multiple interpretations of these.

So we actually find. I'll go to the results, actually the main thing is this distance to work with, interact with public transport. So our interpretation of that is people don't like a long commute by public transport during the pandemic, don't like being spoken to, exposed to infection. The car, its travel difficulties, I guess interpretation there is more they've had five years of this horrendous commute by car, and they just fed up with it and they jump at any chance to work from home.

Okay. So I've already described some results, and I said, important effects of characteristics are mainly seen at home, in the office, we see fewer differences across people. I mean, there's more subtleties to this. So, as I said, we see parents at home, they have a negative effect compared to non parents, relative to non parents.

Although conditional parenting, the female penalty is seen in the office. Okay, as I said, managers, employees, large firms, they have this kind of positive effect relative to non managers, relative to employees at small firms. In the home only, but not in the office. And then I'll show you the results in a moment, but other effects kind of not discussed today.

So housing characteristics, so we have, there's a question on having enough desk space or having desk space that's important at home. Obviously not in the office, that kind of as a falsification test of our data, I suppose. And then I think some interesting results, so the cognitive function is actually quite strong in the data, and it's surprising.

So we find that the returns to being smart, seem to be blunted at home, which I think. I think I was kind of expecting the opposite. I mean, there's, you know, there's a narrative that, being at home allows us to do what kind of thinking work. But this is obviously all conditional on the job type, but within job type.

Yeah, returns to cognition seem to be less effective at home, but it's the same in the office. And then personality, we actually see positive effects kind of across locations. So it speaks to more the kind of pandemic environment in general, of agreeableness and conscientiousness. And we discuss an interpretation of those in the paper.

Okay, so I'll show you the tables from these, okay, so at the top, we have this negative parent effect. A home, but not in the office, I said the female penalty. It seems I've done this as a male here, but, yeah, think of it as the female penalty is in the office.

And then managerial returns to kind of, or the positive effect of managers for relative to non managers is at home, not in the office. And log size of firm, again, home, not in the office. Okay, and then finally having desk space at home, but not in the office.

Okay, so to round up, so we examined productivity during COVID19 and its relationship with the evolution of working from home. Productivity was highly hedge genius, and working from home tech have evolved as individuals experience productivity realizations. And the determinants of productivity vary across locations.

>> Audience  13: Thanks for that, a lot of interesting data there.

You talked about attenuation bias, in your assessment of the impact of productivity during the pandemic on later, take up a work from home. And you stress measurement error in your, the role of measurement error in your measure of productivity. But that's not the whole idea, there's two sources of measurement error cuz there's a conceptual measurement error.

Because what should drive the change in behavior isn't the idiosyncratic productivity realization, it's the surprise component of that. If it's just that you're more productive in work from home mode, but you already knew that before the pandemic, that's not news. There's no reason for that to drive behavior.

So in the work I've done with these first three guys over here and others, what we look at is not the idiosyncratic realization. But the surprise component of the realization of productivity and work from home mode, and relate that to later work from home outcomes. So what I'm trying to say is there's not just a measurement error source of attenuation in the key relationship you're estimated.

But the concept that you're actually sticking on the right hand side is not the theoretically appropriate one.

>> Ben Etheridge: Yeah, right, point taken. I mean, certainly the first lockdown, we would say it's all innovation, it's all kind of unforeseen, I guess.

>> Audience  13: No, no, that's not right. The event of the pandemic's unforeseen, then the impact on, that's different than the surprise component of my productivity.

You can implicitly assume that the realized, the idiosyncratic component of the realized difference is identical to a surprise, that's fine. But that's an additional, that's a maintained assumption. And to the extent that assumption is violated, you're gonna get a second source of attenuation bias.

>> Ben Etheridge: Yeah, okay, so if I understand you, but let's say that no one knew how productive they were gonna be at home in the first.

 

>> Audience  13: Yeah, that's the maintained assumption.

>> Ben Etheridge: Yeah, sure.

>> Audience  13: But presumably that's in practice, some people had some sense of how productive they would be. And so even conceptually, even if you had no measurement error in your productivity measure per se, there's still this conceptual source of measurement error.

Which is leading to a further attenuation bias. That's the only point I'm trying to make.

>> Ben Etheridge: Sure, okay, sure. And I guess leaving, exiting the pandemic. Okay, so let's go back to the first lockdown versus later on. So even if they had some expectation that they knew how productive they are at home, we're looking within the pandemic.

So maybe they don't know how productive they are at home in a pandemic. Leaving the pandemic, we're going back to a world.

>> Audience  13: To put it in a constructive way, you've got this very strong maintained assumption that idiosyncratic realized variation is identical to the surprise. So you could just vary that assumption and say, okay, suppose the signal, the noise, think about exactly how to do it.

But you could make alternative assumptions, and that would generate an implication for how much attenuation bias you're getting from this particular source.

>> Ben Etheridge: Sure, okay, thanks.

>> Audience  14: So I was curious if you could say something more about your finding that manager that for managers working at home was particularly positive.

And, I mean, that seems a little bit counterintuitive if you think that, yeah, mentoring people, supervising people is different when you're remote one. And second is, so we have this result in our survey of working arrangements and attitudes that among people over 50, managers really wanted to be back in the office, but non managers didn't.

So I think trying to interact this performance of managers against other characteristics like age or, okay, perhaps occupation might be something interesting to look at.

>> Ben Etheridge: Right, okay, yeah, sure. I mean, so I still hadn't seen the result, band managers wanted to go back into the office, right, sure, okay.

Yeah, I mean, there were, of course, I don't want to downplay that. I don't want to kind of, I suppose, there are so many aspects to kind of productivity and as I said, it's kind of speed of task completion. I mean, there is one interpretation that managed in particular, there's so many distractions.

And so in terms of speed of task completion, when you're at home, you can unclick your turn off broadband or whatever you're doing. Maybe you can kind of switch off the phone. So, yeah, yeah, I mean, it's, yes. I mean that yeah, no, I'd have to think about it more and I'm trying to not.

I suppose there are so many aspects of productivity and we're focusing on one of them. Maybe that's both a strength and a weakness of the study. You can think about doing these kind of. It's kind of making people think about these kind of quite narrowly specified tasks, maybe rather than their job more broadly.

 

>> Audience  15: Yet really interesting one question is, we get asked a lot and I guess many people get asked is about self-assessed productivity. So whenever I talk to non-academics, though, I was kind of skeptical of self-assessed productivity. One thing that made me think about this, as well as Natalia and Emma's paper from earlier, because there's two measures of output.

One is what I achieve per day, but that ignores. Maybe I mentor and provide input to others. And the second is also provide input for others. And I wonder, I don't know what people mean when they respond, but if I'm like a manager, I may, a lot of my output, I may be better at doing the day-to-day thing, but not as good at training and mentoring others.

It's a kind of an open. It's like, how long is a piece of string. But for us, I don't know anyone else here knows of anything where they compared self assessed to realize. Because it's kinda hard to know what people mean.

>> Ben Etheridge: Yeah, no, so I agree on the kind of in the, my interpretation is people are not thinking about the mentoring when managers ask this question right to you.

So yes, I mean, yeah.

>> Simon: Thank you, Ben, super interesting.

>> This is Simon. One question about defining that productivity outcomes at home are more positive at larger firms, could it be driven better management practices, better work-from-home technology at these large firms?

>> Ben Etheridge: Yeah, right. Yeah, sure. Yeah, no, no, that's the interpretation I think these large firms were able to invest in.

I mean, it's been documented and discussed, yeah.

>> Audience  16: Large firms also have open office floor plans which have been shown to actually be detrimental to productivity and interaction. So maybe working from home is just more productive because you get more space to do it. Raj, go back to your US patent and trade office a study and ask and redo it.

But ask people for their perception of like. So if you think back to the question about received versus actual, there probably are ways to do it, but you have to do this sort of work where you're looking at patent output, patent quality on the productivity on one side, but you're also pulling workers at the same time.

It's an interesting idea because I've been reliant on the survey aspect of it. I've never seen the two done side by side. It'd be really interesting for somebody to do that. Yeah.

>> Host: Great, so I think that's it for the sessions.

Part 5:

>> Jose: All right, so the first session of the afternoon, we'll start with Lindsay Relihan telling us about the impact of work from home on brick and mortal retail establishments. So take it away.

>> Lindsay Relihan: Thank you so much to the organizers for having our people on the program and for putting us on right after lunch.

That's very optimal. This is joint work with my co-authors at JPMorgan Chase, including James Duguid, who's also here today. So I think a great starting point for this project is to think about how most of the economic research on cities over the last sort of half century has really focused on the productive benefits of cities, like how cities make firms better at making a bunch of stuff.

But in the last two decades, what we've started to also focus on are how cities have become places which bring extra consumption benefits, right? So these are places where you can go out to great restaurants or great bars, or see a concert, go to nightlife venues. And these kinds of consumption benefits have also been driving people to locate into big cities and to especially locate in very dense cores of the downtown areas.

And COVID-19 has the potential to sort of disrupt that part of us, modern city life. So of course, many Americans are now working from home. So you're breaking where you're working. And lots of other research has shown that this has caused a huge shift in where people are also living.

So first and foremost, moving mostly from downtown areas more towards the suburbs, and also secondarily moving from some large cities towards smaller cities. So what this paper is going to do is really try to explore what this fundamental change where people live and where people work from work from home is doing to the location and composition of retail establishments.

When the shock to retail markets is really coming through the customer base and locations of their customers, rather than through their exact workforce. So retail workers largely remain in person. So there are three ways we sort of think of being the primary channels through which work from home.

And this fundamental separation of these locations could affect retail markets. So the first is through the fact that a substantial share of shopping trips are going to take place starting from the origin of a customer's home. So retailers really value being super close to where people live. And so if you have a bunch of people changing where they live because of work from home, then retailers are gonna be drawn to follow them.

The second way that this might affect retailers' desired locations is through the fact that many consumption trips are tied to work locations explicitly. So this might be because you pick up coffee on the way to work or you go to the grocery store before you go home or you're going out for lunch with your coworkers or happy hour.

There's lots of this kind of consumption that happens because of where you work. And now if you're not going to those office places, then that kind of workplace related consumption might not be taking place in the same way. And then the third way is that people who work from home might just have really fundamental different levels of demand for different kinds of products.

So if you were somebody who used to go down to the office and you went out to Chipotle for lunch. But now you work from home, the question is, well, will you go to Chipotle in the suburbs, or will you just make a sandwich at home, right? That's a different kind of substitution.

So we could think that there might be lost demand for these kinds of consumers. And what we're gonna show essentially is that these three channels are all present in the data. And this kind of impact that we'll show you really has ties to many of the various implications of work from home that have been surveyed.

And Gill Duranton and Jessie Handbury's work on the structure of the city. What's the value of different kinds of real estate in different locations? How do labor markets function, and the productivity of the city.

>> Lindsay Relihan: So there are really two research aims, actually, of this paper. One is this really centered around what is the effect of work from home and brick and mortar retail.

And the second goal of the research is actually to add to this huge data building exercise where we're going to use credit and debit card transactions data to build a panel of store entry and exits to track retail establishments over time. It's like a new data asset to understand where these kinds of stores are located across time and space in a detailed way.

There are three main areas of research that I feel like we connect really well to. The first is sort of the early efforts to understand the pandemic's effect on retail markets. A lot of that early work, including ours, is really trying to understand sort of the scale of the crisis in the moment.

There's been a smaller amount of work that's looking at sort of longer run effects on small business closures, but that's not necessarily strictly retail, but other kinds of small businesses. And of course, there's great work on the effect of work from home on real estate more broadly. Several of those people here in this room and presenting other works.

But most of that work is focused on is the effects on the office market, so people leaving the office. So that's gonna obviously be a direct link there, and then effects on residential markets. But more broadly, I think this research really speaks to how do technology shocks reshape cities.

And so we're part of a much larger literature that tries to understand that kind of dynamic from other kinds of shocks, like transportation or online retail. Just a brief word about the data. So we are working with all of the debit and credit transactions of about 70 million Chase customers from 2017 to 2022.

So it's a really big data set. So Chase is the largest consumer bank in the US, for those who aren't familiar. And so it covers a huge, substantial portion of the residential base in the US. We see a really rich set of fields associated with each of these transactions, whether or not the card was present at the sale terminal, the ZIP code location of the merchant.

And a textual description of the merchant, so that you bought at Starbucks number 123 and ZIP code 12345, for example. And then, because we have lots of information on the customers, we know their exact address, but we're going to use their ZIP code in this study. So we're going to be limiting, in this first pass, to 16 of the cities with the largest Chase customer footprint, because we're going to be relying on these customers to essentially act as surveyors to tell us where stores are located.

So we want a really dense customer base to get full coverage of the kinds of stores that might be in existence. And from these two sort of filters, we're gonna end up creating a quarterly panel of 1.7 million establishments at a quarterly frequency through the fourth quarter of 2021.

And here, with this data, what you should really think about an establishment is as a unique combination of merchant name, establishment number, ZIP code, and product. And we're gonna do some benchmarking to the county business patterns data, which is the census product with the finest geographic granularity to show how we compare in terms of levels and growth to an established product.

So here is the benchmarking results for 2019, where on the left side, we're showing you the levels comparison across ZIP codes. Where each bubble here is a ZIP code with the establishment count coming from the CBP in logs and our establishment count in logs on the y-axis. And the growth comparison here is for growth rates in each ZIP code from 2018 to 2019.

So one thing you'll notice here is that in levels, we do a really good job of matching this existing census product. We do a particularly good job in matching for restaurants where Most of these businesses are going to be employee based, and the census products are actually mostly using things like payroll reporting to describe where establishments are located.

So they'd be more likely to miss things where an establishment did not have any employees, or they were a single sole proprietor business. And so if you have lots of sole proprietors doing things like corner bodegas, like grocery stores, which is up here in the top left, or personal care services, like a single person cutting hair in a living room, that's something that the CBP might miss.

But if you have a card terminal, like a Square terminal on the end of your iPhone, it's something that we pick up in our data. Now, you'll see that there's very little correlation in terms of year to year growth with the CBP. One other important limitation to the census product is that it really only does a comprehensive survey every five years.

And then uses some other sort of marginal administrative data sets to try to track companies when there seems like there might be a probable change in their organization. So they're really not set up for this very high frequency measurement, and measurement in a fine detail in changes. And so we think that we're actually showing here that we're gonna make a contribution to be able to track this at a high frequency.

So another thing that I think is important to point out is that this is a different way of thinking about where stores are located. This is one of the favorite stores I've ever looked up in the data. This is the suburbs of Columbus, Ohio and the grandly named Pizza Chateau.

I don't know if you can see that name, but this is obviously somebody selling cigarettes and pizza and hot dogs from their front living room sort of store. And this is something that might not very well be picked up by traditional data sources, but it's something that we see really well in the data and is actually an important source of consumption for the people that live in this neighborhood.

Another thing that can be different about traditional products is highlighted by this picture, which is perfect for this crowd. This is the Ferry Building down on the Embarcadero here in the Bay Area. And this is one of the ZIP codes in which we show the greatest difference with the CBP in terms of the count and growth of establishments.

And one of these reasons for this is that you see there's a lot of farmers markets in the summertime, and inside there's all these really small merchants. These are temporary establishments, right, that might not be reporting these as employment locations, but they're still generating transactions, these locations. So they're actually places where consumption is happening in temporary stores that might not be picked up by traditional sources.

Great, so, as a gut check, we can show you some of the basic establishment patterns that we see in the data over time. And the top panel here, what we're showing you in the light blue line are the active establishments benchmarked against the 2019 Q4 establishment count. And then in the blue line, we're looking at the establishments who are in existence.

So the gap here is sort of establishments who have closed down because of the pandemic. And you can see that about 15% of stores completely closed for all of Q2 of 2020, but only about 6% of them closed permanently at that time. And that has narrowed as we've gotten farther away from the acute phase of the pandemic.

In the bottom, we're looking at entry and exit rates relative to the stock in the previous quarter. And you can see that right when the pandemic is most acute is when we're seeing this huge surge of exits. So this sort of just matches some basic features that helped us get some confidence in the measurement that we're doing.

So now getting into some of the basic details about how work from home has affected brick and mortar retail. as a first basic fact, we're gonna do some cross-city analysis, where on the x-axis, we're looking at establishment growth. And on the y-axis, we're looking at population growth. On the left, we're looking at the pre pandemic period for 2018 to 2019.

And what you'll notice is there's a lot of crowding here, where very few cities are experiencing extreme population growth or decline. In fact, some of the bigger cities, like San Francisco and New York, are already experiencing some small declines in population. But everywhere, establishments are growing really strongly in the city.

It's not really related to how much population is growing or not. But in the two years of the pandemic that we're focused on, you can see that now there's a super strong correlation between population and establishment growth. And what we think this is showing is that the kinds of customers that these retailers are really focused on attracting and maintaining are exactly the kinds of customers who are now sorting across cities.

So they're going to want to be following these kinds of customers in a way that they would not have had to before. Now, the rest of the analysis is going to really look at what's happening within cities. And to do this, we're gonna normalize the distances in all the cities where the downtown ZIP code is gonna be a distance of 1.

And then the farthest ZIP code from that within the CBSA is gonna be a 1. So every city will be on a 0 to 1 scale, and then we're going to estimate average population and establishment growth at each of those points across all of the cities. So this is probably my favorite picture of the research.

So I'll go through it a little bit slowly. On the left, we're looking at population growth and establishment growth. And then these curves here, you can think of as the growth curves relative to before the pandemic, at different points along the pandemic. So in the orange growth curve, what you're gonna see is what was initially happening in the very early parts of the pandemic in terms of growth of population and establishments across these cities.

And you'll notice that we see very little movement initially in population because we're focusing on our customers' permanent addresses reported to the bank, and those have not really moved at all relative to the pattern yet. But because we're using real time tracking of establishments, we see immediate changes to the growth curves of establishments across the city with about 13% of downtown businesses immediately closing permanently with almost a much smaller sort of negative 1, negative 2 effect out towards these city edges.

And the downtown areas, like the closest 10% to the downtown, that's where really this enormous negative effect is highly concentrated. Now, in the most recent quarter of our data, you can see that a lot of things have changed since those initial stories. So our customer addresses have now caught up, and our population growth curve is now dramatically rotated.

Where in the downtown area, more than 9% of people have permanently left, according to their addresses reported to Chase. And in the farther suburban areas, you're seeing population growth rates in excess of 5% and 6%. And that's a pretty linear relationship. But for establishments, what we see is more of a convex shape, where the downtown closures are still fairly pronounced, but there's not a great amount of growth in the outer suburbs.

It's really this inner suburban area where you're seeing the most robust growth. So now, we'll dig into a little bit of what that is by parsing out establishment growth by particular kinds of products. Sorry, I'm gonna skip that for a second. So in the top left, we have grocery, and underneath that, we have restaurants.

And these are the two product categories that are dominating the overall pattern because they're accounting for the largest share of the number of establishments. And what you'll notice is that grocery stores seem to be able to pretty well follow the population even out towards the exurban areas. But it's restaurants who have really been only able to grow in the sort of inner suburban ring.

Now, there might be a lot of reasons for this. One might be something like zoning that prevents restaurants from going out really far into the suburbs. But also things like population density requirements for restaurants that are not the same for grocery stores. So it's really restaurants that are driving that convex shape, while grocery stores are able to go all the way out to the experts.

A couple of other interesting patterns that come out, if we look at leisure and professional consumer services. So leisure, you hear you should think of things like gyms or bowling alleys, other kinds of like local entertainment outlets and then professional. Sorry, I wanted to say, yeah, professional consumer services is things like vets and daycares.

So everybody who bought a pandemic puppy needs a new vet. Even though they're still down overall. They do have this monotonically increasing relationship with distance away from the CBD and even out in the outer places, we are already seeing positive growth. This seems to be also a kind of product that would still be demanded by those who work from home, and maybe even more so.

That contrasts with clothing here in the top right under that, personal care services. These are product categories in which across the board you really see mostly quite negative effects and they do not seem to be able to follow the new work from home population. So I don't know what you guys were doing in the pandemic, but mostly what I was doing was programming in my pajamas and not getting haircuts.

And I think that that's closely reflected in these two figures that you don't demand as many of those things when you switch to being work from home. One other interesting note you'll see is here that pharmacy don't have any growth curve cuz there's basically no statistically significant change across space.

And we think this partly reflects some of the licensing requirements for these kinds of products. And then finally, I'll say for general goods and home goods, these products are also experiencing negative growth all throughout the city. Something to dive into a little bit more in future work. What is a good, I think, hypothesis for this is that these kinds of products were already heavily competing with online retail before the pandemic, getting hit by a double shock like this one.

They would have a harder time sort of coming back from that and following consumers. Great, so at the end here, what we'd like to do is a simple decomposition analysis to try to formalize some of these relationships that I think have been compelling in just a graphical way.

So what we're gonna do is measure just the simple OLS regression, the effect of work from home on establishment growth over these two years. We don't have any direct measure of work from home in our data. So what we're going to do is sort of combine a couple of different measures.

The first is to sort of measure directly population change within our data at the ZIP code level, and save. Most of the population change over this time period is directly due to work from home. And then we'll use measures from the American Community Survey to try to understand sort of the behavioral impact of the local population, either in terms of residents or in terms of employees shifting to work from home based on their pre pandemic occupation composition.

And then we'll include some other neighbor controls like online and offline spending growth, and some characteristics, basic characteristics from the American Community Survey of the local population, and then including city fixed effects. So in the interest of time, I'll just talk about the full model, how am I doing on time?

Great, I'll talk about this last column where we include all of the controls. One thing I'll say is that once you control for population growth, there's no extra impact of a downtown core sort of dummy. All of the negative impact for downtowns has really come through this loss of population.

And then of course, if you have more people, population growth is positive. You get more establishment growth in the data, which is what you would expect. But the work from home residential exposure and work from home employment exposure measures are really strongly negatively associated with establishment growth during this period.

So people who are working from home or not going out as much to do things like clothing shop or go get their haircuts, or even sort of interact with local economies. They're not there to buy books or flowers in a sort of tangential way. And then similarly for the employment exposure, which is actually driven entirely by restaurants, this is really about people not going out to eat associated with their jobs.

We don't really see much of an effect of online spending growth, it's actually slightly positive. Our major working theory of this is actually most of the online spending growth during the pandemic that was sustained was in products like online retail or online groceries and online restaurants. Which if anything, require more grocery stores and restaurants to service, unlike sort of the tradable goods sectors which grew in online retail space over the last few decades.

So in conclusion, what we find here is that work from home is really driving a dramatic reallocation of retail to new locations and changing the product composition of those stores. That's driven both by changing in residential and employment locations and the kinds of goods that people who work from home demand.

There are a lot of implications here. One interesting time that I'd like to point out is that there's a comment made yesterday about how we think that work from home might be an exacerbator of economic inequality. And I think that might be the case that some of these results might lead to as well.

There's a great paper by Conrad Miller who showed that during the suburbanization that was driven by the growth of the highway system, that really exacerbated the black-white employment gap. And here, if we have a bunch of retail stores moving to the suburbs. But the people who normally work in retail, who are low income and non-white and living downtown, can't move to the suburbs in the same way.

We might see further exacerbations of the employment gap between these low socioeconomic status individuals and those who live in the suburbs. And then finally, in urban economics, it's quite common to not think too much about amenities in a location changing over time when we model cities. But I think this really speaks to the speed with which amenities can actually change in a location and the importance of modeling these kinds of dynamics in a rich way in our urban models.

Thank you so much.

>> Speaker 3: I was curious to dig into a little bit more about how some of the differences in terms of which establishments are able to move into the suburbs are affected by staffing and recruiting workers, and particularly how, like, affordability of housing in a given area or accessibility by public transit might factor into, like, what types of businesses are able to open, particularly in suburban areas.

 

>> Lindsay Relihan: Yeah, so, sorry, my voice is struggling a little bit today. I think that this is the concern that I was trying to speak to right at the conclusion. That the kinds of people who normally work in the downtown retail corridors might not be able to follow those kinds of retail jobs.

And so what you might see is that the retailers who are able to move to the suburbs are those who can afford to pay higher wages to workers who have access to cars, who have access to housing in those locations. And so the kinds of retailers who can pay those higher labor costs are likely to be the ones who can get there.

And I think when you're talking about, like, a small sort of resort town like that, that larger company can really internalize that need and build housing for workers. When you're talking about sort of the suburbs of New York or the suburbs of San Francisco, that seems less likely that they would build actual housing for those workers.

But I think this would give extra motivation for some of the other discussions we're having in housing to increase the housing supply of the suburbs, do things like reducing single family zoning, doing things like figuring out how to improve suburban public transportation, like bus schedules, and that would be, I think, a greater call to do that kind of thing here.

Yeah, sorry.

>> Speaker 4: So I had three questions. The first one is more about the conceptual question. So I was thinking about how I can think about your results as defects of work from home or defects of COVID. COVID can be contagious, so people can now kind of don't want to travel too much, to go out, something like that.

Because in the CPS there is the work supplement data and food supplement data. And by linking these two, we can see actually work from home workers' consumption pattern. And that was in 2004, but actually work from home workers have a higher consumption share of spending. So I was thinking about that kind of data result, one time I kind of looked into.

And the second question I had was not only the establishment growth, maybe you already look into this. I was curious about entry and exit, maybe separately, because especially COVID can make establishments being devastated. The last one is more about the amenity channel. So sometimes when I think about the amenity channel, the elasticity of substitution across differentiated goods really matter.

Because if that is really low, then people have more value on differentiated goods, and then you can drive more establishments coming in. So depending on the spatial variation of where people leave with the establishments, I was curious about how you think about this variation, heterogeneity.

>> Lindsay Relihan: Okay, let me make sure that I check off all three, so is it COVID versus work from home?

So we tried to formalize a little bit that with the regression by using these sort of pre-pandemic measures of the work from home exposure. Other people have used that as instruments. I think they're pretty convincing that it's capturing this differential impact driven by work from home, rather than sort of a short term COVID effect.

The third question was about amenities, not a substitution, I think that's super interesting, requires a much fuller model. Look for the next paper, which I hope we'll do that. And then the second-

>> Speaker 4: Entry, exit.

>> Lindsay Relihan: Yeah, we can easily do some of these by entry, exit. We've never, I think, done that, but that's a low hanging fruit that we could add as an appendix figure, for sure, yeah.

 

>> Speaker 5: Such an interesting paper. So this is probably not in your data, but do you predict that the retail space is going to become more mixed use, is what they say, isn't it, about central business districts in terms of office space versus residential? And you can sort of see some trends, can't you, with retail stores have already been experimenting with this.

Clothes stores now sell other bits and pieces. Do you have any view on that? That the nature, not just the location of retail, but that the nature and the design of retail is going to change as a result because retailers have to get the footfall and so they have to offer more things?

 

>> Lindsay Relihan: Yeah, I think that actually speaks to my other line of research was a lot about online retail specifically. And sort of both shocks, probably, moving retailers to be more non-tradable, to be more service-like, to have a product that's not easily replicable by being online or being at home.

And so if work from home is both, like driving people to do more leisure, and online retail is also driving people to do more leisure, like, that's the way to get people out the door to your store is to be more of an experience. I think your question also speaks to a disagreement I feel like I have with Arpit about whether or not we're gonna get caught in an urban doom loop.

So I think that downtowns should be really focused on being places that people want to live and consume services as a way of mitigating this crisis. And I think it's really possible that you could see downtowns become more about both amenity provision and residential locations as a way to continue to grow.

I think I'm gonna, okay, yeah, I think he had his hand up for a long time, so I was gonna ask him first, yeah, hey.

>> Speaker 6: Nice paper. Correct me if I'm wrong, but it seemed to me that you focused on the extensive margin, not so much on the intensive margin.

My impression is that a lot of this relocation of demand happened by existing firms selling less in downtowns, and existing firms selling more in the suburbs without necessarily entering or exiting. So that's question one. Question two. I was really puzzled by Figure 2, because Figure 2 shows that there was a much bigger decline of establishments in central locations before COVID than after COVID.

So that's something I didn't quite understand.

>> Lindsay Relihan: So I think establishments were growing everywhere in the city. Maybe we're conflating the population versus the establishments. Population was already declining downtown before COVID.

>> Speaker 6: Yeah, we can talk about Figure 2 later.

>> Lindsay Relihan: We'll talk about Figure 2 later. Yeah, we decided for this project that we really wanted to do the extensive margin of store entry and exit because we feel like that's an interesting sort of untold story here.

And we're interested in using this sort of high frequency panel and a lot of different applications to understand where stores are located. But understanding that intensive margin could also be interesting in its own right. It's just something we didn't do in this paper, yeah.

>> Speaker 7: Thanks, really interesting paper.

Just two quick questions. So I'm interested in if you can distinguish between brick and mortar businesses versus kind of online startups, right? So are you just selling online? And so is this actually a real store that's coming in versus leaving, or you switch from being a brick and mortar to only online?

And then also maybe the shift towards durables, where those types of things might, big shift in durable consumption, need more space, potentially pushing.

>> Lindsay Relihan: So two things, so we see whether or not the card was present at the transaction. So what we're really measuring here are stores where physical in-person transactions are taking place.

Or have stopped taking place. So on the margins where we might have some mismeasurement is if one small restaurant decided to become a ghost kitchen, right? And so a physical kitchen with employment still takes place downtown, but it just never makes an in-person transaction. We view that as an exit when it might not happen in reality.

But we still view that as sort of a marginal, probably very, like, obscure edge case. So we think we're doing a really good job of tracking whether or not a store is actually open for business for retail customers. The second question was about durable. So I think one interesting thing that might be creating some of these patterns across different kinds of products is the fact that the least durable intensive establishments are likely the ones who can move the fastest in the sort of like, short to your window.

And some of these home goods providers, right, where you need big warehouse spaces and you need showrooms or stuff, they might be the ones who might be the slowest to respond, absolutely.

>> Jose: Thanks, so the next paper in the session is by Jesse Matheson, who is gonna tell us about crime and the rise of working from home.

 

>> Jesse Matheson: Great, well, thank you very much for having me. It's fantastic to be among so many people with similar research interests here. So this is a paper that I've been working on for a few months now with Brendan McConnell, James Rockey and Argyris Salakis. Okay, so I mean, we, there's lots of evidence that working from home, I don't need to convince anybody in this room that it's really increased since the pandemic.

To give you an idea of what this looks like in the UK. So if we look at the quarterly labor force survey, about four times as many people say that working from home is part of their normal job than before the pandemic in 2019. Similarly, we can look at the UK survey of working arrangements and attitudes, and we find that the average worker is reporting that employer planned working from home is going to be about two days a week moving forward.

So these are changes. They've been big changes that have happened, and they seem to be really sticking at this moment. So now this has led to changes in other sorts of economic activity. So by changing where people are spending their time, we're seeing spillovers into other parts of the economy, like we just saw in the last presentation.

So in local service spending, we also see evidence in real estate demand. So I'm going to look at a different part of the economy here. So I'm gonna look at the criminal activity that happens within an urban setting. And in this paper, I'm gonna provide the first evidence on how the rise in working from home has changed post-pandemic criminal activity.

And I'm gonna focus specifically on burglaries when I look at this. Okay, now, I mean, working from home is changing lots of things in the economy, so why are burglaries of particular interest? And I'll give you three reasons why. So the first is, the decrease I'm gonna show you is big enough that this is of first order importance.

So this decrease is 30% drop relative to 2019. And that, in and of itself, begs some explanation. If this were due to a police intervention, for example, everybody would be very excited about how successful that is.

>> Jesse Matheson: So now, also, studying this change should be shedding some light on how working from home is going to affect urban activity in general.

We should understand all margins of that. So this is just another type of economic activity that's happening within the city. That is criminal activity, and that's an important part of how our cities form. And finally, so, because the change in working from home is not similar across all neighborhoods, neither will the decrease in burglary.

So this decrease is not going to be evenly realized. So we wanna ask, will this exacerbate urban inequalities that are already existing?

>> Jesse Matheson: Okay, so here's the figure that kind of gives you the idea of what I'm going to be looking at here. So the dark lines, so this is showing from 2017 up until the end of 2022.

So the dark line in this figure is showing us what happened with burglaries. So we can see as we go along here. So we hit the first shaded line. This is the start of the first national pandemic, or, sorry, the first national lockdown. So this is when all citizens in the UK were required to spend time in their homes and only work on site if you were part of a few selected jobs.

Now, the second shaded line here is going to be when the national lockdowns officially ended. Now I'm gonna refer to this period here as being the lockdown period. Now, officially, there was a reopening during this period, but for our purposes, that's just going to be the lockdown period.

The post lockdown period is going to be everything after the last lockdown ended and the pre-lockdown period, everything before. Now, what I want you to take away from this figure is, first of all, there was a big drop in burglaries. So a 30% drop in burglaries. So this equates to about 635 million in revenue if we just think about the property that was stolen from these locations.

So that doesn't include probably about a million hours in police time and any of the non-pecuniary damage that was done to the victims here. The other thing I'd like you to take from this, is notice that unlike other property crimes, which I have here. So burglaries have remained stubbornly low, so they're not showing any signs of recovery.

But the other property crimes are going back towards their pre-pandemic means.

>> Jesse Matheson: Okay, so here's the empirical strategy we're gonna use. So, for almost 7,000 neighborhoods in England and Wales. So this is almost every single neighborhood, we don't have data for the Manchester area, it's the only missing part.

We're going to construct a neighborhood level measure of working from home potential, based on pre-pandemic residential occupational distributions. This is gonna be combined with monthly crime data and data reflecting other neighborhood characteristics. And then we're also going to supplement this with crime data provided by the Metropolitan Police Force.

So this is the London police force, which is going to give us details about exactly what time of day crimes were committed. So the baseline strategy is basically going to ask, does a neighborhood's change in crime vary systematically with its work from home potential? And we find out the answer to this is yes.

So a neighborhood work from home potential. So, evaluated at the sample average of about 36.5%, that means about 36.5% of all work can be done from home in a neighborhood, leads to about a 15% decrease in burglary rates relative to 2019 levels. So just based on our estimations here, we can explain about half of the drop in burglaries.

Event study analysis is going to show that this is a very stable relationship over time. And the relationship is going to be driven entirely by burglaries committed during the weekdays, during working hours. So we see no action outside of those hours. So we're really contributing to two strands of literature here.

So the first is kind of the economic and societal consequences of the rise of working from home. And then there's also going to be some information here about the economics of crime and criminal decision making. Okay, so I'm gonna start by just going through a conceptual model, which is gonna help us frame our thinking around this.

Of course, I'm going to keep things fairly high level, though. So the model basically can be thought of as modeling burglars based on the actual behavior that we see. So this is a quote from the Metropolitan Police Force, which is basically saying that burglars, one of the key things about burglars is they don't want to be seen doing what they're doing.

That is, they're going to avoid houses that have people in them, and they're going to avoid going into busy communities where somebody might see what they're doing and move on to other areas. So we propose a spatial model of criminal search. So we're gonna have two neighborhoods here which experience an asymmetric increase in working from home.

So one neighborhood experiences a larger increase in working from home than another. And each potential criminal must decide whether or not they want to engage in burglary or take some outside option and how to allocate their time searching across the different neighborhoods. Now, working from home enters this model through two channels.

The first is, it just reduces the number of empty houses in a neighborhood. So in a house with somebody in it is not a suitable target. The second is conditional on finding a suitable target. Working from home increases the probability of getting caught, because it increases the probability that somebody's gonna see you engaging in this activity.

And then we can solve for the equilibrium of this model, where we have an equilibrium where no burglar is gonna prefer their outside option. So it's incentive compatible. And any burglar that takes the outside option would not rather be a burglar. And then there's going to be a spatial equilibrium where the expected return to searching in each neighborhood is going to be equalized.

Now, given this model, if we increase working from home, and it increases differentially across the two neighborhoods, so we find three results. So, first is, it reduces the total number of burglars. The second result, and this is the really important one, because this is the one we can observe in the data, is it reduces the number of observed burglaries.

That is, burglaries conditional on somebody even finding an empty house, not just criminals. And that reduction is going to be greater in higher work from home neighborhoods. And then finally, under some reasonable assumptions, it's going to shift criminal search away from high work from home neighborhoods towards lower work from home areas.

Okay. Okay, so our measure from work from home. So I'm going to spend a bit of time on this. This is going to be calculated using methodology that we outlined in an earlier paper. And the basic idea is this. So the work from home potential is going to be a function of the number of people in each neighborhood in a given occupation times how much work in that occupation can be done from home.

So a Dingel and Neiman type measure adopted for UK SOC codes. And we're going to use four digit SOC codes to calculate this. And then the number of the amount of work from home is just the weighted value of that index divided by the total number of people who are employed in the neighborhood.

And like I said, we're going to calculate this for 6837 neighborhoods. So that's everywhere except Manchester in the UK. Now, a neighborhood is defined by what we call a middle super output area, which reflects approximately 3500 residential properties. Okay, so to give you an idea of what the spatial distribution of working from home looks like.

So I've got two police force areas here. So the first one is the Metropolitan Police Force Area, this is London. And the second one is the Merseyside Police Force Area, in other words, Liverpool. Now, just to the shading in these two figures. So the bins are selected based on the figure itself.

So a dark area within Liverpool is going to be a different rate of working from home than a dark area in London. But basically, working from home varies between about 20% in a neighborhood to a maximum of about 74% in the City of London. So it's this variation that is going to allow us to identify the effect on burglaries.

Okay, so our core crime data comes from police.uk. So one of the great things about the UK is we have great access to crime data. So this is publicly available data and it gives us the offense type by street by month level data for all police force areas, excluding Manchester.

We're also going to be using restricted access data from the London Met. So this is going to have all that information, but it's also going to give us the time and the date of the offense. And then of course we're going to also include a number of neighborhood covariates, which I won't go into too much detail about here.

Okay, so here's our basic estimation strategy. So it's a difference in difference strategy. So I'm going to interact our work from home potential measure with a lockdown dummy for the lockdown period and a post lockdown dummy for the post-lockdown period. Okay, so I'm going to jump to the results now.

So here's our main table of results and our preferred specifications. We're playing a little bit around here with the fixed effects and the inclusion of control variables. And the results are fairly stable when we move across these different estimations, and our preferred specification here on the column 4.

So the way to interpret this would be a neighborhood work from home potential of about 36.5% leads to a 15.2% decrease in burglaries in the post- lockdown period as opposed to relative to pre-lockdown rates. And that's very similar for the lockdown period. So just as we saw in that initial figure, there's actually not a huge difference between the relationship of working from home potential and what happened to burglaries, whether we're in the lockdown or the post lockdown period.

And here's an event study graph. A couple things to take from this first. I mean, if we look in the pre-period, it's not really clear that there's any systematic pre-trends happening here. And we just see a level shift happen as soon as we enter into the first lockdown.

And it really doesn't recover from that level shiftdown. So these results are really quite stable over time. And so here's just the results for London. And what I want to show you here is, so we've got working hours. So this is burglaries that happen during working hours. And here's non-working hours.

So this is outside of the hours of 8 till 6 and on weekends. And really, all of the action is happening during the working hours. And here's a figure which just shows. So this is during the lockdown period and in the post lockdown period. And we've got here working hours in these two figures and then outside of working hours in the other figures.

Now, there's a little bit, it looks like during the lockdown period there may have been a little bit of a shift towards non-working hours away from working hours, but that completely disappears when we look at the post lockdown period. And there's no clear differences across days of the week here that we're observing.

Okay, so we do a number of robustness checks, looking at just specification issues with this, and we have a number of additional results, which I'm not going to go through in too much detail here. One thing I would like to pay a bit of attention to, though, or bring your attention to, is what happens in a housing price analysis that we look at.

So I'm just going to jump to that. So we're interested in trying to get at some of the welfare considerations of what happened here. So what's the value to these reductions in burglary? And one way we might be able to get at it is by looking at, well, what happened to housing prices across these neighborhoods.

So we estimate a hedonic housing price model in which we interact work from home potential with these time dummies, and then also include what we call the neighborhood's ex-ante burglary risk. So that is what was the risk before the pandemic of a house being burgled in that neighborhood.

And then we're going to allow these coefficients to vary across different property types. So basically, we're asking did housing price change with work from home potential in the post-pandemic period? And does ex-ante burglary risk matter for these changes? And here's what we find. So we have detached properties here, semi-detached or duplexes here, and then terraced houses here.

And apartments on the end. And what we can see, so a couple of things. So, first of all, I mean, work from home seems to matter. Now, the entire level shift, I wouldn't want to attribute that at all to burglaries. There could be a lot else going on that's correlated with work from home.

So this is just ex-ante burglary risk on the x-axis. So a higher value means that higher risk of burglaries. And we can see that that relationship is monotonically increasing for all of these properties, except for apartment buildings. So in apartment buildings, it doesn't appear to matter at all, okay?

And now, one theory would be that's because, well, if you're in an apartment building, you're just less exposed to the risk of burglary to begin with, whereas a detached house would be much more exposed to burglary risk. Okay, so I'm going to leave it there. So work from home potential does appear to really matter for these changes that we're observing in burglary.

In fact, we can explain about half of the total 30% decrease in burglary, and this is just through looking at variation within a given police force area, okay? And this result has been fairly persistent over time, right? Yes, I will stop there.

>> Arpit Gupta: Great paper, I'm wondering about the possible violations of spillovers and satval in this context, because you obviously have more people in the work from home areas that are coming as a loss from the areas that they previously showed up to.

So the crime effects could be driven both by a combination of less crime in the affected areas and more crime in the kind of downtown kind of commuting core. So one way to address that might be to have a control sample that's not all other areas, but areas that experience no change in occupancy, right, to kind of establish a consistent benchmark.

I think these effects are really important if you think about downtown areas, like San Francisco, right? They seem to be experiencing more crime, because they're losing the occupancy and the eyes on the street effect.

>> Jesse Matheson: Yeah, that's a really good point. So we do look at spillover effects, too.

So one of the results that we find for spillover effects is if you're in a high work from home potential neighborhood, but that neighborhood is surrounded by other, even higher work from home potential neighborhoods. Then we do estimate a positive spillover effect that, in fact, completely offsets, well, in the worst case scenario, completely offsets the negative effect of working from home.

Yeah, so this is something that I mean, in a main specification, we really, I mean, we want to think about, how do we parcel that out in a sense? And yeah, your suggestion is definitely a good idea.

>> Speaker 10: So two thoughts, a comment and a question. So the comment is, probably, burglars are probably the only occupation category who cannot work from home.

 

>> Speaker 10: So I was thinking about that, and my question is similar to the previous question. So I was thinking about the negative spillover to vacant commercial establishments, but also I was thinking about spillovers to other forms of crime. We saw an uptick in the aggregate trends towards other forms of crime, and I'm wondering if your data allows you to say something about that.

 

>> Jesse Matheson: Yeah, we're talking a little bit about this earlier. So we have looked at, I mean, these other property crimes. So that would be the most obvious place for a burglar to go is into some other form of property crime. But there's not anything obviously coming out of that.

Now, one of the problems is, so the methodology that we're using here is really well suited for burglaries, because burglaries have a very fixed location, whereas something like even auto-theft is very difficult to start to pin down using the methodology that we've used. Now, we're not seeing any crimes that have seen an uptake over and above their pre-pandemic levels, which might be what you would expect to see if there's going to be a shift over.

Now, the one exception that we haven't really looked at in too much detail with is drug-related crimes. We're not completely happy with the publicly available data, but, I mean, if anybody's listening from a police force and has more detailed data for us, we'd love to look at that.

 

>> Speaker 11: I had a question specifically, when you talk about burglary, I imagine in my head that means burglary of people's houses, breaking into someone's home or apartment. There's other types of crime that is picking up in the news these days, which is specifically around either, in San Francisco, we call them porch pirates, where people come along and steal the increasing volume of packages that get stolen.

So that's an increase in residential neighborhoods when people are working at home. There's also this kind of coordinated flash mob style break-ins to stores, commercial stores and car break-ins. I'm guessing those are not part of you're thinking.

>> You're talking about home.

>> Jesse Matheson: That's right, yes. Yeah, this is home burglaries.

We have looked at commercial burglary. So we have commercial burglary data for London, and we're just not seeing anything really obvious that's been going on there. We did expect to see a spillover into commercial burglary, cuz there's potentially lots of empty retail spaces now. We're not seeing anything really obvious there.

As far as the grabbing an Amazon package on somebody's doorstep goes, that's probably not gonna show up in our data, cuz it just wouldn't be high value enough to be reporting to the police for insurance purposes, whereas a standard burglary would. Now, what this will include is if somebody breaks into your detached shed or something like that, so anything that happens within your property area, even if it's not in your home specifically, will be showing up in this data.

 

>> Steven Davis: I wanna go back to the spatial spillovers. So I'm thinking about it correctly since your main line specification here doesn't have the spatial spillover, so you're overstating the effect because you're not capturing the fact that if I have high work from home rates in my neighborhood, some of it spills over to other neighborhoods.

So there's many ways to get the spillover effects, but a simple way that would allow you to control for that effect is to construct an index of the work from home intensity in the nearby neighborhoods around the neighborhood of interest. That controlling, using both those coefficients then and that kind of specifications might be expanded, I think was a better way to get a handle on what the overall impact on aggregate impact is on burglaries than the statement you've got here.

But in addition, it speaks to another issue, which is how mobile are criminals? They live somewhere, too. It's not just about equalizing the expected returns across neighborhoods. They have to factor in the cost of getting to the neighborhood that they wanna burgle. The work from home phenomenon alters that calculus for them.

Separate point on the spillovers, other types of crime. The one that, at least in the US context you would think is potentially important is robberies. But of course, the UK, guns are not so available. But in the US, they're quite readily available, so you can rob people at gunpoint.

Not that I would advise that, but you can.

>> Jesse Matheson: But you could.

>> Steven Davis: But it happens a lot in the US. And so I'd be really keen to see a similar analysis in the US where you see whether the spillover of robberies is actually happening in a big way as criminals are displaced out of burglary.

 

>> Jesse Matheson: Yeah, that's a really good point. And I do get the impression that robberies are more of a problem in the US than they are in the UK. And it could just because of access to weapons. Thank you very much for the suggestion about spillovers, we do have in our spillover specification.

So we do look at contiguous neighborhoods and we're looking at relative work from home rights. Because of course it needs to be not just the work from home rate in the next neighborhood, but relative to the size of that neighborhoods. But yeah, it could be as simple as just including for all contiguous neighborhoods, the work from home rights that we see there.

 

>> Speaker 13: So really nice paper. I thought that the results by specific property types were sort of a roundabout way of controlling for density. And I was thinking, why not just directly measure the work from home density, rather than work from home share? Cuz right now what you're doing is you're comparing a rural area where 20% are work from home with a place in London where 20% are home.

But you could just more directly say, okay, I want the number of work from home per square kilometer in this case, right?

>> Jesse Matheson: That's a really good idea. And our results are much stronger in rural areas, which makes sense, because, yeah, there's less, well, eyes on the street, so to say, yeah.

 

>> Speaker 14: So this may be a minor point, but it seems that the work from home effect should be smaller in neighborhoods with lower female labor force participation. Because the women are at home, and so the work from home doesn't necessarily deter the burglars because there's already somebody there.

 

>> Jesse Matheson: Yeah, absolutely, I mean, this is something we should be factoring in here, is thinking about. So what did work? Well, what did people at home look like in these neighborhoods before the pandemic? Because it could be that. Yeah, that could be correlated with work from home potential, for sure.

 

>> Speaker 15: I had two questions. One, are you able to get data on apprehensions for these burglaries? Can we show that, I would imagine equilibrium, you would still see an increase in apprehensions if we're making the cost higher. And then second question is, I found it interesting, no, sorry, I have three questions.

Second question. I found it surprising that there are enough burglaries during the daytime to have made such a huge change. Were you surprised by that? Does that vary a lot from one place to another with the daytime versus nighttime.

>> Jesse Matheson: I can't say I really had a prior other than my own experience.

Any burglary I've experienced has actually been during the daytime. So I guess it sort of fit with what I went into this. But about 50% of burglaries would be during the daytime, actually, yeah. And it's just because of empty houses.

>> Speaker 15: And then finally, what do you, what would you guess can explain the other half of the variation that you see?

You stressed that you're kind of within a neighborhood. I wondered, do you think there's cross neighborhood variation going on that you're not able to observe? And have you thought about the story for it?

>> Jesse Matheson: Yeah, absolutely. I mean, I wasn't super clear about this when I went through the estimating equation, but our variation is really based on deviations from time trends within a local police force.

So it's actually a very specific type of variation we're estimating this on. So it could be that if we were able to be more generous with the variation we were using, that would pick up a lot of the other half, the change.

>> Jose: So I think we need to move onto the next paper.

 

>> Jesse Matheson: Thank you very much.

>> Jose: All right, final paper in the session is presented by Simon Krause about the future of work and consumption in cities in Germany.

>> Simon Krause: Thank you, Jose, for the nice introduction, and thank you to the organizers for having us on the program. We're really delighted to be here.

So this project, The Future of Work and Consumption in Cities After the Pandemic: Evidence from Germany, is co-authored with Victor Alipour, who is with us in the audience today. Oliver Falck, Carla Krolage, and Sebastian Wichert. As you all know, the COVID-19 pandemic has disrupted the organization of work in various ways, but in particular by inducing a lasting increase in working from home.

We have several experts here in the audience running great surveys for the United States, but also around the world. And I brought to you a chart from Germany, which basically resembles many of the patterns that you're seeing around the world as well. So this is a chart on the share of work from home employees in Germany over time, employees that work from home, at least partially.

And we see that the work from home share has increased from a prepandemic value of about 5%, which is in the same ballpark as in the United States. With the onset of the pandemic marked by the red line here, to about 34%, then we had two lockdowns in Germany.

So one in 2020, another one at the end of 2020, around winter 2020-2021. But the interesting thing that we see here is that work from home in Germany, as in other countries like the United States, is stabilizing at a much higher levels. So I hope you can see here the graphics here on that projector are not as great as in slides.

So obviously the green shaded areas here. Since April 2022, there were no more pandemic restrictions in place in Germany, and you basically see a stabilization of the work from home rate at about a quarter of employees who work from home partially. And we also elicited the post COVID plans of employees and employers in Germany for the near future, which also gives us a value of about 24%.

So we do see a very durable shock to labor markets. And the arguments that we're going to focus on in our papers that work from home carries the potential to alter the geography of economic activity in cities. There's a great and growing body of research here. And we have many of the contributors to this field here who show that work from home changes not only the place where we work, but also where we live and where we spend.

So what I will do, I'll refrain a bit from summarizing the literature here. But just focusing on what we try to contribute to this literature is to provide new causal evidence on the impact of work from home on the geography of consumer spending in cities. In particular, we are focusing on offline spending at retail locations in post pandemic urban agglomerations.

So our data includes, at a very fine spatial level of AIP codes, data on work from home prevalence, as well as consumer spending in major German cities between 2019 and currently until the end of March 2023. That will be expanded soon as well. And what we do in this paper is we try to make two contributions here.

So the first one is that we identify changes in the spatial distributions of consumer spending. We show, firstly, that offline consumption shifts to areas with previously lower consumption. This is a story about the doughnut effect, where consumer spending shifts from the city center to residential areas, to the suburbs to the outskirts of cities.

But we also see a sustained shift to online spending in Germany. The second contribution that we want to make, and this is the main part of this paper, is we established a causal link between work from home and consumer spending. So you've seen the chart. We know it around the world that there has been sizable work from home growth during the pandemic.

And what we exploit for identification is that there were spatial differences in terms of work from home growth due to the untapped work from home potential before the pandemic. So in our data, we define, or we have a measure of the work from home potential that was already determined by the jobs that people were doing at their place of residence before the pandemic.

And we show that this high untapped work from home potential not only predicted work from home growth. But, and this is the main result, we see that for one standard deviation, higher untapped work from home potential before the pandemic, that local spending increases by 3 to 4%. And this is not, these 3 to 4% spending increases are not something that we only see during the pandemic.

But actually, we see it right until the end of our timeframe in March 2023. And based on the durability of these findings and having people here in the audience and the conference organizers who argue that work from home will stick, we also cautiously project that the effects will likely persist thanks to the continuously high work from home uptake.

Now, to give you an overview of the setting and the data. So, here on the left, we have a map of Germany. And in our sample, there are five major metropolitan areas in Germany, including the cities of Berlin, Hamburg, Munich, Stuttgart, and Dresden. These five metropolitan areas comprise about 17% of total population in Germany, and 810 ZIP codes.

The ZIP codes is the area of spatial analysis that we're using. And we have data from Mastercard on consumer spending at the daily level for those ZIP codes. So the spending data, let me elaborate on that for a second. Includes both debit and credit card transactions from residents.

So, in Germany, there are more people using debit cards, for example, than in the United States. So we have data as well in our data. And the data for offline consumption is daily anonymized and aggregated. We see it for total consumer spending per day in a given postcode, but also for individual sectors.

In addition to that, we also have information on online spending. But the online spending is not linked to the geographic location. So it's linked to the metropolitan area, but not to the ZIP code, that's what I wanted to say. The timeframe for the data is from January 2019 and runs through March 2023.

In addition, we have a very nice dataset on work from home and area data that we collected from a company that focuses on spatial data analysis, infas 360. And they ran a representative survey for us on work from home patterns and collected a lot of information at the ZIP code level for these five cities.

In addition to that, we also have, from them, but also other sources, wide range of indicators on population, settlement, business, and land use characteristics. So, to start with a stylized fact of what's going on, I brought you here a chart that shows the changes in spending volume by the pre COVID consumption intensity.

What we do here is we categorized the ZIP codes in those five metropolitan areas by their 2019 consumption average. And then we normalized it to an index that takes a value of 100 for the year. And what you see is that in 2019, these lines pretty much co-evolved, they were roughly parallel.

But then with the onset of the pandemic in 2020, we do see a clear divergence between the red line, which are the low pre-COVID consumption intensity areas. So those areas are mostly residential, low density areas, low density of inhabitants, but also low density of businesses. And we see that these areas throughout the pandemic, but also now in the open period, without any pandemic restrictions at all in the last year, see a huge increase.

And actually, the gap widens between them and the other areas. So they're currently 40% above the pre-pandemic level in terms of spending. Whereas, if you look at the blue line here, these are the high pre-COVID consumption intensity areas in the city centers. Those areas are marked by high population density, high business density, and actually, the low spending level is still at the end of our timeframe, 4% below the pre pandemic average.

We see that finding if we look into spending segments for a second, we see these findings across the board. So, for example, we see it here in panel b. I hope you can see it is for grocery and food stores. Same result, as well as for eating places and also for apparel.

So the consumption changes are pretty broad based, and they also, when it comes to the reduction and consumer spending in the city centers, they resemble the doughnut effect that Nick and others have shown. So, we brought it to you here for Berlin. So if we just run a regression or just compare the growth rates and consumer spending for certain distance rings, this is for the city of Berlin here in blue, a 2 to 10 kilometer ring.

And gray the 10 to 30 kilometer ring, and in red the greater than 30 kilometer ring to the urban center. We do see that all of these values are negative, meaning that the growth rates in the city center was lower than in the outskirts of the city. So we do see a doughnut effect in consumer spending for these cities.

This is also mirrored, for example by a lower pedestrian frequency in the high streets in the city centers than before the pandemic. If we take a more structural or a bit more of an econometric approach to that and run a regression here using a full variation of the 2019 consumption intensity variable that is a standardized variable with mean zero and unitary standard deviation.

We see that by the end of our time period here in the post pandemic period that we're focusing on a one standard deviation higher consumption intensity before the pandemic. So the closer you get to the city center kind of is associated with 15% less consumer spending. We characterize these consumption intensive areas that are losing out with a whole lot of indicators, socioeconomic indicators, population structure area indicators, industry composition.

And what we see here are some really interesting stylus results. For the sake of time, I cannot get into it, but the aspect that is most striking for us is, while these socioeconomic indicators, population structure area indicators, and so on, they are pretty fixed in the short term.

But what is changing is work from home growth. And what we see in terms of work from home growth, that is quite interesting. So, high consumption intensive areas that see the spending declines are correlated with lower work from home growth. In other words, the areas which have a lower consumption-intensity, and I've seen more consumer spending growth, also saw a higher work from home growth.

And this is exactly what I would like to focus on now for the remainder of this paper. So what we're really after here is to link work from home to the regional shifts in offline consumer spending in the cities. And obviously, this is a very challenging process, because work from home uptake is likely correlated with other sources of spending disruption during COVID and so we proposed two solutions to address that.

Solution number one is that we want to estimate intention to treat effects using a measure that we call untapped work from home potential. Untapped work from home potential is the local share of residents with a teleworkable job, but who did not work from home. This measure is calculated at the zip code level based on the job distributions and comes from our survey data.

And what we see in the left panel here, panel A is quite interesting. So, before the pandemic, we had, on average, an untapped work from home potential of 60%. Just to make sense of that number, if we think about zip codes that had 50% teleworkable jobs before the pandemic.

That number of 60% untapped work from home potential means that only 20% of people in that zip code also work from home, whereas the other 40 percentage points that could have worked from home did not. But then what happened during COVID is that the untapped work from home potentially got exploited, and that's also what we're showing here in panel B.

So untapped work from home potential is a really good predictor for work from home growth. Now, you might still be worried that the changes in consumer spending due to work from home might also be influenced by some other supply side disruptions, for example, business closures. Also structural factors, for example, population structure that influences behavioral adjustments to the pandemic and also the current situation.

And so what we're doing here and proposing as a second solution is to control for these supply side and structural factors that might be correlated with untapped work from home potential as well as the time trend. So we are estimating a regression here where we have log consumer spending in a given zip code and month as our dependent variable.

And our independent variable that we're interested in is the untapped work from home potential in that zip code. It's a standardized mean zero unitary standard deviation variable per zip code. And we interact that with time trends as well as we interact the supply side and structural controls. So what we have in here are time invariant controls for, for example, the consumption intensity, local population, including density, age structure, education, income, furniture, but also business structure and land use regulation in these postcodes.

And this is the main result. So what we find here are differential spending effects of untapped work from home potential. What you see in both charts. So, panel A are our estimates for weekdays, panel B for Saturdays of the untapped work from home potential times month dummies. Interestingly, before the pandemic, up until February 2020, we see that these effects were insignificant and have evolved in parallel.

But then with the onset of the pandemic, we see that for weekdays, there is a jump in the effect. And we are kind of skipping this part here where we still had lockdowns and just want to focus on the last 12 months of our data up until March 2023, when there were no more pandemic restrictions in place.

And there we find a very interesting effect of a significant effect of about three to 4% higher spending in a given postcode for a one standard deviation higher untapped work from home potential. Interestingly, on Saturdays, and very fittingly with our work from home story, is that these effects are insignificant on Saturdays.

In terms of the effect differences, weekdays versus Saturdays, they're statistically significant. So the weekday is statistically significantly different from the Saturday effect. We also have a broader table here of the regression results of the intention to treat effect. And what I would like you to focus on is here the line of the post COVID period where we see the significant effects for the weekdays, insignificant effects.

Sorry, significant effect for all days and the weekdays, insignificant effects for Saturdays. And then when we add more controls to our analyses, the interesting thing is that our effects for the post COVID period and also the last open period before, they remain very strongly significant. Verse for the other periods before where when there were lockdowns in place, the effects become insignificant.

Potential explanations for that are, for example, the business closures during lockdown periods that induce a shift to online spending, as well as short term work schemes that were in place in Germany. And that reduced the correlation between untapped work from home potential and consumption. Because I mentioned the shift to online that was particularly prevalent during the lockdowns, you can also see that here quite nicely.

So there was the first lockdown in early 2020, and then the next one in the winter of 2021. And you really see kind of like a mechanical increase in the share of online spending that has now come back to a lower level, still a bit elevated compared to the pre-pandemic level, but it's kind of gone back to its long-term growth trend.

And just to add to that, so if we look at the full distribution, and if you move from the 25th to the 75th percentile in the distribution of untapped work from home potential, data is associated with a 26% increase in distance to the city center. So work from home is kind of fostering, sorry, it's increased associated with 26% increase in distance to the city center and causes a 15% increase in local spending.

So in that sense, work from home contributes to the doughnut effect in consumer spending. Additional mechanisms for the work from home induced relocation of consumer spending are also increased migration to the suburbs. So people moving out of the city centers to the periphery, as well as a stop in the urbanization trend of employment growth that we've seen in Germany before the pandemic.

To now a situation where employment growth in the city center anti periphery is pretty even. So this also contributes to the effect here. Now to conclude, what we find in our paper on work from home and consumer spending in cities is that. Work from home causally shifts local spending towards less spending intensive suburban residential areas that contributes to donut effect in cities.

We cautiously project that the effects will likely persist because work from home seems to be here to stay after the pandemic. So in our survey that we run for this project and that found a projected future work from home rate of about 25%, we see it in our ifo Business Survey that we're running every month.

Where we see a stable work from home rate of about 25% since April 22. And you also see it in a survey of working arrangements and attitudes in the United States, where full work from home days are pretty stable at about 25% to 30% since mid 2022. And that, of course, altogether has very important implications for the future of cities in terms of urban planning, local businesses, municipal finances.

But also ultimately the relocation of people and jobs. Final point, what's next to our project? So we're currently in the process of obtaining more and better data. So we want to measure work from home not only based on our survey, but also on cell phone mobility. And also extend the sample of consumer spending from currently five metro regions to 50 metro regions in Germany.

With that, I thank you very much for your attention and look forward to your comments and questions.

>> Arpit Gupta: Yeah, so I think your results also have implications for inflation. So think about what's happening. One set of regions is experiencing a huge demand increase. If there's a local Phillips curve that's gonna result in inflation.

Another set of regents is experiencing a large drop in spending and prices are sticky downward. So in aggregate, you can have an increase in inflation even if people are just moving spending from one location to the other. There's a similar research in macro showing this with the changeover from services to goods causing inflation over the pandemic.

So I think it's a whole paper out there. I think one thing you could do in your setting might be just to look at the effects on inflation in the kind of high spending versus low spending areas. So how much of the increase consumer spending you're finding actually show up as higher prices versus higher quantities.

 

>> Simon Krause: Thank you, Arpit, that's a great point and that also shows the wider ramifications of the developments that we see. I think that's definitely something we can put on the to do list for the version of the paper. I don't know who was next, please. I think you have to microphone.

 

>> Speaker 17: You mentioned migration at the very end. I'm wondering if you looked at the correlation between the work from home potential and migration. Because if you see that people are moving out from the high work from home potential areas, then your estimates are actually a lower bound of what actually the effect is.

 

>> Simon Krause: That's a great point. So we're currently still in the process of properly analyzing the migration data, but purely from simple correlations, it seems to be exactly that way. That we have a higher work from home potential in the city centers compared to the periphery, and we do see an outflow of people from the city center to the periphery.

So if that is the case, our numbers could be interpreted as a lower bound. Thank you.

>> Speaker 18: Hi, I wanted to ask or check my understanding. It seems that you had been using the destination of transactions to measure spending growth. But the origin of transactions, that is where people are living to measure work from home potential or untapped work from home potential.

And I was wondering that while most of spending, or a lot of spending that happens around the home, if changes in commuting might have, I don't know. Hidden some of the things that were going on in the data or caused a weird interaction between and what you're observing and what you want to observe.

 

>> Simon Krause: So thank you. Great point. We do have some data limitations here. So ideally, from an ideal point of view, it would be great if we could track individuals with their place of residence as well as the kind of destinations where they spend. Unfortunately, for data privacy reasons, Mastercard is only giving us the destinations.

For the work from home potential, we also have the data at the workplace, but I think because we're looking at the effects of the untapped work from home potential that matters where people live. And so yes, it is an implicit assumption that the destination of spending and the residents of the people are connected, but we feel somewhat confident that we can make this assumption.

But I'm happy to discuss that further with you.

>> Steven Davis: So a suggestion about how to provide more context for the episode you're looking at. And so you show us a lot of information about the relationship between work from home and the reallocation of consumer spending. Spatial reallocation. But it would also be very useful to show us about the magnitude of the spatial reallocation shift.

And a simple way to do that, you could do this separately for each metropolitan area, is compute month to month or quarter to quarter changes, absolute changes in the share of spending by zip codes. Okay, then show us how that's evolving over time, I suspect, but I haven't seen anybody actually show it.

That not only was this work from home phenomenon the key driver of the spatial reallocation of consumer spending, but that it was an enormous shock. It was enormous in an absolute sense, it was an enormous shock to the spatial allocation of spending of a sort that we've rarely seen in recent decades.

You only go back to 2019, but that's fine. Show it so you understand the time. I want to see this time series plot for each of your five pictures. I featured your five cities to get some size of how big this spatial reallocation shock was compared to the benchmark, the pre pandemic benchmark.

 

>> Simon Krause: Great suggestion. Thank you, Steve. We'll definitely do that. So I can tell you from us analyzing the absolute data, like the plot I showed to you, was normalized to compare the differences. But yes, also at the absolute magnitude is quite sizable, but still also, after pandemic in Germany, at least, maybe the effect.

Effects in the US that we see for some cities, such as San Francisco or New York City, might be stronger, seem to be stronger than what I read in your paper. So we still have the fact that most of the spending occurs in city centers, and yeah, like the residential areas have seen tremendous growth, but from a much lower level.

 

>> Simon Krause: Yeah, we'll do that.

>> Speaker 19: Thanks, really interesting. So I'm just curious, what would happen if you, say, would you repeat this exercise outside of the urban areas if you were to look at the kind of more rural areas? And the reason I ask is that in kind of small town America, there are these huge booms in terms of local sales tax revenues.

And so kind of wonder if there's any potential in the rural areas. Maybe there's not enough variation in work from home, but these are places that might not have any brick and mortar or very few. And then I kind of wonder what would happen if there's maybe smaller mom and pops, and do we notice this show?

 

>> Simon Krause: That's a great point. Unfortunately, we don't have the data for that, but I would suspect that the effects look quite different to what we see in the large cities. So what we'll definitely be able to do with the coming data update is to look at smaller metro regions, but the purely peripheral small towns.

But I'll think about what we can do there, nice.

>> Speaker 20: That was really great, there's one thing I was surprised about. Just make sure I understood it, you saw a big increase in further out suburban spending on the weekdays, but not on the weekend, right?

>> Simon Krause: So, yeah, we do see the increases on the weekend, but in terms of the absolute spending, it's quite interesting that, so did the work from home effect is pronounced on the weekends.

But if we look at absolute spending, we do also see like a reallocation of spending from the weekdays to the weekend, which I think also kind of fits

>> Speaker 20: There's two totally different things going on. One is I living in the suburbs, I just don't go into the center of town, so I go buy my coffee at Starbucks, etc, so that's a weekday effect.

The other thing is I permanently move, which you expect to see seven days a week, and it looked more like the first. And I think they're two very different, but they both give a donut effect, but one is only weekdays, the other is permanent.

>> Simon Krause: Right, so, yeah, the first one is much stronger in our data, which also fits with the preliminary analysis of the migration data that we did in Germany.

So there has not been that much migration from the city centers to the periphery yet, but it is increasing over time. So there was almost nothing in 2020, a little bit in 21, now it seems to be another chunk of migration in 2022. So that second effect over time, I think, if you look forward, could become stronger.

>> Jose: Okay, I think it's time for a break, so let's come back at 3:30.

Part 6:

>> Speaker 1: All right, so the final session of the day again on effects of the city. We'll get started with Arpit Gupta, who will be telling us about converting brown offices into green apartments.

>> Arpit Gupta: Great. Thank you so much for the kind introduction. Thanks so much for having our paper on the program.

This work is joint with Candy Martinez, who's a PhD student at Columbia, and Stayman Norberg, also at Columbia. And our paper is also a little different from others on the program because this is a normative paper as well as a positive one, in that we are advocating for more conversion activity, as well as describing the conditions under which it can be realized.

And this paper is actually gonna be out in the next few weeks through the Hamilton project at the Brookings Institute. So it's sort of explicitly commissioned and produced as a policy-oriented piece. Let me give you a little bit of the background for why we're studying this question. This is drawn from our earlier work, which we had presented here last year.

This is drawing from scoop data on back to work plans. So it's updating a little bit the specification we had earlier. And it's just sort of showing office demand at the firm level connected to firm level choices in remote work. And it's not surprising that the firm is gonna be in the office most days of the week.

Their office demand doesn't change that much, not that surprising. If you're going fully remote, you get rid of your office. Maybe more surprising is in between. So even hybrid work, which is a very dominant form of work, as we know, in the hybrid and the remote work era, entails some loss of office space demand.

And so we think this is really hitting the office market. So just this morning, for example, I was looking at the office delinquency statistics. So office loans are very quickly going more delinquent at a pretty steep pace. So as of today, about 5.6% of office loans are delinquent.

Last year, it would have been less than 2%. So there's been a big increase in defaults on these loans. And here we just sort of see that whether we look at industry, whether we look at cities, places that have more remote work, exhibit lower office demand. So we're seeing these trends in office use play out as a function of how much power mold places are in ways that imply that office demand will therefore be down substantially in a lot of cities.

So here in San Francisco, for example, we're now at like a 33.9% office vacancy rate. So one in three square feet of office space is vacant in downtown San Francisco. All right, so that was all kind of earlier work which had established the doom and gloom of office.

And so for change, we decided to be constructive about the problem and think about, okay, first, let me dwell on the doom and gloom for a bit. So we've got this lower demand for office coming from remote work. We've got lower valuations on these buildings coming from the demand shock, also coming from higher interest rates is a big effect.

Of course, these trends have really big implications for cities, which draw a lot of their revenues from these buildings. At the same time, there's not enough housing. So we've got an affordable housing shortage. And then finally we also have a climate crisis going on. So in New York City alone, buildings are actually responsible for 70% of the carbon emissions in New York City.

So we've got this stranded asset problem in multiple dimensions. These assets are stranded because there's less demand for them and also because they're polluting. In New York City specifically, we have a carbon tax that's coming in, which will mean steadily increasing carbon taxes for these buildings. All of these mean the assets are less valuable, less usable for their originally intended purpose, and raises the question of what we're gonna do with all this space.

So our proposal is to turn the offices, particularly turn the dirty brown offices that are environmentally polluting, into clean, multifamily buildings. And to do so, we're gonna do three things. First, we're gonna identify whether this is physically possible or not. Then we're gonna analyze whether it's financially feasible to do.

And then we're going to explicitly analyze the policy levers at the local and state and federal levels which can accelerate this conversion activity. So more of it happens. So this is basically the feasibility part of it. And this is across the country where we've identified candidates for conversion from dirty offices to green apartments.

I'll go through the criteria that go into this, but it's basically here just to illustrate. I think there's a lot of scope across the country to convert office buildings all over the nation. So not just San Francisco, New York, but also possibly other cities as well. Here's what it looks like just in New York City.

So on the bottom there in the financial district, we've actually been converting buildings for several decades, and we've had even local public policy programs intended to encourage the conversion of those office buildings. They have a particular typology that actually makes them very suitable for conversion into multifamily. But you can see from the dots here, we think there's still some more room for conversions in that sort of downtown financial district area, as well as in Midtown, particularly East Midtown, as a lot of these very low value office properties that would be suitable for conversion to multifamily.

Okay, so let's first identify these candidates. So just a little bit of descriptive statistics here. So in New York City, we're actually getting through local law 97, a new carbon tax. So this is going to impose steadily increasing fines. So in 2024, 2030, we're going to get more fines on buildings that violate emission standards.

And so these are envisioned or coming in many different cities across the United States. Some exist internationally. And here, the point is just the distribution of these fines is heavily concentrated in these older buildings, these sort of pre-war buildings. And so these are also the buildings that we know from a prior work, at least in demand due to remote work shock.

So that kinda works out. We've got this kind of concentrated cluster of buildings that are both dirty and people don't wanna be in them. So let's figure out what else to do with that space. So how do we identify a plausible candidate for conversion to multifamily building? Well, we're going to take some of these commercial districts.

So here, Midtown and downtown Manhattan, where we think that the policy imperative from an externality standpoint is the largest, right? So we just saw earlier that we think that commuting kind of brings a lot of externalities in terms of crime, in terms of consumption. Therefore, the vacancy problem in these downtown areas is the strongest, and hence there's the strongest public rationale for doing something with the space.

We're taking older buildings, Benchmark Harris before 1990. But we have a whole sort of decade of construction analysis. We kind of exclude the highest quality buildings cuz those kinda seem to be holding up. We're taking minimum 25,000 square feet. The big requirement, really the big requirement here is the floor plate.

So the floor plate is the dimensions of the building. And when this floor plate dimension gets really large, it makes the economics of conversion a lot harder. You essentially either have to then dig a cord through the center of the building or down the side in order to release more surface area.

So you can have all the bedrooms adjacent to the corridor. And so we kind of impose a requirement. You can't have a distance that's too far from the corridor of the building to the window. That kind of also means we're picking up more of those very old skyscrapers that have already a narrow floor plate.

We also want to pick buildings that don't have too much of a tenant population in place. So buildings where you don't have too many existing tenants. Actually, we've seen some cases where actually the existing tenants wanna leave, and so they actually pay you to get out of the lease.

But in general, we're gonna assume that if you have a lot of tenants, it's gonna be hard. So we're gonna focus on buildings that are more vacant. Then finally we're looking at brown offices. So we're looking at buildings which violate New York City's emission standard for a Benchmark.

Of course, there's gonna be more conversion possibilities if you don't require that. And we're gonna be focusing on conversion because it's also a lot better from an environmental standpoint, taking into account the embodied carbon in these buildings. So if you tear down the whole building, of course there will be other tear downs as well, you basically have a lot of carbon, the steel, the concrete in the building that you're getting rid of, and you have to kind of build that all from scratch.

You save a lot of environmental benefits if you do it through the conversion route. All right, so this is what we get. So on the far left is what we observe in New York City. So we're kind of starting with the total number of buildings at the top.

We're steadily employing these different requirements for what counts as a possible candidate for conversion. And so we think we can produce basically 38,000 apartment buildings in New York City with buildings that we've identified as possible conversion candidates. What we then do is kind of employ a similar algorithm for the United States.

So the difference between New York and the whole nation is we have better emissions data, in particular in New York. So we use the emissions data that we have in New York to kind of fit a model for what is a likely brown building, and we apply that national model out for the whole country.

So that adds up to over 100,000 apartments, which would save sort of half a million tons of CO2. This, again, is the graph of all those buildings. And this can be scaled up given we have sort of incomplete data coverage in Comstack, which is our data provider for this.

So if you kind of count for the fact that we don't observe all buildings in our sample, we think that this is representative of about 400,000 new apartment buildings that are plausible candidates for conversion, given our algorithm. So, to put that into context, new apartment deliveries in a typical year is 260,000 new units.

So if we converted all of these at once, it would be a large multiple of the existing number of buildings, and it would replace a pretty substantial amount of carbon as well, if we're simultaneously remediating the carbon expenses of these buildings while we go through that conversion process.

So typically, the carbon emitted by buildings is coming from the boiler, the heating system, things like that. And as you're doing the conversion, you can kind of replace a lot of those electrical, heating, cooling, et cetera, systems as you go. One challenge I'll come back to is this loss factor.

You're not getting as much space as a apartment building as you are with the office. Okay, so that's just from a physical feasibility standpoint. Just can the building physically be transformed into a multifamily building? That doesn't mean it pencils out financially. And so that means the next step of our analysis.

And so what we're actually gonna do here is a pro forma real estate analysis, where we put in a lot of assumptions about how the building is gonna be in the pre conversion state as it's in office. Take into account all the headwinds we talked about. We're gonna compare that to a plausible financials for the building, if we convert it to figure out whether this whole thing pencils out, whether it makes financial sense to convert the building in this way.

And this is going to be one of the main public goods we hope to release as a part of this project. This is all going into a calculator, which will be on a website through the Brookings Institute. And the idea is, any developer can type in whatever parameters they like about a particular building to see whether conversion makes sense for them.

A local policymaker can use this tool as well to figure out, is there a shortfall in the conversion, not penciling out financially. If so, what would therefore be the required subsidy necessary to make the project npv feasible? Right, so this is all, not just a set of results, but a framework to think about the conversion potential.

And so we just kind of throw in a variety of assumptions here. Again, these are all controllable with dials in our eventual calculator, which basically get at what's happening with buildings before COVID after COVID, etc. Okay, so there are a variety of things that are going on here.

We're basically putting together a financial model that takes into account the higher interest rate. It just got a little bit higher last week. We take into account the remote work impact here, which is impacting the rents that you can collect if it's an office building. And so we think that the higher interest rates, the higher the remote work, these are all impacting the property, impacting the property value of the building, and then finally the climate part of it.

So, taking all that into consideration, the interest rates, the remote work, that climate tax is coming in. We think there's about a 60% financial loss for buildings here. This is benchmarked all in New York City. But show you some numbers across the whole country. Okay, so the idea is that if you take these shocks seriously, you think the building should be worth a lot less than it currently is.

Can you then, as a developer, buy the building at the new valuation, spend the money to convert it into an apartment building and that's expensive, and then turn this into a green apartment building? If so, will the resulting cash flow stream make this project NPV positive, right, and financially worthwhile for development purposes.

And so we also put in assumptions here about what the rent is gonna be, what the financing cost is, so on and so forth. These are all both assumptions that we kind of do our best guess at, but also toggleable dials again in this financial calculator. So the upshot is basically it can pencil out, but this is, I would argue, actually a little bit of a knife edge case.

So we're getting a little bit of NPV here. So this is a building that was about $100 million in valuation. It's like a standard kind of assumption here. We think the NPV of conversion for this sort of standard building is about $4 million. The reason I argue this is actually a little bit of a knife edge case is because in order for this value drop to happen, you're gonna have to realize a pretty large financial loss.

So that may require a foreclosure or default event for this building to actually realize that loss. And then developer has to go through it. They're going to have to go through all this expense. At the end of it, they're getting 4 million in NPV, 17% IRR. So it kind of makes sense for them, but it's a little bit dicey.

So it suggests that conversion can just pencil out, assuming the large value decline. And the reason it doesn't pencil out even more easily is office rents are actually pretty high relative to apartment rent. So you basically need to write down the value of the office a lot and then convert it into a luxury apartment in order for this thing to even be close to working out.

All right, so if you tell a policymaker this, the first thing they're gonna tell you is, okay, well, that's gonna generate luxury units. But we don't want luxury units, we want affordable housing units. Okay, so we consider affordable housing mandate. What if you then require the building owner to provide affordable housing?

Affordable as in requiring that the rents on the unit not exceed the local area median income. So it's a very common requirement through inclusionary zoning and other mandates recently. So we basically conclude it doesn't really pencil out if you do that. You kind of really need to have these units be either market rate or have the government provide more subsidies if they want these units to be affordable housing.

And yeah, we have an estimate of exactly how much. So in order for that to happen, you can do a few things. You need property tax abatement. So that's basically a program that would allow for the government to cut your property taxes if you generate affordable housing units.

That's actually something New York City did through the 421 G program, specifically for office conversions back in the nineties. It's something New York City used to do with the 421-a program. It's something many other cities have done in various contexts. You could also subsidize the debt that's going on here.

So part of the problem here is how expensive the interest rate environment is. The government can subsidize that. And finally, you have the Inflation Reduction Act. We're going to highlight some provisions in the Federal Inflation Reduction act, which we think. We think can cover the cost of these conversions, provided they also meet that environmental perspective.

So part of the point here is that this conversion is something that policymakers should be really thinking about because they get more tax revenue this way. If it happens, right? If the value of the office goes down, you lose all that revenue. If you can convert it to higher use, you kind of get more property tax revenue.

Across a lot of different scenarios. We do a lot of other sensitivity analysis. How sensitive are these results to conversion costs, rents, all that stuff? Not surprisingly, the results are very sensitive to all the inputs and assumptions going into it. And so we think of this whole analysis as less a definite indication that these buildings definitely should be converted and more a kind of proof of concept that conversion can pencil out under the following assumptions.

It would affect however many units and policymakers that really want to push this need to think seriously about the types of incentives necessary to really make it happen. So that's gonna be the last part. What policy levers really need to happen. So this is a very active and ongoing discussion in many cities.

Basically, you need to change the zoning and building codes to make it happen in the first place, ensure that the building owner can do this conversion as of right to avoid all the delays and construction costs, lawsuits and so forth that derail these kinds of construction programs in general, you need to change building standards to be more flexible in allowing conversions, particularly with this bedroom issue.

So Charlie Munger, right, loves these windowless dorms. That's sort of one way to get these conversions to pencil out is remove the requirement that every bedroom have a window and providing additional density bonuses, getting rid of parking requirements, these things like that. Cities are actually moving already in this direction.

Like in New York, Eric Adams has announced a big new zoning overhaul that also gets rid of the parking requirements, allows for SROs, and is also moving forward on the office conversion side. Last thing I wanna emphasize is Inflation Reduction Act has a lot of money that's being spent.

And we've identified a number of provisions which basically are large pots of money which can go towards this purpose because these are, are funding sources intended for different spending sources that are energy efficient. If you take a dirty office building, fix the environmental issues, it qualifies for these subsidy programs.

And so we think that there's a lot of money out there from the federal source which can also go and augment the local policy stimulus in order to make more of these happen. All right, so that was all our paper on conversions. I do wanna spend the last couple of minutes and just pitch another paper, which is a work in progress.

Here we're measuring the productivity impact of remote work from GitHub data. So looking at commits, coding data commits. And here the comparison is that the two lines at the top there are basically fully remote and hybrid work. And the bottom line is the in person workers. And this is comparison of how many inputs you're doing on GitHub compared before the pandemic as a proxy for productivity, right?

And so we see two things. One, we see greater dispersion. So remote workers are doing more things at all hours of the day. We know that from earlier work, and we also see a big increase in this productivity measure for the fully remote and for the hybrid workers.

So it's consistent with a substantially positive impact of remote work on productivity. Here. Let's go back to Brian's point earlier about scaling, firm scaling. So here we're measuring basically employment counts for firms, and this is looking at a coefficient of whether your firm goes more remote. So the remote firms basically scale a lot more when they go remote.

And it's kind of coming after sort of flat pre-trends before the pandemic, their kind of employee count sort of rises. This is very much work in progress, but I figured I'd use the opportunity to pitch it to you a little bit and try to solicit more feedback from you on where we should kind of take this project.

All right, thank you all very much.

>> Speaker 1: So, can I ask the first question? I had the following thought experiment in mind, which is if the alternative to converting these buildings is basically tearing them down and building something entirely new, the carbon benefits are potentially much bigger because steel and concrete production is so carbon intensive.

So I was wondering if you'd considered that or if there's some way to quantify that in a way that might be useful to policymakers.

>> Arpit Gupta: Yeah, so there's some existing work that quantifies this embodied carbon problem. So how much carbon is existing in structures? So it'll be common for buildings that have tons of embodied carbon, it just varies so much by existing buildings.

So for sure that's in the background, for us, that's just sort of the reason why we focus on conversion from the get go as a way of trying to remediate the expense, the emissions. But I take your point. We can probably think a little bit more about doing a more like, for like, comparison and really quantifying what the total carbon benefit is relative to tear down.

 

>> Speaker 3: There's another carbon benefit to assess, which is, is it worthwhile to go from brown offices to green apartments? Or would it make more sense from an economic perspective to go from brown offices to brown apartments. Okay, so you positioned this paper as about going from brown to green.

But you told us nothing about whether there's actually what the economic payoff is or what the carbon reduction payoff is from the brown to green conversion. I have no idea. It may be that these things that are marginally pencil out to a positive NPV conversion might be a lot more attractive if they didn't have to go to green.

They just went from brown to brown. And you still avoid the tear down.

>> Arpit Gupta: Yeah, so that's the emissions line there. So that's basically the tons of carbon that we save if we go through ahead with the conversion.

>> Speaker 3: Okay, but what's the cost side of that? If we went from brown to brown, we still don't have to tear down the buildings.

That's Jose's point. But does it make the conversion a lot cheaper? How much of this conversion cost is really-

>> Arpit Gupta: Most of the conversion cost is not coming from the carbon remediation.

>> Speaker 3: Okay, well, then that point should be drawn out. That seems like a key point to draw out.

And I didn't, maybe it's in the paper, but I didn't get it in the presentation.

>> Arpit Gupta: That's a good point.

>> Speaker 4: So I'm hearing a lot from policymakers in the UK and also in Canada about things like this. But nobody seems to be pulling the trigger on it too much.

Now. One of the main reasons for doing these conversions is actually because there's a social value to having a lively city center. Would there be any way to estimate, say, the multiplier effects that just come from housing more people in city centers?

>> Arpit Gupta: That's a great point. Something we've thought about.

I think it would be possible to do that, especially with historic conversions that have happened. For example, in downtown Manhattan. What the multiplier was, I think we didn't feel we had a good enough identification strategy to really sell that. One other point to mention just about other countries that are doing it, learn.

Over the course of this project is in the Netherlands. They actually have like a central government office dedicated to this. It's like a team of people that go out and help people convert their buildings. And 15% of the apartment of the housing constructed in the Netherlands over the last ten years comes from office conversions.

So there are other international examples, actually, of policy makers and countries that have done this very successfully.

>> Speaker 5: I was wondering if you've thought about sort of what are the equilibrium Equilibrium counterfactuals. So if these buildings, if they're not converted to residential, then you'd sort of expect the office rents to fall.

It's possible that there are suburban locations where residential construction would take place, except for the office demand that's happening now. Or maybe you think the buildings will stay vacant. I'm just sort of thinking from the perspective of a broader urban equilibrium. What do we expect the spillovers to be to the suburbs, to the size of the office market in general.

Since I think that'll help understand what the net effects will be on carbon reduction and so on.

>> Arpit Gupta: Yeah, that's really part of the motivation here. So what we're identifying is basically like ten to 15% of office buildings are financially feasible, or, sorry, are physically feasible, of which about half of those are financially feasible for conversion.

So it's definitely not most office buildings by any stretch of the imagination. But think of a market like New York, is like in the low 20s in terms of vacancy rates. Or a market like here, that's in San Francisco, that's like 33% vacancy rates, right? If you take an office market like that, and you take, let's say, 5% of the office stock and convert that, you've now kind of changed a sort of disastrous office environment to one that moderates substantially.

So you can sort of, even with not that much conversion, the hope is you can kind of potentially stabilize a lot of what's going on in the office market. At the same time, you're adding more residential, you're adding more eyes on the street, you're kind of addressing some of the fundamental problems of the downtown and sort of help revive the whole city, even if it's not like a ton of units.

So definitely that kind of equilibrium sort of thinking is behind this. We haven't formalized exactly what it would be, but that would be like a great have to think about.

>> Speaker 6: If I could piggyback on that. Interesting laying out of the issues, particularly as somebody looks at these conversion issues.

But I was curious how you might get a more dynamic result through phrasing it as equilibrium. Because there's an implied assumption here that the current status of remote work is gonna continue. And the current status of what is perceived as lack of office demand in urban centers is gonna continue.

And that is not at all clear, at least looking locally. And sort of the related question to that is there is some argument that can be made that the reason why we have all sorts of problems, both in the office space area and in the housing area is because of the imposition of various things that interfere with the private sector response.

It sounds to me like you're proposing more interference with the private sector response. Have you looked at what might be a reasonable, this is where it piggybacks off of the earlier question. A reasonable private sector response.

>> Arpit Gupta: Yeah, so first, just to clarify, a lot of our objective really, it's gone here.

Anyway, a lot of our objective really is to remove the zoning regulatory impediments that are preventing some of these conversions from taking place purely from private developers. So I'm totally in agreement that a lot of the ways that we're gonna get this done are by removing government that is currently a blocker in this situation.

I also think back to your first point about, are people going to go back to the office? This is actually a great way to get people back to the office, because we know from work people here, like commuting distance, right? Is a key barrier behind people getting to the office.

And so if you build more apartment units downtown, that will produce a population that is more inclined to come back to the office. I think in terms of, is that office oversupply or not? The other thing we know from our earlier work is it's really the newer buildings that are holding up the best.

So I think even if office demand kind of comes back incrementally from these new workers. That will result in either new office construction or it might change like a little bit on the margin, whether you wanna be in a building from the 80s or the 70s. We're really focusing on the buildings from the 30s and the 20s and so on and so forth, where we think this demand has really fallen away.

 

>> Speaker 7: Hi, a few years ago, a co-author and I published a paper, it was sort of a theory, simulation paper. Analyzing the effect of taxing emissions in cities where the tax is levied on buildings and on automobiles. And so we came across some engineering literature saying that heating and cooling costs from building, and thus emissions, depend on the surface area of the building.

And that in turn implies that the way to tax buildings is not simply by square footage, but in a different way that I won't tell you about it. You can ask me later if you want, but what's the form of this emissions tax? That's my bottom line question.

 

>> Arpit Gupta: It's on the emissions, it's on the actual emissions for the building. So they have building level data. How much emissions are you putting off? And so-

>> Speaker 7: Is that right?

>> Arpit Gupta: Yeah, and so, already happened, cuz they have the energy usage, the heating, the cooling, etc. So they have all that data.

 

>> Speaker 7: Okay. So it's based on energy usage, not actually measuring emissions.

>> Arpit Gupta: That's right.

>> Speaker 7: Okay.

>> Arpit Gupta: Energy usage, that's right, I should be clear.

>> Speaker 8: I had a couple of comments, questions. One was when you said vacancy rate to me, there's a couple of headings. Vacancy, meaning it's empty and nobody's there versus it is currently on a lease and someone has it on the sub lease market.

And it's still also physically, and there's no humans in there, but it's being paid for. What definition would you use?

>> Arpit Gupta: So, like for San Francisco, I think it's like a 33.9 availability rate, including the subleases. I think it's at the high 20s with an actual vacancy, yeah.

 

>> Speaker 8: Okay, thanks, Juan. To the conversion to housing as a number, did I hear you say it starts with, first, you have to realize you're gonna lose a lot of money on the value.

>> Arpit Gupta: Yeah.

>> Speaker 8: And so the owner of the current building has to have a very quiet conversation with themselves about losing some large amount of money before you can start this whole premise.

Did I hear that right?

>> Arpit Gupta: So that is happening in some cities, San Francisco is actually the best example. Office per square feet have gone from 800 before the pandemic to now you're seeing actual sales at 200 square foot. So there are certain cities where that process of acceptance, denial, etc., is moving along more rapidly.

 

>> Speaker 8: Okay, thanks. And the other last one is, we hear you're talking about housing. Are there other uses that you've seen might pencil out better? I'm thinking like community college use or alternate school campus use or things like that.

>> Arpit Gupta: Yeah. It's outside the scope of this paper, but I absolutely agree with you.

I know in New York there's been interest by educational institutions, by hospitals. Lab space has been a huge demander of urban space, you know, everything from warehousing space, asylum centers as well. So there are a variety of uses, I think, that can be done with this space. And the big picture is cities are constantly churning in renewal, right?

So the space in Manhattan was once factories, is now lost, is now something else. So this is just part of the general process by which cities turn over.

>> Speaker 1: All right, I think we need to move on. Thanks so much. All right, so second paper of the session is gonna be presented by Jack Lieberson on the short and long run effects of remote work on US housing markets.

 

>> Jack Liebersohn: Thank you so much for having me. As we've seen from all these sessions and from a lot of research by folks in this room, the rise of remote work has changed how people wanna Consume their housing. So it's changed where people want to live, the movement towards suburbs, exurbs and possibly entirely new cities.

And it's also changed how much housing people want to consume. People have home offices and during the pandemic, spent more time at home. It led to an increase in the size of homes that people wanted to consume. Now, as any good urban model will tell you, this led to changes in the price of housing, both in relative terms across space and in absolute magnitude magnitude.

Housing costs went up dramatically during the pandemic and have stayed higher ever since, despite the rise in interest rates. The argument that we're making in this paper is that the long run implications of remote work for the us housing market are going to look very different than the short run implications have been.

So to give you an example of how that might play out and some of the mechanisms that we're studying here, I'd like you to consider a city like Houston. Historically, Houston has been relatively affordable, has had relatively affordable housing for the middle class. And in part that's because construction is easier in Houston as compared to places like, say, the Bay Area or the La metro area where I live.

However, during the pandemic, there was a rapid increase in housing costs, rents and house prices, even in relatively easy to build places like Houston, Dallas, other places in Texas. Now, in the long run, we think that new construction will accommodate some of the changes in demand. Even though that new construction didn't happen during the pandemic, simply because construction takes a while.

What that's going to mean is that in equilibrium, more people are going to move to Texas because housing will become more affordable. And that'll have effects on the sort of general equilibrium of the housing market in the United States, including places like California, where people are moving from.

So those are the sum of the mechanisms that we're gonna study. The basic point is that construction takes a while. And so the short run equilibrium that happened during the pandemic is not reflective of the long run equilibrium that we expect in the housing market. So our paper has two parts, and I'm going to take you through each of the parts one by one.

In the first part, we're going to take as a given the changes in population movement and the changes in housing costs. That occurred during the first years of the pandemic, the rise of remote work. We're going to interpret those changes through the lens of an urban model and use the structure of our model to back out what we think are the fundamental shocks.

So what do I mean by the fundamental shocks, the shocks to housing demand to live in each city, in the country or each location, and the changing demand for larger homes. So we're gonna back out demand, or the change in demand to live in every US county and also within each county.

How much bigger homes did people demand? Then we're going to feed those shocks back into a version of our model that allows for more construction. What do I mean by more construction? Well, we're gonna assume that in the short run, new construction was not possible, but in the long run, it is possible.

And so when we take those shocks and feed them back into the model. What we arrive at is a different urban equilibrium with a different level of housing affordability in each location, and different patterns of where people live. So there are two key facts that we want to feed into our model, and that we want to interpret through the lens of our model.

The first is the aggregate increase in rents that occurred from 2018 to 2021. So overall, we think that the rise of remote work led people to demand larger homes, both for home offices and because they're spending more time at home. And so they might want to consume more housing in other ways.

The second is the sort of set of cross sectional facts that house prices changed, that housing costs changed in relative terms across locations. So one of the big cross sectional facts that we're interested in interpreting is the donut pattern that we've seen documented in many papers by Nick, by Arpit, by Jan, by other people in the room.

With the less need to commute into downtown offices, people were able to move further away from CBDs, and prices in suburbs and exurbs rose relative to housing costs in CBDs. Okay, we microfound a model. It is sort of, if you've seen urban type models, it looks very standard.

We log linearized the model. And what I'm going to take through is sort of the five key equations that define the urban equilibrium in our model. And this is the result of a log linearization. , in the model, every individual chooses how much housing to consume. So you can think of how many square feet and also which location they want to live in.

For our empirical application, each location is going to be a county, but in principle, we could have estimated this with zip codes or msas or whatever you want. Okay, so the first equation is the housing demand equation. And it says that the amount of housing per capita. So you can think how many square feet is decreasing in local rents but then there's also a housing demand shock epsilon.

And this epsilon is going to be one of the shocks that we're going to try to back out from the data. And this is specific to each location. So epsilon has a subscript i, i is a location. The second equation is the housing supply equation that says that the change in total housing in each location is governed by the housing supply curve and a housing supply shock c.

Each location has a housing supply elasticity sigma. And ultimately we're going to use the, ultimately we're going to assume that in the short run each location has housing that's supplied completely inelastically. In the long run the elasticities will be positive. Third equation is the location demand equation. This says that the change in population in each location is decreasing in local rents with an elasticity mu.

There's also a location demand shock eta. If some places become more attractive because of the feasibility of remote work, the adas will be more positive in those places. Finally, there is an outside option that determines the population in each place, and that's U tilde. U tilde is determined endogenously in the model.

Fourth equation says there is a housing adding up constraint which says that the total amount of housing in each location is the change in population plus the change in housing per capita. And the last equation says that everybody has to live somewhere. So the total population in the US equals the total population in each city summed up.

Oops, okay, there we are. Okay. So those are the five linear equations that determine our model and that we solve. Now, I'd like to speak a little bit about the structure of the model and why we picked this model compared to some models. It's relatively simple. We only have three shocks, a demand shock, a supply shock, and a location shock.

The model is not dynamic, although we do a dynamic extension in the paper. The reason why we like the simplified model is because we can calculate formulas that back out the shocks from observed moments in the data, and we can also calculate analytic formulas that show what the effects on average rents are based on those shocks.

And so we can calculate pretty much everything we do, analytically, and we think that gives a lot of insight into the structure of the models and how different sort of counterfactuals will affect the urban equilibrium. So what are the forces that govern average rents in the model? Well, there are two important forces.

One is the demand for living in particular places. If more people want to live in places where housing is more elastically supplied, that will lower average rents. Why is that? Well, if we take one person and we move them from an inelastic place like the South Bay to an elastic place like Urbana-Dhampaign, house prices will rise a little bit in Urbana-Champaign, because housing is elastically supplied there.

House prices will fall a lot in the South Bay, because housing is inelastically supplied there. And that means that in equilibrium, the average house price in the US will fall. I should say. I'm saying house price, but it's really a model of rents. House prices are dynamic, and we extended to that, but I should really be saying rents.

So the point is that the location of where people want to live affects the average rents and the average levels of affordability. You can probably see where this is going. If there are some sort of shock, like a remote work shock, that allows more people to live in places with an elastic housing supply, that will lower rents overall in the country.

Now, there's another force, which is the demand for housing size. If people demand larger homes, that will mechanically increase rents. If there's a shock such as a remote work shock that makes people demand larger homes, then people will consume more housing and that'll raise housing costs. Ultimately, we're going to see that those forces go in opposite directions, and the exact magnitudes will depend a little bit on the calibration.

Okay, so what about the short run? Well, in the short run, we assume that housing is supplied perfectly inelastically everywhere in the US, and that's how we're going to interpret the years, sort of 2019 to 2021. It's a stark assumption, but we think it's relatively reasonable. Not very much housing was constructed then.

If we modify it by allowing a little bit of housing to be constructed, it doesn't change anything quantitatively. It just makes our formulas a little bit more complicated, okay. We think this captures one of the key facts about the pandemic, which is that even in relatively affordable places, historically, even in those places, rents rose by a lot.

And the sort of best way to explain that is to say it was hard to build new housing there, even though in principle it should be possible. In the long run, the forces will be very different because the shift towards elastic areas will affect aggregate house prices. As those elastic areas are able to finally build housing in the both the short and the long run, the increase in demand for larger homes will raise prices.

So that force is still there, but that will also be ameliorated in the long run as more homes get built or as larger homes get built. Okay, so now I'm returning to the five key equations. I'm just showing you the same equations as I showed you before. The only difference is that the second equation is zeroed out.

It says that in the short run, no new housing is constructed. The elasticities are set to exactly zero everywhere, and there is zero supply shock. That's a short run of housing markets. With these five equations, now four and a few lines of algebra, we can back out the structural shocks that we care about.

The two structural shocks we care about are given on the first line, epsilon, which says how much housing people consume, want to consume a bigger home. The second structural shock, Ada says, where people want to live. What you can see is that both of these shocks are linear functions of the changes in population and the changes in housing cost to measure as rents in each location.

The exact linear combination of housing costs and population that tell us each shock depend on the calibrated parameters lambda and mu. Lambda and mu are the two elasticities of housing demand, the elasticity to move to a particular place, and the elasticity to consume a larger home. So let me, before getting to our estimates and telling you our preferred calibration, I want to give you a little bit more intuition for what those two elasticities are.

Okay, so I'm gonna start with mu. This is what we call the location demand elasticity that says, if rent is higher in a particular city, how will the elasticity to move there change? So in a Rosen-Roback type model, mu is infinity. What does that mean? That means that if rent goes up in one city, compared to an identical city, people will move out until the rents are once again equal, the sort of Rosen-Roback assumption.

If mu is zero, then people won't move. Regardless of how high rents are. They don't care about rents. They really want to live there. We think that a high mu is reasonable because we're talking about the long run. We actually have another paper where we try to estimate mu, and we come up with large values.

It doesn't have to be infinity, but a mu of something like ten is basically infinity for the quantitative purposes. And again, that's over the very long run, we're thinking maybe 30 year timeframe. The second elasticity is lambda. That's the housing demand elasticity. It says how much of a smaller home people will want to consume if rents are higher.

If lambda is zero, that means that everybody consumes a unit, just a home. If lambda is 1, that corresponds to Cobb-Douglas housing consumption. We think lambda equals two-thirds is reasonable. That's based on a bunch of previous research other people have done, and it's what we choose in the other paper.

But the results are not very sensitive to lambda, they're more sensitive to mu. Okay, so with those calibrated parameters and the formulas I gave you a few slides ago, we can back out the fundamental shocks to housing demand and location demand and how they changed from 2020 to 2022.

We use rents data from. From Zillow. And we use the population data that Nick helpfully provided as a public service to the profession. I guess FOIA'd and USPS provided it with some help. That's right. Now, for the long run, we don't assume that housing is inelastically supplied. We say that the housing supply will instead be equal to its historical levels.

And we use the elasticities from Baum-Snow and Han as a sort of benchmark. And we do some work to reconcile those with the elasticities from Sais, which we think can be reconciled if we think about the timeframe that they're estimated on, right? Net change of address requests from USPS tell us population.

Okay, so in the paper we show various maps of where the demand shocks are. Not surprisingly, they look a lot like the maps of population movements and rent changes. So you see exactly the same doughnut pattern, the same decrease in demand for CBDs, the same increase in demand for sort of suburbs and exurbs.

Let me advertise the paper. You should look it up and you can see some nice maps. I'm gonna skip those in the presentation. In the interest of time, we plug those back into the model. And as I promised you, we derive formulas for the size of these two effects.

The intuition I gave you a few minutes ago said that if more people demand housing in elastic areas, that will decrease aggregate rents. And in fact we show that the size of that channel is exactly equal to the negative covariance of the demand shocks and the elasticities. And that's scaled by the sum of the average housing supply and the average intensive margin housing demand.

So the effects aren't as large if the average housing supply is higher everywhere. So as demand shifts towards more elastic areas, the aggregate price will fall. This scatter plot shows the relationship between what we think are rent changes over this time and elasticity. And you can see there's a positive relationship, although it's sort of curved.

What this means is that the demand shocks were quite negative in the most inelastic parts of the US, basically CBD's of major coastal cities. The demand shocks were more positive in the sort of suburban areas in the middle, and they were roughly flat in very rural areas. This means that there's a positive relationship between elasticities and demand shocks.

And what that will mean is that the effect of location demand will be to make housing costs more affordable in the long run, even though that didn't happen in the short run. We calculate that sort of using our benchmark parameters, the covariance is 0.0023. We calculated sort of average housing elasticity of 0.25.

And so the net effect will be a long run, half a percentage point lower price due to remote work relative to the pre-pandemic benchmark, okay? If you think remote work will change where people wanna live by more than we saw, because it'll expand, then this estimate can be scaled up accordingly.

If you think it'll return to a lower level and how location demand shocks will decrease, you can scale this down proportionately. The second effect is that people demanded larger homes. And this channel is actually very easy to calculate. It depends on the average housing size demand shock, so the average value of epsilon.

And we derive a formula for that. And again, it depends a little bit on what we think of parameters are. For lambda equals two thirds, we arrive at a real rent increase of about 0.045. That is 4.5 percentage points. Keep in mind, that's way down from the short run effect, which was eight percentage points.

Right. And the short run rents went up by about 8% over this time period. In the long run, we think that'll go down to 4.5%. Right, and sort of we can calculate something like 4.9 percentage point long run channel. Okay, so the net effect of remote work will be the sum of these two channels.

In both cases, these, these channels mean that in the long run, housing will be more affordable than it was in the short run, because new construction gets built to accommodate the changes in housing demand. The fact that people want larger homes will increase rents. The fact that people are willing to live in more elastic areas will decrease rents.

We think the net effect is still likely to be positive, but not nearly as positive as it has been. And depending on parameter values, something between two and four percentage points down from the eight percentage point short term effect we saw. Okay, so there's a lot more in the paper, implications for prices, evidence that our results are starting to happen.

We wrote this paper about, we started writing it about a year ago, and some of the things we predicted about sort of long run population movements are in fact starting to happen, which is really exciting. So I encourage you to look up the paper for more detail, and I'm happy to take your questions.

That's all I have. I'll let whoever controls the microphone.

>> Speaker 11: Thanks. Great. So one, really interesting, I don't know this is part of it, or one claim I heard is that hybrid, as much as remote, is bringing a bunch of land into use that was previously unbuilt on.

So there's a paper claiming if you look at areas that have something like one and a half to three hours commute out from the city center previously, no one would really live in that commute into the center because it's too far. If you're only going in twice a week, say, that's now within space and apparently around a lot of big US cities, that stuff vacant.

So I don't know, is that part of elasticity? Is that like.

>> Jack Liebersohn: Absolutely. That's exactly what we sort of are expecting and hoping to find.

>> Speaker 11: Fringe Houston and Fringe LA and these areas that are just so far from-

>> Jack Liebersohn: That's exactly right. So a lot of the change in population is from CBD's to sort of fringe part of cities.

We were, we expected very rural areas would see population increases, but that's actually not true at all. And what determines the covariance between elasticity and demand is really the change from CBDs to exurbs. So exactly what you're talking about, and that's driving all of the results. This sort of rural areas effect is not nearly as important as I expected it to be going in.

It turns out it doesn't really matter very much at all.

>> Speaker 12: So one thought was, as workers move towards these new areas and the schools become better and the services around those areas become better, would the elasticity parameters endogenously change in these new places over time? Because you are assuming the same parameters over the two periods, correct?

 

>> Jack Liebersohn: I think that's very possible. One of our referees made a similar comment. And so our appendix c endogenizes the elasticities. Ultimately what we learned from endogenizing the elasticities is we're talking about maybe 5% of people moving here. So we don't think it's So big that it'll affect the elasticities that much.

The elasticity estimates from SAIs and from Baumsnow and Han are a function of density. And so you would think that exactly what you're talking about happens as the density changes mechanically, the elasticity will change.

>> Speaker 3: So what you showed us are reds, right?

>> Jack Liebersohn: Yes.

>> Speaker 3: Okay, so if I take your model seriously, and I think I can use it to back out.

Well, I have to make another assumption that the adjustment from the short-run equilibrium to the long-run equilibrium happens at a constant pace.

>> Jack Liebersohn: Yes,

>> Speaker 3: if I take that assumption and combine it with your model, I think you can back out an implied time to the long run equilibrium with a simple asset pricing formula.

So you gotta set a discount rate and so on. But you can do this city by city, right? You take an asset pricing formula.

>> Jack Liebersohn: Yep.

>> Speaker 3: And combine it with your model and this constant adjustment pace assumption, there's only gonna be one value of the time it takes to get from the short-run to long-run for each city that will be consistent with those equations.

 

>> Jack Liebersohn: I see.

>> Speaker 3: So it would be useful to do that, I mean, there's gonna be the discounting, the asset pricing model, that's gonna matter. But the variation across cities will presumably be similar across a wide range of different assumptions you might make for the different asset pricing formulas.

 

>> Jack Liebersohn: That's interesting, so what we did was assume that all cities will happen at the same rate, and then calculate the implied discount rate, which is a little bit of a different exercise.

>> Speaker 3: Inverse of what-

>> Jack Liebersohn: Exactly, what we did is the inverse exercise.

>> Speaker 3: It seemed like you wanted the asset pricing ought to be uniform across cities, and then it might tell us something interesting about what the market is expecting the adjustment period to be city by city, which is also interesting.

 

>> Jack Liebersohn: That's a very interesting exercise, that's a great idea, thank you, yeah. Sorry.

>> Speaker 6: So I had two small questions.

>> Jack Liebersohn: Perfect.

>> Speaker 6: Okay, my first question relates to Nick's, in that I'm assuming you're working at the county level because of data restrictions, but that should change over time and hopefully over the lifetime of the projected project.

And I feel like it's important to try to get a little bit smaller than that because of how much of the moves are within city. And so I'm really concerned that you're sort of underestimating the scale of the effects that you're interested in because of the county restriction.

Second thing is, I think that it'd be nice to say something about the interaction with the commercial real estate market here. Because they're both, we're talking about the conversions that are happening downtown, or we're talking about retail going to the suburbs. These land markets and real estate markets don't happen in isolation here in this process.

And so I don't know how to think about the interaction with the commercial real estate in the simple setting that you're trying to set up.

>> Jack Liebersohn: Those are both great comments. I suspect that as you look at finer levels of granularity, you'll see the same pattern of people moving to more elastic places sort of happening over and over again, especially if you look at something like LA county, which is giant.

And I think that'll be future work, but it's definitely future work we wanna do because there's no reason the model has to be at this level. So it's a great suggestion. All right, your other suggestion was about commercial real estate, which I totally agree, this is sort of aside of the things which we've totally ignored, but I think is probably really important.

Yeah, thank you.

>> Speaker 1: All right, thanks very much. Great, so final paper of the session presented by Charlie Porger on Remote Work and City Structure.

>> Charly Porcher: Thank you very much to the organizers for including the paper in the program. This is joint work with Fernando Monte and Stefan Rossi Hensberg.

So I'm conscious I'm the last career between you and dinner, so I'm gonna try to remain as entertaining as I can. So we're all here to discuss the consequences of the recent dramatic adoption of remote work. And what we do in this paper is that we want to reflect on one of the possible long-run consequences of remote work, on the structure of cities.

And at the heart of what we're going to think about is one of the most central ideas in urban economics, which is that people benefit from being close to each other. If you go to the city center, you benefit because you meet other people. And what remote work changes is that it allows us to live in the same city without being in the same place when we reduce.

And that could mean that we might lose on these interactions. And this idea leads to a coordination problem. I benefit from going to downtown areas if others do, but if others don't, then I benefit less, and maybe I decide to also stay home. And this coordination problem could lead to multiple stable equilibria at the level of a city with one featuring high levels of commuting or one featuring low levels of commuting.

As we know, when we are context, when we think about context with potential multiple equilibria, initial conditions matter. And this is where the pandemic lockdowns come into the picture. The lockdowns moved us and forced us to move to an initial situation where we were not commuting. And if we lived in a city that could sustain these low levels of commuting, then it might have remained at these levels and not have returned to higher levels of community.

And so what I'm going to show you as one of our main empirical results is that we see very heterogeneous patterns if we look at different cities in terms of activities in the city centers throughout the pandemic and the recovery. And if we think about visits to city centers in some cities, particularly smaller cities, we'll find that they have mostly returned to their pre-pandemic levels.

But in the largest cities, we see that they remain much below pre-pandemic levels. And so if we, we're going to argue that if we think about this, the responses of these cities through the lens of this coordination problem. We can predict pretty well which are the cities that came back and which are the ones that didn't, depending on whether we think they could sustain these low levels of commuting in the first place.

Another consequence of thinking about remote work through this coordination problem is that it has implications for the strength of agglomeration forces in a city. And therefore it has implications for aggregate productivity and ultimately for how welfare is affected by whether a city has switched or not to these low levels of community.

So what we're gonna do in the paper it's try to think conceptually about this question of coordination problem. And so we're going to build what we think is the simplest model that can capture this idea of remote work in a city with the key trade offs at play.

And that can help us understand what would determine how a city responds to the types of lockdown shocks that we were talking about when the decision of whether to be remote is endogenous. And so we're gonna consider a monocentric city model in which all firms are located in the center, in the CBD, and people are going to live in one unit of housing.

So you'll see a lot of the characteristics of the model are simplified to their most so that we can capture the role of this coordination problem. And so people can decide whether they want to deliver work remotely or in person. If they work in person, they have to commute to the city center.

And if they work from home, they can stay some fraction of the week at home. And wages are going to be a function of the productivity, they're going to vary by the delivery mode. They're going to be reflecting the fact that the productivity of work in the office or at home might be different.

And the key trade offs that people are going to face is that remote work offers some attractive features. It offers lower commuting costs, potentially lower rents if people move to the outskirts of the city, also higher amenities from spending more time at home. And as I say this, I'm kind of thinking of all the papers that we've talked about that kind of fit into these different important aspects.

But that comes at potentially the expense of lower wages and lower downturn amenities. And so just one slide to highlight the key, the critical elements of the model. We're going to think that office work, the productivity of one efficiency unit of work in the office is a, but it's z at home.

And this could be potentially different. And the number of efficiency units that people can provide is going to depend on the number of in person interactions that people have, right? With some elasticity, Delta. So this delta here is going to be an agglomeration force in the sense of an urban model.

And so I'm gonna depict in blue everything that's an agglomeration force, and in red everything that's a congestion force. And so if you're a remote worker, we're gonna say that you can spend a fraction that's fixed a fraction mu of the time at home. And so for the LC, the number of commute of people who decide to be commuters.

They're going to receive a wage that's going to be the baseline productivity of office work multiplied by the number of efficiency units that they can provide. And that number of efficiency units is going to be a function of basically the typical number of people that are present in the office.

So taking into account full time commuters and part time commuters and the remote workers, if they choose to be remote, we're going to index them by m for mixed, they have this kind of part time commuting allocation. Their wage is going to reflect that with being the average between office work, office wage, and at home wage.

So now this is gonna be a dynamic model. And so people are going to make decisions every period about whether they want to readjust their labor delivery mode. But even within period, people are making some choices, some of them. So the choices they make is where do they want to live in the city, how far from the city center and how much they want to consume in terms of housing and goods.

So if we think, for example, here about the remote workers that I'm calling m here. There's a cop douglas form of their utility in consumption of housing and goods. But the more I think, interesting part here is what happens to their utility coming from being at home. And so for a fraction of the week, they enjoy this amenity from being at home.

And then the rest of the time they enjoy part of the amenities from being downtown. Part of these amenities might be increasing in the number of people that come to downtown, right? So if I'm more likely to have lunch with my friends, the more people come to the city center, that could be a positive externality here.

But there's a cost to going to the office. Particularly there's a transportation cost that will increase with the distance between your home and the city center with some elasticity, gamma. And then potentially some congestion in the form of higher traffic with more people commuting, right? And so this kind of what we're going to find ultimately, is that we're not going to be able to reject in the data that the net effect from these amenities versus congestion is different from zero.

So really the key elasticities that will matter for what we have to say today will be delta and gamma, this distance elasticity of transportation costs. If we wanna think about commuters would just replace mu by 0 in this equation and we get what is their flow of utility.

So then the next step is thinking about how people readjust every period, their choices. And they're going to weigh these flow utilities between remote work and in person work. With also idiosyncratic preference shocks that they're gonna draw every period in a very typical discrete choice model style, subject to fixed switching costs.

Okay, so what this gives us, right, is an economy that will be governed by one key state variable, which will be the number of commuters at a given point in time. And that variable is going to evolve over time and lead the city to converge eventually to a stationary equilibrium that I'm going to characterize now.

And so this is really the advantage of having this kind of simplicity that we can characterize the city through this only state variable. And the key result from this analysis I'm going to summarize in this graph here. That tells us under which condition does a city potentially feature this multiple stable equilibria.

And so what we see is that different regions, depending what determines whether a city can have multiple equilibria. Depending on the strength of their agglomeration forces on the x axis, and the productivity of remote work, z, relative to office work. And so what we find is that if the agglomeration forces are too weak, then there is a unique equilibrium.

But if they're strong enough, and if the relative productivity of remote work is kind of in a medium range, so that it's not too low, so that no one wants to do remote work, it's not too high, so that everyone wants to do remote work. If it's kind of in this between zone, then we get multiplicity in this form of this kind of cone of multiplicity.

The next step really is for us to see the parallel and how to understand how to apply this concept to the data and see if this simple theory, once we try to get some numbers on these key parameters, whether it helps us understand what we see in the data.

Let me just fix ideas by one example before I move to the data. It's just considering a city with low agglomeration externality. And this is a phase diagram. So it tells us the number of commuters tomorrow as a function of the numbers of the number of commuters today.

If the agglomeration forces are weak, then the lockdowns, what they do is that they send us all home. So LC goes to zero, and then once we release the lockdowns, the city comes back to its prepandemic equilibrium, there's no other possibilities for this city. If we have a high agglomeration externality, this is the typical shape of the phase diagram.

And so this is where we see that the lockdown, once released, is going to bring the city back to a different stable equilibrium with low commuting values. Kind of another question is, do we see some sort of a parallel to this in the data? So some descriptive evidence that I want to show you.

This is now data from. Cell phone tracking data provided by Safegraph, where what we do is we count the number of visits to the city center over a year in each block groups in these cities, controlling for variations of a month in the number and also the changes in where people reside in order to remain consistent.

And what we see is know relative to. So everything is expressed relative to January 2020. And we see patterns that we're pretty familiar with now from different data sets is that in April of 2020, there was a drop of about 80% of visits to the city center across the board.

So, you know, in all of these three example cities here, if we think about New York, we see that there was a partial recovery, but by mid 2022, it was still well below 50% of trips to the city center. And it's a very similar story for San Francisco.

It looks quite different if we think about Madison, Wisconsin. And so this is something that we can see more systematically across cities. If we now divide cities between the large cities, those that have more than 1.5 million employment, and the small cities that have less than 150,000 employment, we see this clear divergence, right, where the initial drop in the visits to the city center is between 20 and 30% during the early month of the pandemic.

And then the largest cities stabilize around 60% of mobility or visits to the center, whereas the smaller cities are pretty much back. So this is a very big difference, right? It's 40% variation between the large and the small cities. Moving to housing now. So a lot of us have been talking about housing today.

So I'm going to go quickly because this is an exercise that has been performed particularly for big cities. I'm going to tell you what happens also in small cities with the sample that we can use. But we're going to basically also look at the low level and at the zip code level.

This is going to be housing prices just because there's a little bit more coverage of housing prices. And so we're going to estimate the distance gradients with respect to the city center. And I'm plotting here everything relative to January 2020. And I'm also multiplying everything by minus one in order to interpret a decline in this curve as a decline in the relative attractiveness of the city centers.

Okay, so we see that in New York, the gradient went down and stayed down, and similarly for San Francisco. And for Madison, maybe there was some drop in the early stages of the pandemic, but by the end of 2022, it was back to pre pandemic levels. And again, we can do this systematically across cities, and we find this kind of similar divergence between the cities that have more than 1.5 million employment, the big cities that went down and stayed down in terms of these gradients, and the small cities where things were pretty much back to pre pandemic levels.

Okay, so this seems we see emerging two broad categories of cities, right? Those that returned to pre-pandemic attractiveness of their city centers in terms of housing prices, in terms of commuting, and those that remain stuck with very low levels of visits to the centers and attractiveness of the city centers in terms of housing.

And so the next step for us is, you know, take our data, our model to the data, try to calibrate it, and, and put some numbers into these parameters for different cities and see whether we can explain whether the cities that we see not coming back are those that we predict are more likely to be in that cone of multiplicity, right?

That they could be stuck in this coordination equilibrium with low commuting. So for that, we are going to need to turn to the data. And this is where I want to emphasize that this exercise is a way for us to test the predictions of the model, whether the theory has a bite in the data.

But there's a lot of the parameters that we need that collectively as researchers, we have very little idea about their actual magnitudes. So a lot of work today shows that we have a lot to learn about what determines these different parameters. So nevertheless, we're going to do our best with the data that we have available.

And so we're going to turn to the national Longitudinal Survey of Youth and the panel 79. And for that, with that wave, we have information about the number of hours that people report working from home. And we're going to use it to calibrate two key parameters. And we're going to simplify things a little bit.

We're going to assume that we're switching costs between remote work and in office work is going to be the same across cities, and so will be the elasticity, the dispersion of those preferences. And so with the NLSY, the number of hours worked from home that we see in that panel, we're constructing transition probabilities, transition at the individual level into and out of remote work.

And, for example, the first thing we do is we ask when the wage differential that we see people getting between in person work and remote work is larger, by how much do transition into remote work increase? We get an elasticity of about two. So that's the first set of parameters, and then we need to get everything else.

And so all of the other parameters we're going to obtain as city specific. Okay, so they're going to be different for each city that we have in our sample, and we're going to end up with 274 cities at the end. For the agglomeration externality, we're going to use estimates that Stefan Rossi-Hansberg, in his previous work, has obtained from industry level, and we're going to aggregate them at the level of a city using the industry composition in terms of employment.

I guess one of the very tricky parts is how do we calibrate the relative efficiency of remote work, the z over a? And so for that, we're going to try to leverage the information we have in the NLSY. We do have variation at the individual level, in the number of hours that people spend at home.

We see their earnings. And so we estimate occupation specific remote work penalties or potential premia, and we use those occupation specific numbers to aggregate them at the city level. And on average, we find about 10% discount for remote work through this method. And importantly, it's controlling for persistent and observed heterogeneity between people using the panel dimension of the data.

And then the last set of parameters we need is the elasticity of commuting cost to distance. And here it's going to be pretty standard. We're going to go to the ACS, get census block group data on rents, and do a gradient regression to estimate those gammas by city.

Okay, so what do we get? Well, we're going to get predictions about which cities fall in that cone, right? Which ones are in that cones? And the question is, is that prediction about whether a city is in the cone of multiplicity associated with what we see in terms of that drop in commuting?

Or in visits to the city center. And so this is the, I guess, the main test of our results. We see that the cities that had the strongest reduction in trips to the CBD. So you think, San Francisco and New York are those that we predict are the most likely to be in that cone.

And this is the result from a bin scatter where we fit a probate model to this set of cities. So we see this as pretty encouraging that, yeah, cities that are pretty much back in terms of commuting, we don't tend to predict that they should be in a multiplicity situation.

Okay, now we can also summarize what we find in terms of the share of cities in the cone by employment. But I showed you this very strong relationship between the patterns of recovery and employment. And so similarly, we find that the biggest cities are more likely to be in that multiplicity potential, right.

And so in a sense, this is telling us that the size of cities is a good summary for the set of parameters of characteristics that make cities more subject to this coordination problem. Okay, so last thing I wanna show you is what happens if we push the model again, with all the caveats that I want to really clarify.

That, this is really just if we go all the way and ask what are the welfare consequences of these switches to a different equilibrium. So this is something we can do for all the cities that are in the cone, for which there are two possible stable equilibria, one with high commuting and one with low commuting.

And just ask omitting for now transitions between these two equilibria. Just thinking about the stationary equilibria, what is the welfare difference between those two equilibrium? And we find some results that are between zero and 3% decline in welfare. And so, these are not zero, but also not huge, right?

But what we learn from this exercise is that we can also decompose what is driving those changes. And we see that there's two competing forces, right? In a city like New York, people are no longer going to the office. And so this is predicting a loss from these interactions, externality that makes people earn higher wages going down.

And in fact, wages would be going down quite a bit, but that's compensated by gains from not having to commute, right? Enjoying those amenities and also the option value of switching into in person in the future, if you desire. And so that's kind of why, in the end there's quite a lot of action, but it's kind of a reallocation about where do people get their utility from.

They tend to benefit now more from staying more time, spending more time at home, and willing to accept these lower wages, and that's kind of the main point of that exercise. So let me conclude, what we do in this paper is that we just entertain this idea that is very central in urban economics, that people should benefit from being close to each other.

And that's a very essential point if we wanna think about cities. And so we take it at face value and we ask, what are the consequences of remote work for the workings of these types of interactions, right? And the crucial consequences is that remote work allows us to remain in the same city, participate in the same housing markets, the same consumption markets.

But not necessarily have to interact in person. And that can change, in fact, the way cities could operate, right? And in a sense, could lead to what we see as this coordination problem, where it could have dramatic effects in the number of people who choose to go to the office.

And that new form of multiplicity can only exist if the strength of the remote work technology is productive enough, right? That's what the cones highlighted, that you need that technology to be in exactly the right area. And we find in the data, when we take this to the data, clear evidence that there's very heterogeneous patterns in the recovery across cities.

And that this theory seems to do quite well in explaining that heterogeneity. Okay, I'm gonna stop here, thank you very much.

>> Speaker 3: Thanks this is a very interesting and ambitious paper. First, an observation about the data. The safecraft data seemed to be telling a somewhat different story about the relationship between the increase in work from home or stickiness and density than we see in the sway data.

So in the sway data, across the entire density distribution in the United States, we see higher work from home levels now than what prevailed before the pandemic. Which seems to be different than what's coming out from the safe graph data. I'm not saying one of the other's right or wrong.

That's actually a critical thing for the evaluation of your model. I don't know what's the source of that disconnect between the two data sources. But I would get a less favorable assessment for your interpretation, I think, using the sway data or the HPS. Well, we haven't looked at the HPS.

Then that's just an observation, and I don't know.

>> Charly Porcher: No, thanks, this is definitely something.

>> Speaker 3: Yeah, we're thinking about. Second, in your model, there is a region that doesn't appear to show up in the data, as I understand it, which is the high. The set of cities where the agglomeration economies are so strong that even if you hit them with a pandemic shock, you'll bounce back, okay?

That's a feature of your, is that a possible outcome in your model?

>> Charly Porcher: You hit them with the lockdown and they bounce back.

>> Speaker 3: And they bounce back to the high agglomeration, high density outcome.

>> Charly Porcher: Right, those are not-

>> Speaker 3: small cities in your model. What I don't understand is why, according to your models interpretation of the data, there are some highly dense cities that also bounce back.

Because the agglomeration economies were, I just thought I didn't have, I just didn't understand. And then the third thing is just another observation. If we take your interpretation of the data seriously, then all we need to do in San Francisco or New York is force everybody to come back.

For, say, a year, long enough to overcome the switching costs that were part of your model, and they'll stick there. And I guess the quiet, I don't really believe that, but maybe you do. Well, but to be fair to him, you would need a coordination mechanism across businesses, across people that actually makes them come back.

So it's not clear how you would engineer that. I guess you could do it in DC with federal government employees, that would be the closest.

>> Charly Porcher: Right, yeah, so just to respond very briefly. So, on the part that you mentioned earlier, right? So why is it that the big cities are those that we think are not coming back?

This is really crucial, right? And the idea is the stronger the agglomeration force, the more likely do you have to have this bang bang type of mechanism, right? And so it's really, if you think about it through this bang bang kind of situation wherever, because there's very little frictions to sticking.

That's why you get a stable environment when agglomeration forces Courses are strong, yeah. And on the coordination, I agree with you, right. This model has very stark predictions for what happens if you manage to overcome this coordination problem. In a sense, everything goes away. And so that's something that in the end, if we were to see some examples of successful, coordinated return to the office, that would be a really great test of whether that's a story.

And I know that in DC there's definitely perhaps the city where there's the strongest momentum to try to do that, but I don't think it's happened yet at all. So we'll see. Thanks.

>> Speaker 15: Yeah, I have a follow up comment to this discussion, actually. So you also see different results across cities when you look at different geolocation data providers.

So, placer AI is another data source. And they actually argue that New York, for example, has come back much more in person. And actually it's the Sun Belt cities that have a lot of remote workers sort of consistent with that. You see MTA traffic really has been coming back in New York City.

And if you look at ACS data, you see that New York City actually doesn't have, is not on the top list of remote workers. It's really Raleigh suburbs, right, Atlanta suburbs. Those areas really have a ton of remote workers. So, yeah, I know in your model.

>> Charly Porcher: But just to clarify on the data, it's a really important point and something that maybe I was kind of lost in the way I was describing.

But we're really looking only at visits to the cities center, and a city center is a very narrowly defined area, right. And so in the case of New York City, it's important because if we think about it's mobility to Midtown and the financial district only, right. And so I think there's a lot of mobility that has recovered in the city overall that could be masked by.

 

>> Speaker 15: But that's exactly what's come back in the latest subway data, right. It's really traffic to Midtown, the financial district, that is kind of going up. So I think the broader point about equilibria is kind of well taken, but I think the full story about what region are actually gonna have the most remote work is a little bit more unclear to me.

I can actually imagine a future in which is actually these Sunbelt regions that have very large homes, longer and longer commutes to get to the office. That actually wind up having more remote work in the long run, whereas it's, potentially even the large cities. Because people have small homes might have shorter commutes if they have a public transit and so forth, might actually come back to the office greater amount.

So I think that the full story is yet to be written, and I think it's worth comparing different data providers that might result in different conclusions about how many people are coming back, even to the urban cores.

>> Charly Porcher: Thank you.

Part 7:

>> Speaker 1: All right, well, thank you so much for being here this morning and making it to the first paper and thanks to the conference organizers for having our paper on the program. So this is joint work with Matt Gustafson at Penn State, Dan Weagley, Georgia Tech, and Zihan Ye at Tennessee.

So we know the pandemic changed many different aspects of our life, okay? So these are both financial, real impacts of the pandemic. And as we saw over the last day and a half, we saw a lot of the implications lie at the center of the city, okay? So the municipalities are very important to understanding the overall impact of the pandemic.

We also know at the same time that the shift to remote work greatly influenced migration. So in earlier work with Dan, we showed that there's these work from home capable high income individuals moving across state lines. We've also seen the doughnut effect by other people in this room.

And it really comes to the fact that we can decouple the job from the location, and people are selecting the location where they want to be, okay? At the same time, there's some debate about the permanence of remote work and therefore the migration response. So this reallocation of people and human capital potentially impacts future revenue sources for these municipalities, okay?

This has broader implications for how cities invest moving forward. It has implications for companies as well as these individuals. What we're going to do in this paper is use municipal bonds as a laboratory to better understand the effect of migration on its implications for the city, okay? The nice thing about municipal bonds is that they're forward looking in nature and they have this downside risk sensitivity.

And so we can see that the areas where people were flowing out from should be more affected if it's consistent with this idea that this migration is having a real impact, okay? So there's no real clear prediction for how this is going to impact these local municipalities, okay?

On the benefit of this in migration is that these people tend to be higher income individuals, okay? So that's going to lead to more consumption locally. At the same time, as people move in, there's a greater need for infrastructure, okay? So you have to build new roads, you have to build new schools, you've got to extend sewage lines, right?

And if there's some uncertainty about the persistence of their shock, there's some good theoretical work that if this is just temporary, then these cities will overinvest, which will cause distress for these areas where people are moving into. In this paper, what we ask is we're going to look at the relationship between the COVID induced migration and the municipal bond yields.

So that's our measure of fiscal health here. So how has the COVID induced shock to preferences affected municipal fiscal health? To what extent does this shock represent a downside risk? Are these effects probably for this audience? Are these effects related to the shift to remote work? Just to give you a preview of our results, we do find that migration is more informative predictor of municipal bond spreads during the pandemic, and this is driven largely by downside risk.

Then we utilize the forward looking nature of bonds and provide evidence that this migration shock has affected the longer maturity bonds. Okay, so that gives us some evidence that there's some permanent effect of this remote work shift, and it has the greatest impact on the medium run risks to these municipalities.

We show that the shocks are affected to the broad transition to remote work, so areas that were more exposed to remote work are more sensitive to this migration shock. And something I won't have time to talk about today, but the shock also affected how municipalities were issuing debt.

So the municipalities are also responding to what they're observing in their local communities. Okay, so let me dive into the data here. So for migration data, we're gonna use the USPS change of address from May 2017 to December 2021. Here specifically, we'll be using the permanent change of address.

Okay, so we're going to try to isolate people that were permanently moving in response to the pandemic. We've shown in other work that this is highly correlated with other migration data. The nice thing about the USPS sample is that it's a broad sample. For municipal bond yields, we'll use the MSRB.

For other bond characteristics, we'll use Mergent and Bloomberg data, and then we'll link up other data from the BEA, the BLS, and the Census of Governments, okay? Our final sample ends up being a bond month level trade weighted average yield spread over the maturity matched after tax treasury bond yields.

So we'll account for the fact that municipal bonds aren't taxable, and we'll have several bonds, for each municipality. Next, what we do in the paper is we first show that changes in population during the pandemic are more informative for fiscal health. That's a good litmus test for us.

But ultimately we think that migration is more informative than population changes, particularly in a pandemic, right, where things are happening. So we're going to try to create an unexpected measure of this migration shock during the pandemic, and we'll use per capita inflows to try to measure the intensity of this migration for a given area.

And we want to isolate a very short time period during the pandemic to get this shock such that we're not over controlling for the migration that occurs later. So we don't want a simple mechanical relationship between migration and yield spreads, okay? So we use April to September of 2020 as our measurement of the initial response to the pandemic.

The nice thing there is that the majority of moves in any given year occur between April and September. And we'll use the change from 2020 relative to 2017 to 2019 to get this unexpected measure. Okay, so our final measure is the average monthly net inflow per capita in 2020, less the average monthly net inflow per capita pre pandemic.

Okay, and I'll talk a little bit about this in a minute. But the results are robust. So if you're concerned about the specification, we can specify this several different ways. Okay, so a couple points to make here. First, this is just a picture of the geographic dispersion of our migration shock.

And what you can see is that even within state, there's a lot of heterogeneity. It's not like people are moving to just one particular area. Okay, so I should say that this is a time invariant county measure, given the way that we're measuring it. Okay, but you can also see some of the trends that were highlighted in the news as well, right?

So the lighter regions are going to be more net inflows. So we can see Idaho, Wyoming, Montana tended to have more inflows in areas like southern California, San Francisco, New York had these outflows. So this measure kind of lines up with what we know in general. The other thing I want to discuss is the magnitude of the shock.

What I think is really interesting and what really drew me to working on this topic is the extent to which we saw this migration shock. So other events, and historically, we've seen kind of moving from one place to another, but this is a very widespread shock, okay. So to get a sense of that, what we're going to do is we're going to look at migration changes year over year.

So 2017 to 2018, 2018 to 2019, and we're going to take the absolute value of that shock and compare it to what we saw during the pandemic. So this black line is the CDF of that shock. And what you can see is that the dispersion of migration is very different than what we saw before.

Okay, so this is really quite a large shock in terms of its Its magnitude, but the dispersion is also quite large here, okay? A quick overview of some of the summary statistics, so our final sample ends up being 3 million bond month observations. If we think about it as kind of a closed system in the US during the pandemic, the average migration shock sits around zero.

So that kind of jives with what we should expect in terms of prior to, we'll end up using kind of a diff and diff setting here, okay. So areas that experience a high shock versus a low shock and show that over time. And so you might be concerned that these areas are just fundamentally different.

In terms of the risk, in the municipal bond market, they're very similar, okay? So there's a slight difference in the spread, but in terms of how these bonds are rated, they're very similar, okay? So that's saying ex-ante, and this rating and spread are measured prior to the pandemic.

So ex-ante is areas in terms of their risk look very similar, okay? What we can also show is that consistent with other papers in this literature, is that people were moving out of expensive areas. And out of these urban centers that had higher employment growth prior to the pandemic, okay?

So it's all kind of consistent with what we've seen in the literature. So our research design is going to test for the effects of municipal bond yields on these migration patterns, we'll control for the usual predictors. So this is going to control for bond characteristics as well as local economic effects and the economic shocks.

So we're trying to really hone in on this migration effect, okay? We'll use a Poisson estimator here, so luckily we got a little preview of that yesterday, this is gonna deal with the skewed nature of spreads in the data. It's consistent on bias and efficient, okay? I can show you as well, it doesn't matter if we use regular OLS and use log spreads, the results are consistent, okay?

Here's our main specification, this is the dynamic setup that we've got. So spread at the bond county month level on the left-hand side. On the right-hand side, we'll have a monthly indicator interacted with that migration shock that we calculated, okay? We've got a set of controls here and we'll standardize this migration shock just for easier interpretation.

Importantly, we include QSIP fixed effects, so that's gonna be bond level fixed effects. So that'll take care of the differences across bonds and we'll have time fix effects as well. So the bond market, the municipal bond market in particular, experienced kind of an initial shock when the lockdowns happened, and we'll deal with that a couple different ways.

And ultimately, we cluster at the county in year month, okay? So this is kind of our first main result here. What we're doing is we're plotting that interaction between the bond month and the migration shock over time, okay? Prior to the pandemic, you can't see the colors here, but prior to the pandemic, there's really no difference after controlling for our set of controls and fixed effects, okay.

The gray area is the start of 2020, and then the blue area is our period of measurement, okay? Over that period, we see that these areas that were more negatively hit had higher spreads, okay? Importantly, though, after our measurement period, what you see is a slight attenuation in that effect, but ultimately a strong and persistent effect through the end of our sample okay?

This is consistent with the idea that this migration shock had a real impact on the municipalities. Next, we move to a static setup. This is just gonna be an indicator for the COVID period. Importantly, what we do is we omit the period January to September of 2020 when we are measuring that migration shock.

So this is not gonna be a mechanical relationship here. And what I wanna say about this is when we put it in the setup and then translate that into economic magnitudes. Is that we find about a six to nine basis point increase for every one standard deviation in this migration shock, okay?

So just if you're not familiar with the municipal bond literature, this is consistent with other predictors of municipal bonds, okay? So this is a real tangible effect that these counties are facing. When we think about the size of the municipal bond market, this is also going to increase the cost for these municipalities issuing debt, okay?

So they're going to have to incur this cost. So the COVID induced flow is net inflows per capita, okay? So the areas where people were flowing out of is negative, resulting in higher spreads, okay? Or greater risk or worse, fiscal health. So, as I said, we'll talk a little bit about the robustness here.

We can measure this several different ways, okay? So if we include all of the data using all the migration through the end of 2021, the results hold are a bit stronger. As you might expect, we can use 2019 as the only year of comparison, so if you think that things were trending differently and we're kind of mismeasuring on that front, it doesn't affect our results.

We can also follow exactly how Romani and Bloom measured migration, and the results hold. We also do a lot of additional robustness here. So we can exclude different bond characteristics, for example, callable bonds. So if you think that this is something about the bond market and the bond characteristics, our results go through, obviously, because we're measuring it at the bond level.

New York City, LA, San Francisco will be overweighted in our sample, we can exclude them, and our results still go through. A different variation of that is to subset on counties below mean population, so this is not just an urban effect, okay? Of these largest cities, we can reweight this sample.

We've also included controls for COVID-19 cases, the stringency of lockdowns, the general pandemic mobility, which is related to those restrictions or pre COVID immigration rates. So during the pandemic, there was a lack of in migration, and so areas that were more urban, that had more migration inflows in the pre pandemic might be more exposed.

We can control for that, our results still go through, okay? We also controlled for ratings in a variety of different ways, and so this is not purely about the ratings of these bonds. We can look within the ratings and show that these outflow areas experience worse fiscal health moving forward okay?

The next question is who should be most affected, okay? When we're thinking about debt instruments, the sensitivity is really going to be about the downside. We know municipal debt is notoriously unlikely to default, okay? So should they be more sensitive, okay? What we're gonna do in columns one and two is we're going to include county year month fixed effects, okay?

So that's controlling for everything at the county level in a time varying sense. And we're gonna take two bonds, one that was more highly rated and one that was more lowly rated or had a lower rating thus it was exposed to more risk. And we find the sensitivity, even within a county, is these bonds that were more exposed to risk are more sensitive.

So it's consistent with the idea that this migration shock is really affecting the downside risk. Alternatively, in columns three and four, what we do is we take this shock and we put it into quartiles and we ask. Okay, if this is really about the migration shock, then the areas that had the most net negative inflows, okay, or the areas that were experienced the most outflow should be most sensitive.

And we can see a clear monotonic relationship there, okay? So comparing the areas that had large inflows with the ones that had large outflows, that's about an 18 basis point increase, okay? So this is a very significant increase in their funding costs, okay? So we use borrow a structural model from Goldsmith, Pinkham et al just to give some sense of the magnitudes, right?

So if we're thinking about the price of a bond, it's both about the risk. As well as the expected cash flows. If we hold the uncertainty fix, this implies a decline in cash flows of about 2% to 4%. Alternatively, if we hold cash flows fixed, this implies an increase in risk between 1% and 1.5%.

The nice thing about municipal bonds, and the reason that we're using it is not only are they forward looking, but they have a term structure to them. You have bonds that mature at different periods. To get some sense of the permanence of this is we can utilize that term structure.

And when we break it down, using some methodology from the municipal bond literature, we find that the greatest risk is really in this five to ten year period. So what that's saying is that there's a lot of uncertainty in both the cash flows and the general risk in that period.

So this is going to be a long term event, at least as the municipal bond market views it. The part that this audience is probably most interested in is this idea of the permanent shift to remote work. We're going to tie into that literature. Obviously, several people here have provided some good evidence for us.

What we want to do is we want to ask, if this is really about remote work, then the migration that we observe should be more sensitive to areas that were more exposed to remote work. So we're going to use that measure of work from home capability and show that in areas that were more exposed to remote work, that sensitivity is about 40% higher.

The other nice thing here is what we can see is that areas that were more exposed to remote work had higher spreads as well, about the same size of a one standard deviation in its migration shock. These areas that were exposed to remote work experienced greater risk. In columns three and four, what we're going to do is utilize an interesting aspect of the municipal bond market.

There are revenue bonds and these revenue bonds, their ability to repay that bond is tied to the income source. We're going to use transportation linked bonds. So we can think about tolls, bridges. If this is really about remote work, people are not using roads and bridges as much.

Well, they should be more sensitive. So these results are really consistent with work from home or remote work capability being more exposed to the shock, and that the migration that follows or ensues after the pandemic increases that risk for them. So when people are both capable and actually do move, we experience worse fiscal health.

Let me conclude, the pandemic altered many aspects of life and the ensuing migration patterns we think are really a revealed preferences of these households. When we decouple the job from the location, we find that municipal fiscal health is correlated with reallocation. This is most prominent for counties and bonds experiencing outflows.

The forward looking nature suggests that there is a long term effect here. And so that's consistent with this idea that remote work is really having a long term impact on these municipalities. Overall, we think the analysis informs our future expectations about migration patterns and its real impact on municipalities, which has obvious implications for employees, employers, investors, city planners, municipal bonds.

Obviously these implications were seeing a lot of it. I appreciate any comments you might have.

>> Audience: So one thing you might do, I think this is a really nice piece of work, but one thing you might explore concerns the question of, in the case of outflows, where did the people go?

And so there are two cases, they move to the suburbs. Oftentimes, they'd be in the same county, but other times not. The suburbs may be a different county. Alternatively, they could have moved totally out of the metro area. And so it's possible that a move, population loss that stays within the metro area has less damaging effect to the county relative to one that population loss, that means migration out of the metro area.

 

>> Speaker 1: I totally agree.

>> Audience: Then you'd have to track where, origin, destination for the flows.

>> Speaker 1: Totally, and I totally agree. And we're not going to be able to capture any of this within county stuff given what we're doing. So the USPS data will not allow us. That's just a simple data limitation on that front.

But we do have data from a moving company. Most of those moves are going to be across state lines. There's also the IRS panel that we could play around with as well. I guess from county to county.

>> Audience: My understanding is that postal service data shows you that origin, destination, zip codes, doesn't it?

 

>> Speaker 1: Yes.

>> Audience: The other data has the cross-section by month, and you can add them together. So if I see John Smith is here on this month and John Smith is somewhere else, so I know they should be the same. They don't totally align. We're trying to figure out from data that they're highly correlated.

I mean, the t stats like a million, but the r squared is 0.5. But you're using, I guess, the USPS stuff.

>> Speaker 1: Yeah, the Fourier data.

>> Audience: Which they look pretty similar, but in the second you can look at where they're going to. Well, I mean, we use this stuff in our AJ paper, and I guess I'm just revealing the fact that I was not the data guy.

 

>> Speaker 1: I think that's interesting. At least from the interstate level, we could kind of compare it, juxtapose it against what we see using USPS.

>> Audience: Thanks. Really interesting. One of the massive shifts in kind of local public finance was the role of intergovernmental grants. Whether it was through PPP or Cares, some of these things we know they were correlated with population or income and kind of other observables.

And there were also grants that were coming from the states to the local governments. The highest level of grants we've seen in a long time. Do we know anything about how those are correlated with these migration measures, or to what extent? Are we worried about that as a confounder picking up in the municipal bond markets?

 

>> Speaker 1: I think that there's some reason. The nice thing is that the money tended to flow into higher population areas. If anything, that was kind of counteracting our effects. Obviously, those were areas with negative outflows and they saw a drop or an increase in their spreads. And so if anything, it's a mitigating factor, but we could definitely do more to kind of account for the flow of funds.

 

>> Audience: Two related comments. First, I was happy to see that you looked at the interaction effect between credit rating and response, but in one sense, you over controlled when you did that because you swept out the between county variation over time, the movement over time within counties. You're only capturing the differences in credit quality, credit rating across bonds in the same county that moves in a different way over time.

Which is fine if you're trying to make the case that credit quality matters, but if you're trying to characterize the role of the pre-pandemic fiscal health in the municipality on its response, that's a bit over controlling.

>> Speaker 1: I totally agree, and I should be very clear. Our main specification does not control for ratings, particularly for this point.

 

>> Audience: Your main specification doesn't get the interaction effects.

>> Speaker 1: That's true.

>> Audience: It's controlling for level effects.

>> Speaker 1: That's true.

>> Audience: Anyway, that leads to my second comment, which is. You could do a fuller job of characterizing the shifts in, say, a duration adjusted yield across municipalities. And there's really, there seem to be three main things happening.

There's the size of the migration shock, there's the exposure to work from home, and there's the pre-fiscal health of the municipality. And you can think about characterizing the shift in the entire distribution. Much of your talk focused on conditional means, but as you made clear, we don't expect the conditional means to capture the full force of what's happening.

So I'd encourage you to show us what happened to the entire distribution of municipal bond yields, perhaps duration adjusted, as a function of the three things I just mentioned. And the reason that's interesting, is because that presumably is a guide to what's happened to new funding costs. Or as a function of these three sets of characteristics, and just show us that entire distributional shift in a picture, that'd be really useful.

 

>> Speaker 1: Yeah, I think that's a very nice suggestion.

>> Audience: That was a fantastic paper, very cleanly done, one thing I'd love to see, which I think is also something linked to the next paper, would be running all the way up to the present day. Basically, what you have is, as you point out, a kind of an insight on how the markets thought how permanent this thing would be.

And so it would be great to take the final, imagine you had perfect foresight, so take the final total migration, and just regress it month by month up to 2023. Yeah. Cuz I think, you kind of went up to the end of 2021, and it looks like it took a while for them to, I wasn't quite sure what was changing, but in my mind, Gina, I think she's here.

But yesterday, on Wednesday I had this comment like, what we don't know it's permanent. And it's really interesting still, and hearing that, and I'm not sure when the market went from thing, this is temporary to permanent. So my guess is it takes a while for it to price in, and then whether it's priced in, more or less, I wasn't sure what else, but you have some long duration bonds, so therefore the duration isn't changing that much.

 

>> Speaker 1: Yeah, the duration, I mean, the term structure, we can get some of that, but the evolution of the expectations over time, it's partly a data limitation. So just recently, they came with the first half of 2022, and so we're just kind of waiting for more data. But it's certainly something that we need to follow up on.

 

>> Audience: It is a perfect window to look at how long it took for people to figure this out. And I agree with each, Steve, maybe you could look at some, like low grade, some of the more sensitive things. Just see, it's fascinating to see how long it took people to figure this out, my sense is it's still not totally agreed upon.

 

>> Speaker 1: Yeah.

>> Audience: So I guess you'd still see pricing effects coming.

>> Speaker 1: Yeah, and we can, with more time, we can also look at the issuance behavior a little bit more closely, so I totally agree, it'll be something.

>> Audience: The other point, sorry, it's just to state the obvious, but it makes it worse for these cities.

 

>> Speaker 1: Yeah.

>> Audience: It's the obvious thing, but, actually oddly enough, last night I was watching some interview on Wall Street Journal with London Breed, and it was like all these troubles, on top of it now the financing costs have come up.

>> Speaker 1: Yeah, and so, some other work has basically shown that when funding costs go up, services get cut, and when services get cut, people move out, and so you get into this kind of perpetual motion, so.

 

>> Audience: This is a very nice work, so you mentioned that municipal bonds are unlikely to default. So I was wondering from the capital supplies perspective, is the liquidity risk playing a role here? So think about that, counties that are affected by the negative flows, then investors might actually, unwilling to provide capital for the, either secondary market bond, traded bonds, or new issuance.

 

>> Speaker 1: So we haven't used liquidity on the left hand side, to understand how the effects on liquidity itself. I can tell you we control for liquidity and time varying in nature, and so this is kind of over and above that. But I agree that understanding the liquidity effects could be interesting in and of itself.

 

>> Audience: Sorry, some of these markets, who's actually buying and selling? Do you have any idea who it is that is buying and selling?

>> Speaker 1: So we have the trace data, we haven't torn into it, what we've done is classified this by institutional, given the size of the trade, classified it into retail institutional, and it shows up in both.

But there's different dynamics, but we can.

>> Audience: Do you know the name of the institution?

>> Speaker 1: Yeah, so the trace data would allow us to identify the data I have isn't anonymized, but we know if it's.

>> Audience: The reason I ask is, we're talking over dinner last night, and someone was saying, some of the investment banks that would trade this a lot had pressure from the CEO's, to return back and push the line that were returning to the office.

I don't know if, say, Goldman's were particularly optimistic.

>> Speaker 1: Yeah.

>> Audience: Not because they were optimistic, but because David Solomon was pushing them to return back, and they're told, you've got to trade the party line. I don't know whether the CEO's position on their own firm would affect trading here.

 

>> Speaker 1: That would be very interesting to look at, that's, yeah, you can get that through the Fed, so try to reach out. Well, thank you all so much, appreciate it.

>> Speaker 3: Okay, so now I've got Xiaofei Zhao, who's gonna talk about ICT availability and asset prices.

>> Xiaofei Zhao: Right, good morning, everyone.

First of all, let me thank the organizers for including our paper in this very interesting program. So this is joint work with Jing Gao and Xiaoji Lin, who are also in the audience today, and two other co-authors are Jack Favilukis, from UBC, and Ali Sharifkhani, from Northeastern. So let me first of give you a little bit of motivation on the paper, and we've seen a lot of discussions already in the past couple of days, that remote work has gained more popularity since pandemic.

And it was also very interesting to see the interview on Wednesday. Nick's interview for Jack Nilles` 1975 paper basically decoyed the concept of telework, that is decentralize the work activities at a traditional workplace. And a lot of people in the literature also have recognized the importance of ICT for enabling the work at different locations.

Now, what we recognize in this paper is that we try to say, more fundamentally, is the human capital associated with ICT, that's crucial for maintaining productions at relatively high productivity level during emergency situations, such as the pandemic. That's really the key point we try to make in the paper, right?

So we are going to leverage on this pandemic setting, to study how ICT human capital can affect work from home policies and enterprises throughout the pandemic. So empirically, what we find is that, ICT human capital is crucial for telework flexibility in the labor force and also for the enterprise dynamics during a pandemic.

So we are going to provide measures of ICT human capital, and we have three key findings in the paper. First of all, we find that industries with high ICT human capital would have more persistently, high work performed practices during and throughout the pandemic, okay? Secondly, we find that industries with high ICT human capital, they experience more employment growth and also hours growth throughout the pandemic.

And lastly, we find that as the prices of the high ICT human capital industries significantly outperform the industries with low ICT human capital, okay? So these are the three main key findings in the paper. And then we provide a dynamic model with multiple job tasks to sort of understand the findings.

So there are really two key features in the model. One is that the output is gonna be produced using in person tasks and a flexible task, okay? And the flexible task requires ICT human capital, and it has an option to switch between onsite work and telework, okay. And the other important heterogeneity across industries is that different firms or industries, they can have different ICT human capital share in their production.

And that's gonna generate the spread across different industries for the outcomes we find in the data. So that's basically what we do in this paper. Now, let me first of all talk about how we measure ICT human capital, right. We use two industry level public datasets, the first one is the O*NET that most of us know about.

So this is really helping us identify the job attributes that are related to ICT, so that we can define what are ICT jobs, right. And then we can link that to the industry composition in order to measure the ICT human capital, okay. So let me start with talking about what job attributes we use here, right because ICT is essential for telework, so we look at five job attributes here.

Two of them are related to knowledge requirements for those jobs, such as computer and telecommunications, right. Also, we look at these skills such as programming, and we also look at the importance of work activities such as interacting with computers, analyzing data information. So these are the five job attributes we look at, and we know the importance of those attributes for different occupations.

Then what we do is we aggregate those five job attributes up to the job level. So now we have a job level ICT score, okay? And here I'm just showing you the jobs with the highest ICT scores and then jobs with the lowest ICT scores, which is, these are quite intuitive, right?

For example, jobs related to computer software network, these are sort of a high ICT job score jobs. And then models fallers, I don't know what is a faller here, but helpers, for example, these are more in person jobs that have quite low ICT scores. Now, once we have the job level ICT scores, we need to define ICT jobs.

Okay, so there are different ways of doing that. The baseline definition we have here is we basically classify the ICT jobs with the top 10% ICT scores, highest ICT scores as ICT jobs, okay? And of course, we run different robust checks to make sure results are not very sensitive to this cutoff.

We look at 5%, 15%, and the results I'm gonna show you later are quite robust to this definition. Alternatively, we can also use ICT jobs defined by other people. For example, there's a paper by Tambe and coauthors, they identify IT jobs. We can use their job list as well, the results are pretty similar.

So once we have the definition for ICT jobs, then we can look at industry labor composition. We know for each industry how many people they are hiring for those ICT jobs, we can capture their labor expenditures and then we're gonna capitalize those labor expenditures for those related, those ICT jobs.

And we scale by the total employment in those industries, that's basically our ICT human capital intensity measure, okay. Now, here I'm just showing you some examples of industries. Again, the left panel is based on our ICT jobs defined by the job attributes from the O*NET. The right panel is from the Tambe et al the other paper job list.

Again, it's not very surprising the industry is associated with the computer software data, these are high ICT human capital jobs. Whereas industries such as personal care services, restaurants and eating places, these are sort of the industries with the lowest ICT human capital, okay. Now, with those measures, then we can conduct empirical analysis, right.

So in our paper, we have three sets of analysis. First set of analysis looks at the enterprises dynamics. So we are gonna use the public traded stock returns and public traded corporate bond returns, so the data source are pretty standard. Secondly, which I want to emphasize here, is the workflow measures.

We're gonna use two sets of workflow measures. The first set of measure is three very commonly used workflow measures in the literature. These are based on pre-pandemic data. Remember, our ICT human capital measure calculated based on the data up to the end of 2019, so it's also based on pre-pandemic data.

So we want to understand, to what extent the ICT human capital will be driving the variations in the commonly used workflow measures, okay? And then we also use the second set of survey based practices of work from home during the pandemic. And thanks to three of the organizers, Jose, Nick and Steve made this data publicly available so that we are able to run this analysis.

So we showed the dynamics of work from practices, how that vary with the ICT human capital, okay. And then we also use the BLS data to look at the hours and employment growth. Now, the industry we are defining here is basically 4-digit NAICS code industry. And when we refer to sector, it's basically 2-digit NAICS code, okay?

Now, let's first of a look at the pre-pandemic analysis, right. So here, basically, again, as I mentioned, we are gonna check to what extent the commonly used workflow measures based on pre-pandemic data are driven by ICT human capital, right. So there are 3 measures here, I think the most famous one might be the one by Dingel and Neiman 2020, which is also based on O*NET.

And there are two other measures based on American time use survey. So the table here basically shows a very simple regression. The dependent variable are those workflow measures. The independent variable is ICT human capital, right. There are two takeaways here, first of all, you see the positive coefficient here, meaning that higher the ICT human capital, the more likely you can work from home, more feasibility you have.

Secondly, if you look at the R square, it's actually quite significant, right. Those R square will translate into univariate correlations between 0.6 and 0.7. So that means the existing commonly used workflow measures, large part of variation is driven by ICT human capital. So that's the pre-pandemic analysis. Now, let me show you two sets of analysis during the pandemic.

I'm gonna show you how the asset prices are predicted to vary with the ICT human capital and also how the labor policy will vary across human capital, okay. So let me show you the graphs here. So I have two graphs, the top panel is basically shows you the dynamics of accumulated stock returns from January 2020 up to May 2020, okay.

Bottom panel shows you the accumulated corporate bond returns between January and May 2020. In other words, this shows you how the asset prices move during the height of the pandemic. Now, we are gonna group the firm or industries into three groups. Okay, you can see three lines in both plots, right.

The top blue line is basically the group with the highest ICT human capital. Middle red line is the median ICT human capital group. And the bottom green line is the industry group with the lowest ICT human capital. It's quite clear here the industries with the highest ICT human capital outperform industries with lowest ICT human capital, and the spread is quite significant.

And of course, we're gonna run some formal tests to show the differences. But this shows you how ICT human capital can affect the asset prices during a pandemic. Let me use this one table as an example to show the economic magnitude. So here, what we do is we use for each industry, we're gonna calculate the cumulative.

Terms in the US between January and May 2020, so one industry has one observation here, so it's one cross section, okay? We regret in column 1, we regress the cumulative return, on a high ICT human capital we measure at the end of 2019, you can see the coefficients is positive and statistically significant.

The magnitude would be one standard deviation increase in ICT human capital, correspond to 6% increase in story terms in these five months window, but that's quite sizable, right? Second point here is that, for the workflow measures, we do a decomposition, right? We basically regress workflow measures on the ICT human capital, so that we have ICT human capital component, plus a residual component.

And then we can put these two components in regressions, 2, 3, 4, for the three different workflow measures, right? Again, you can see here, the way the workflow measures can predict returns is largely driven by, the ICT human capital. Now, residual component the predictability, varies cost measures, and also across asset classes as I will show later okay?

So that's the US stock return, now we also extend these stock returns to other countries, in other G7 countries. So, we have the ICT human capital measure for different industries in the US, we can map those to other G7 countries. We can run the same analysis, the results are pretty similar, even in terms of economic magnitude it's similar to the US, okay?

And now moving on to corporate boundary terms, we run the same analysis. We look at the cumulative returns between January and May, regress that on ICT human capital we can see, one standard deviation. Increasing ICT human capital is gonna correspond to a pretty large 4% increase in corporate bond returns during this five-month spindle okay?

So, this is the first set of empirical findings I have, now let's look at, how ICT human capital impacts the labor policies such as the employment growth and hours growth. Again, the setup is similar to what we see for the asset prices, but instead of asset prices we have in the top panel, the cumulative employment growth, across three ICT human capital groups.

In the bottom panel is the hours growth, between January and May 2020 across these three groups. Again, you can see here, high ICT human capital groups, they experience more significant employment hours growth. Or put it differently, they experience less job, in employment and our hours growth, whereas the low ICD human capital group, they experience more job in employment and hours growth, okay?

Again, we can take this to the regression analysis, the results would be basically what you see in the data. There's a positive correlation between the ICT human capital, and the employment and hours growth. We also look at the weekly earnings growth, once we have sector fixed effects that relationship is insignificant.

So that's the second set of empirical finance we have, let me look at the post pandemic analysis. Now here I'm gonna focus on, how work from home practices, is gonna vary across the ICT human capital group, okay? So, this is basically the data based on the survey data, right?

So, this shows the, how workflow practices varying across the ICT human capital group between June 2020 up to very recent 2023, okay? So, there are two takeaways right here, first of all you can see clearly there's a big spread here, right? ICT human groups, they have more work from home practices, whereas lowest ICT human group they have to have a lower.

Now, the second takeaway from this graph here is that if you look at during pandemic, for example, the dash line here shows the end of 2021. If we define that as pre pandemic period, compared pre, we supposed we can see here, for high ICT human group. The workroom practice is actually pretty persistent at a high level, whereas the low ICT human group, there's a trending down, right?

For low ICD human group, probably the benefit of workflow is getting lower, or the cost is getting higher, that's why the practices goes down. So, if you look at the cross sectional spread actually, the post pandemic, the spread in work from home practices is gonna be larger, for the post pandemic period, okay?

So that's really the key, message we like to show in this set of analysis, now we can then run regressions. For example, let me just show the column 1 number here, we regress the work of home, a survey-based measure on the ICT human capital. The economic magnitude will suggest one standard deviation increase in ICT human capital, corresponds to 7% increase in remote work.

And keep in mind, the average percentage of remote work is only 35% in the sample period, so that's quite large. And then secondly, sorry, the other columns just shows where the variation is coming from, we've put in different fixed effects. It turns out the sector level, variation is the most important for this relationship, and we also run the post compare post pre, as I showed you on the plot, right?

The post really cross section relationship is stronger here, and that's really due to the high ICT capital human capital group, has a persistently high work from practices, whereas low is trending down. That's why the spread is widening post the pandemic, now we can also look at the dynamics of the asset prices.

Now for asset prices is a bit less clear in the sense that the top panel is the stock returns up to now, bottom panel is the boundary turns up to now, okay? During the pandemic analysis, we focused on January and May, the effect was the strongest, but if you look, for example, focus on the bottom panel here.

If you look at the bond return dynamics, you can see they are converging, when economies are recovering, then the shock of the pandemic is mitigated. Therefore, the different groups, they are converging, whereas for stock returns, it's less clear because there are other things going on. For example, the AI shock can create a difference here, but that's less related to the dynamic shock.

If you look at the employment hours growth dynamics, we can see similar patterns, right? You can see still it's quite persistent, high ICT human capital group, they have persistently high accumulative growth in hours and employment. But apparently the spread between these two groups is gonna be, becoming narrower across time, right?

These are the empirical findings now let me show you, summarize the key ingredients in the model. And I'll show you a few figures on, how the policy of work informed works and also some impulse responses here. So, we're gonna build a dynamic form model with two job tasks as I mentioned earlier, so there's an in-person task that requires labor, but you can only do that on site, okay?

There's a flexible task, that requires ICT human capital, and also physical capital, but here you have an option for onsite work, versus telework, okay? So, there's a benefit depending on the state of the world, there could be a benefit productivity gain for telework, but there's a cost. If you wanna switch from onsite work to daily work, there's a cost, so there's a tradeoff here, right?

So, the key, as I mentioned also at the beginning of the presentation, the key heterogeneities in the model are the ICT human capital share across different firms. And also, the task share across different firms or industry groups, okay? Even the remote cost we can think of the remote cost can vary over time, as firms adopt to the telework.

There are three aggregate shocks that are important, in generating the right quantitative responses, these include the demand shock, the labor party shock. Is very crucial because during the pandemic, the in-person work cannot be done as effectively as pre-pandemic, so that's a very big productive shock here. And there's also an important uncertainty shock here, in the sense that during pandemic we didn't know how long it would take the pandemic to over.

So that uncertainty is quite important to generate the persistent large effect here, okay? Now, the key policy here is the work adoption policy, it's determined by this relationship. Basically you compare, if I switch to remote work, there could be a productive game, especially during the pandemic, because the televote can sort of maintain relative high productivity.

But there's a cost, so it's really after the cost, whether that outweighs the onsite work output. And that would determine the sort of whether you adopt the remote work policy. So let me show you one heterogeneity in terms of the policy response here. So there are two plots here, the y-axis is the work from home adoption, 0 means that you don't adopt, 1 means that you adopt, okay?

The x-axis means the ICT human capital share, for example, you focus on the red line here. When ICT human capital share is low, the benefit of telework is relatively low, because you don't have that much share you can telework, therefore you don't adopt because there's a fixed cost to adopting, right?

So you don't adopt the telework, but when the ICT share is high enough, there's large enough benefit of teleworking, outweighs the cost. Then you switch, and you can see that this is really the key cross-section in terms of ICT human capital, how that affects the workflow dynamics, okay?

Now let me show you something on the remote work adoption dynamics. This is something that we are still working on, that shows three panels of work from policies here. The left panel basically is pre-pandemic, middle panel is during pandemic, and then the right panel is post pandemic, okay?

So how do we read this 3D graph? On the flat face, there's physical capital and ICT capital, the vertical axis is basically 0,1, work from home adoption policy. You can see before pandemic, the adoption of remote work is quite low, of course, in this set of calibration because the cost, the productivity gain is not as big, because there's a big cost of switching to telework.

However, when you move to pandemic, there's a big shock in productivity, in the labor productivity due to the pandemic, therefore, you have an incentive, right? If you switch to telework, there's a big productivity gain and that far outweighs these adoption cost. Therefore, a lot of firms they will adopt, you can see the yellow shady area, that means that a lot of firms adopting to work from policy, right?

And that's the case, when the ICT capital is high enough, then you start adopting that, okay? Post pandemic, you can see the yellow shady area is narrowed. This is because now post pandemic, the productivity gain of telework and onsite work is much lower because there's no pandemic anymore, right?

But the cost is still high, therefore the work from home policy would be lower than during pandemic. However, if you compare the post with pre, there's still more sort of persistent work from home policy here, is because now the adoption cost is lower. So there are still firms, when they have a high enough ICT human capital, they are willing to adopt the work from home policy here, okay?

So that's the dynamic of remote work adoption that can be generated in the model. So now let me show you the, sort of impulse response of how the real quantities such as output, labor, firm value, and remote work policies react to the shock of pandemics, okay? Here we simulate this for three groups of firms, the red lines represent the high ICT human capital group, blue lines, median, and then black line is for the low.

As you can see here, right? The high ICT human capital is better able to maintain the output labor policy from value, because they are able to adopt more remote work policy. Whereas low, they get a bigger shock, therefore the values drop small, okay? So now let me conclude, right?

What we do in this paper is we measure ICT human capital based on job attributes, okay? Then we show that ICT human capital is critical for telework policy and also the impact on asset prices. Going forward, I think this phenomenon is likely to stay, as we can see, how ICT human group has a very persistently high work from home policies.

And the empirical findings are consistent with the dynamic model for that, right? And these figures basically summarize them in findings as you can see, all right, thank you very much.

>> Audience: Okay, great, thanks, really interesting paper. I think there's a subtle confounder that's omitted from your empirical analysis.

 

>> Xiaofei Zhao: Mm-hm.

>> Audience: That is leading you to overstate the impact of ICT human capital.

>> Xiaofei Zhao: Okay.

>> Audience: Probably by a small amount, we should check, so let me explain.

>> Xiaofei Zhao: Sure.

>> Audience: So the ability to work effectively in a remote capacity depends on the quality of your Internet connection.

 

>> Xiaofei Zhao: Sure.

>> Audience: Okay?

>> Xiaofei Zhao: Mm-hm.

>> Audience: Jose, Nick, and I have a paper that documents that and quantifies that effect. It's not a surprising effect, but bottom line is roughly, as I recall, roughly a fifth of the US workforce has poor quality Internet connection.

>> Xiaofei Zhao: I see.

>> Audience: And the paper also finds that that seriously undercuts their productivity in work from home mode, so that's kind of the starting point.

 

>> Xiaofei Zhao: Yeah.

>> Audience: Now think about where ICT jobs tend to be concentrated.

>> Xiaofei Zhao: Mm-hm.

>> Audience: They tend to be concentrated in areas with better Internet connection.

>> Xiaofei Zhao: I see.

>> Audience: So I think part of the differential that you're finding across these ITC, what you call these ICT groups.

 

>> Xiaofei Zhao: Mm-hm.

>> Audience: Is probably coming from the fact that their workers have differential quality of Internet connection at home.

>> Xiaofei Zhao: I see.

>> Audience: So that's the nature of the confounder.

>> Xiaofei Zhao: Sure.

>> Audience: Now you can control for that.

>> Xiaofei Zhao: That`s right, yeah.

>> Audience: And include interactions because if you go to this, Nick, Jose, and I have a paper in an Aspen Institute conference volume.

 

>> Xiaofei Zhao: Okay.

>> Audience: That measures Internet access quality at the worker level.

>> Xiaofei Zhao: A worker level?

>> Audience: Yeah, at the worker level, you can take those data and aggregate it up to industry.

>> Xiaofei Zhao: Yeah.

>> Audience: And construct measures of worker access quality to the Internet.

>> Xiaofei Zhao: I see.

>> Audience: For your ICT.

 

>> Xiaofei Zhao: Yeah.

>> Audience: And then redo your analysis, I suspect you'll get modestly smaller effects.

>> Xiaofei Zhao: I see.

>> Audience: Then you're currently getting because of that confounder.

>> Xiaofei Zhao: Yeah, I think that's a very good idea, basically do another cut, right? You can think of this as like a double sorting basically.

 

>> Audience: You can do double sorting.

>> Xiaofei Zhao: Yeah.

>> Audience: If you have enough data, that would be even better because then you could isolate each one. But if you don't have enough data to do the double sort, just.

>> Xiaofei Zhao: Cuz we can clearly generate more variation, in the sense that we can construct the measures at the industry at state level, that can potentially give us more variation.

 

>> Audience: Yeah, we'll see if there's enough variation, which I don't know. But I think it's just worth exploring that connection.

>> Xiaofei Zhao: Yeah, and the data again is publicly available now, right?

>> Audience: Yeah.

>> Xiaofei Zhao: Okay.

>> Audience: If you have any trouble, just ask Jose.

>> Xiaofei Zhao: Thank you very much, yeah.

 

>> Audience: So there's something that's still not clear to me after listening to the entire talk. And the issue is, are the ICT people facilitating work from home for other workers, or are they themselves working from home? And I don't know what the paper says, but I think that bears clarification.

 

>> Xiaofei Zhao: Sure, that's a good point, I think in. In our paper, for example, in the model, it's clear these are the workers, they work from home, but there could be complementary facts, right? When they work from home, they can potentially, but we don't explicitly measure that. In the data we might be able to do something along the line because we know the ICT attributes of audit occupations.

So potentially, cuz we now just define a subset of jobs as ICT jobs, right? So if we can somehow link those to the other, we can also quantify say, even if you're not ICT job, what are the contribution, right? Of the ICT to your job, in terms of importance of the ICT scores?

I think you can come up with some measure like that and then.

>> Xiaofei Zhao: Yeah, those industries, they rely on those capital to produce output, that's correct, yeah.

>> Audience: In the model we have two tasks, the second task has ICT human capital, and these guys can actually choose to either telework or onsite work.

So these guys actually are doing that, however, we have two tasks, and then the first task also requires labor. So think about those guys as unskilled workers, so depending on the complementarity and substitutability between two tasks. Then if the telework guys, ICT guys, choose to work at home, presumably if two tasks are more complementary, it's gonna help the in person guys to keep their productivity relatively high, okay?

Does that make sense? I think you could push the model a little bit and understand firms that deviate from the optimal strategy as you guys have kind of defined it. And that would shed some light on whether the management practice of allowing people to work from home is providing value or not.

So, again, kind of comparing firms that have kind of optimal work from home levels versus those that deviate from the optimal strategy and to see how their asset prices and employment evolve. Does that make sense?

>> Xiaofei Zhao: In the model?

>> Audience: Take the model but allow it to define firms that have the ability to work for remote, but have management practices that don't allow it.

 

>> Xiaofei Zhao: I see.

>> Audience: Cuz we've had this back and forth about.

>> Xiaofei Zhao: Sure.

>> Audience: Yeah.

>> Xiaofei Zhao: In the model, I mean that would be clear because if optimal policy is to work from home but you don't, then output would be lower, right?

>> Audience: Yeah. Firms in the industries with high ICT that should be working from home and compared to the firms that have low uptake of work from home, like the high ICT.

 

>> Xiaofei Zhao: I see, sure. And then we also need to sort of bring the benefits of work from, at a worker level, not just to the firm value, right? And that will create a more trade off at the workflow yeah.

>> Audience: Can I ask one quick. Sure. So that was great one quick question is I don't know whether you or anyone else has done.

There's the other kind of classic asset pricing thing which is look at daily returns regressed on ICT human capital times, either COVID infection rates or lockdown. Cuz the prediction would be you want like a shock so suddenly you're forced to work from home. And if you have high human capital you're in good shape.

And if you don't, you're kind of screwed basically. And I guess maybe you've done that, seems. No, we haven't done that yet, but we can do that. I was just trying to think what's the most obvious shock to being either the COVID infection rates. You could even do surprise cuz the COVID infection rates in AR one protein roughly, or an AR two.

And you can just kind of have innovations to COVID basically on something. Suddenly there's news today that, wait a minute, we're all gonna have to shift to work from home, which either a lockdown announcement or COVID infection or something like that. Cuz I would assume that's what's going on early.

That's what you're at home. Some places are fine and others are just completely equipped for it and therefore they're in trouble.

>> Xiaofei Zhao: Yeah.

>> Audience: Cuz the problem is yes, I'll stop.

>> Xiaofei Zhao: Yeah, sure.

>> Audience: All of the data looked like you were asking questions. You're tracking on software engineering specifically.

You looked at the job categories it seems.

>> Xiaofei Zhao: The top ones, yeah but we have 81 jobs as I city job.

>> Audience: I'm curious if you have a split out. I mean, as a computer guy I totally get that. But the secondary thing is what about location independent knowledge workers that are not computer people.

So there's either some administration stuff, there's accounting, there's some human. Are those tracked in there anywhere or if not, could they be in future?

>> Xiaofei Zhao: They can be, as I mentioned earlier. I mean we have 81 jobs classified as ICT jobs. I haven't looked into the full list, but we can definitely capture for any occupation if you want to, we can capture their ICT human capital contribution as well yeah.

 

>> Audience: Okay.

>> Xiaofei Zhao: Yeah thank you.

>> Speaker 3: Thank you.

>> Xiaofei Zhao: Thank you.

>> Speaker 3: So now we're gonna have David, we're all talking about taxes and telework.

>> David: Okay, great thank you very much. This is a joint work with Jan Bruckner and it's part of a long research agenda to first understand the legalities of taxing telework.

Then the theoretical implications, then the empirics so a Jan and I's paper kind of is in the theory, but this is an intense legal area. And so I'm gonna spend Nick suggests in the first maybe ten minutes to get us all on the page about what actually are the rules that we're talking about here.

So that then we can get to the stage at the end to be able to discuss what's the exact data we need kind of empirically test this. And so I'm gonna draw on some of the legal work that I've done with Kirk Stark when introducing this topic. So an important part of work from home and the pandemic more generally is the impact on state and local public finance.

And we've seen dramatic shifts in property tax, sales tax, and potentially income tax revenue that differed from our initial forecasts of what was gonna happen. And those impacts have been extremely heterogeneous across jurisdiction size. With respect to income taxes. The key point of our paper is going to be that work from home is gonna decouple basically the residents and the employment state.

Which is now going to allow taxpayers to have potentially two margins of behavioral responses in terms of their mobility to those taxes. Conditional on their residence, they could change the location of their employer, or conditional on their employer, they could change the location of their residence. And now the key is, depending upon the tax rules that are in place, where is the income actually taxed?

Which one of those things, or potentially both of those things, do you need to move to basically avoid or lower your tax burdens? And so this is a big policy debate, and it's been kind of very popularized in the media, and it's kind of gone so far at the end to kind of go to the suggestion of, well, even if you're on vacation.

And you're physically working from a location, do they have the right to tax your income when you're working from vacation there? And so this is a very controversial topic in terms of state tax law. And the dynamics of this controversy came full force in terms of when the state of New Hampshire sued Massachusetts in the early days of the pandemic.

So what happened was there were lockdowns, and people who were previously commuting into the Boston metropolitan area from New Hampshire were now working from home and in their New Hampshire houses. And previously, Boston, Massachusetts, would have been able to tax that income, but now they were working from home and they weren't physically in Massachusetts anymore.

So Massachusetts, based on their laws, had no right to tax that income. So what Massachusetts did is they just changed their law. And they said, well, we're gonna tax teleworkers. And so even though you don't physically come in here, we're just gonna maintain what was the prior status quo.

New Hampshire said, no, you can't do that. These workers never coming into Boston or Massachusetts anymore. And we have a zero tax, and we've chosen that. So they petitioned to the Supreme Court. Supreme Court said, we don't wanna hear this. Let the states resolve this. And this means we have a hodgepodge of extremely different tax rules across states.

And that's kind of the variation that Jan and I are going to exploit in the model and that, ideally, everyone would want to exploit going forward empirically. So the basic question is as follows. Suppose that you have a software engineer living in Tulsa, Oklahoma, and they're providing services for an employer that's located in Silicon Valley, potentially without ever setting foot in there, or maybe setting foot in for a couple of days.

Where does the actual activity take place? That's where states disagree. So some states will say, well, the activity is based on where the employer is because we are facilitating that business environment. Other states would say, no, the activity is physically occurring where the individual is residing. And so now we have this decoupling depending upon which ways the states have elected to tax them.

I want to just walk you through state tax laws and all of the variety of the possible combinations, and I'm just going to run through a very simple example. Suppose that there's two states in the world. State H. State H is high tax and it has a 5% flat rate.

And state L is a low tax state, and it has a low 2% flat rate. Now, let's suppose that an individual earns $100,000, and let's go through the cases of where this income could potentially be taxed. At the polar extreme, there's two polar cases. Number one, we could have an entirely residence based principle.

So this would be where you live has the entire taxing, right on all of the income that you earn, regardless of where your employer is located. The opposite polar extreme is what I'll call the source principle, and that would be 100% based on where your employer says that they booked you on payroll.

These not need be the same, obviously. You could have all states being source, you could have all states being residents. But in practice, what we have is some states want to do source, some states want to do residence. And so I'll need to just walk you through what happens in those cases where then the issue of tax credits becomes important.

So let's just look at an example where an individual who lives in the high tax state at the 5% rate and they have an employer that is located in the low tax 2% state. Let's suppose that both of these states have already agreed either 100% source or 100% residence principle that they're both going to apply.

In the blue lines, I'm going to plot what the distribution of your tax burden the same person would be if the source principal prevailed. Then in red I'll plot what it would be if the residence principle prevailed. If the source principle prevails, then you pay $2,000. That's a 2% times the 100,000 to your employer's state where it's located.

You pay zero to your residence state for a total tax liability of $2,000. If instead we kept the same allocation of your residence and employer, but we only changed the tax rules, then you would pay to the residence principal. Then you pay $5,000, the 5% rate to your state of residence, zero to your employment state, for a total of 5000.

The key is that depending upon the sourcing rule, if you want to lower your tax liability, either you have to change your residence or your employment. If the residence principle applies, then the source tax rate is basically irrelevant and you need to change your residence. Whereas if the source principle applies, then the residence tax rate is irrelevant and you need to change the job or you need to convince your employer to book you on payroll in another potentially multi-state.

The mobility response could even result from the employer adjusting where they say is your physical work location. So where are these in place? Many of the states have kind of residence only rules. Many payroll tax, however, are source based. At the local level, many local payroll taxes within metropolitan areas are source based.

So I'm talking about states here. You could easily call these states localities. Prior to the pandemic, six states attacked remote workers based on where the employer's office was. I've listed those six states there. Massachusetts has been an addition, and there's currently legislation in other states pending to try and switch the principle here.

And so now the question is, well, what happens if you're in a combination of, say, Kentucky and New York, where Kentucky has a residence rule and New York does not have a residence rule, they have a source rule. And so how do things work there? Well, now we need to either have tax credits.

So the way it would work is, suppose that I am living in the high tax state and then I have an employer in the low tax state. Now there's two possible cases. The resident state can either give a tax credit or they cannot give a tax credit, potentially against the income.

Under the credit case, that's in blue here, the way it works is as follows. You first calculate your tax liability in the employer state, in New York at 2% there, and that's $2,000. I go to my residence state. How much do I owe them? Well, it's a 5% rate.

So I would owe 5000, but I've already paid 2000 to New York, so I'm going to get a tax credit for that 2000. I owe additional 3000 then to my residence state for a total of 5000. The alternative is that there's controversy here. Some states might say, well this, this person's never setting physically in, you have no right to tax this.

Vermont does this with New York. They say, you just have no right to tax this. We're not going to offer a tax credit. And so what happens? That individual pays $2,000 to their employer's state, then their state has a 5% rate and there's no credit. So then the total tax liability is the 7000.

And so you have basically all this wide range of complexities and variation that we now want to exploit in terms of thinking about the mobility responses. And these issues get even more complicated if we start having digital nomads like we heard on the first day, if we have digital firms and if we have multi state firms that can potentially shift their payroll.

And so you could have moves happening necessarily without even the choice of the individual, individual worker. What I want to talk about that Jan and I are going to do is we're going to then try and say, well, let's, for purposes of the talk, I'm going to say let's look at the purely source base, that polar case and the purely residence based case.

Let's see what the behavioral responses and the incidence effects of work from home are going to be. All of the other more complex double taxation and credit cases are in the paper. I'd encourage you to see that if you're interested. And so what we're first going to do is say, well, now we have this big bang.

Work from home comes about. How does that now affect the equilibrium, the spatial equilibrium in the presence of a taxation? How does population and employment change? Then once work from home is established, how does a marginal tax rate in one state change then the allocation of employment and population and wages and house prices?

So the model I'm going to show you today is quite a stylized version of the one that's in the paper. I have employment that's given by l, a population that's given by n, and there's a wage rate w of l, and house prices. And states also differ in their exogenous amenities, a, that are either high or low.

Labor is going to be taxed in this model at an ad valorem flat tax rate. And it's going to finance a publicly provided private good, which depends upon the tax base, which obviously depends on the tax rate, but also on the rules, whether it's the source or the residence principle.

And so I'm gonna show you kind of a quasi-linear utility framework here where we have an additive exogenous amenity. Plus think of this as being your private consumption net of taxes and other expenditures. And then v here is your valuation of the public good. And then this H function is the utility that's derived from your consumption of housing net.

And that depends negatively as a function of n. For purposes of the talk, I'm gonna set aside this public services. So I'm gonna zero out this beta. And we could think of this then as very high income taxpayers to kind of get no valuation from the public services.

The reason that I'm gonna do this is it's gonna just make everything really crisp for me talking to you. If these public services are adjusting as they are, right, that's kinda an offsetting force, right? Higher taxes potentially imply higher benefits. And then you potentially get different responses of the high income people and the low income people depending upon their valuation of public services.

And so all of that is kind of discussed in the paper, but I'll just focus on this very high income case. And so if I assume now that there's two states, state h and state l. So state h is the higher amenity state, which is exogenous. What happens before telework?

Well, if there's telework, you need to jointly determine your absent interstate commuting, you need to jointly determine your location of residence and your location of labor supply to where the firm is. And those two things need to move in lockstep. So we need to have that l equals n.

And so then there is only in terms of spatial equilibrium, one condition that we need to have one equal utility condition which we can then show applying that the higher amenity state in terms of its exogenous amenities will have a higher population, therefore a higher employment and therefore higher house prices and lower wages, given the wage function, has the standard negative derivative there.

And this is basically looks like the standard rows and row back result. And so now the question is, how does tax affect this in the presence of work from home? So now let's open up work from home. And to do this now labor and population no longer need to be equal because they're decoupled.

And so now I need to have two indifference conditions for spatial equilibrium. Number one, you need to be indifferent between the location of where you're working, that's gonna then determine L. And then you need to be equal utility indifference across the places where you're going to be living.

The key is that the form of these equal utility and wage conditions depends upon whether we have the source principle or the residence principle. If the source principle prevails, we need to equalize the net of tax. Why? Because you're gonna be potentially living in, conditional on living in state h, you're gonna be paying different taxes depending upon your employer is.

And so I need that to be equalized. Whereas if your residence principle is applying, if you live in state h, then regardless of if you work in state l or state h, you pay the same tax rate, 1- th. And so then that 1-th is gonna disappear. The nice thing then is that then that also feeds back into your equal utility condition.

If we have this equalization here, then going back to this original slide, this term is going to then basically be equal on both sides and it's going to drop out. Whereas with residence taxation, it doesn't drop out because I have equalization of wages, not after tax wages, which is what's necessary for your consumption.

Okay, so now what is the effect of work from home? Opening up work from home immediately on the allocation of labor and population after work from home happens, holding constant the tax rates at their optimal levels prior to work from home. In other words, in the regime when L equaled N.

What we can do is we can go back to these conditions here and we can evaluate them at the old equilibrium values that we had at when L was equal to N. Keep in mind, this first result said that state h in the pre work from home era had higher population, had higher employment.

So now let's evaluate in that higher h, higher l into this equation here. And what we can see is the direction of the inequality is such that with the shift to work from home, we're gonna have employment that's gonna fall and the wage rate is gonna rise in state h relative to state l.

And because there's this 1-th term under the source principle, but it's not there under the residence principle, we can sign, not sign, we can determine the relative magnitudes with the changes being bigger than under the source principle than under the residence principle, okay? We know that now there's gonna be a shift basically of employment away from state h towards state l.

What about populations? Well, we can now evaluate this equal utility conditions at kind of the old equilibrium value. And what we can show is that, with the shift to work from home, the population housing price are gonna then rise in state h. So what's the intuition for this?

The intuition is, well, I just told you kind of what's going to be happening in terms of the wages. So state h is now becoming relatively more attractive and therefore, its population is going to be increasing. And so if there's public goods under the source principle, this is going to really complicate the analysis because now you have potentially tax exporting that's going to be occurring.

And depending upon that relative magnitude, then that might then complicate whether the wage effect or the valuation in terms of public services matter. I wanna show you two last quick results. The first is an efficiency analysis. You might say, well, which tax regime is optimal? So we can show you a really important result.

Number one, the pre work from home is not efficient. It doesn't satisfy the three efficiency conditions necessary equalization of marginal products, optimal population condition, and kind of a benefit tax for the goods. It satisfies some of them, but not all. But interesting enough, what work from home does is it allows us to actually achieve the social optimal, but only if residence taxation is in place.

Why? Because then we get equalization of marginal products and population clearing and the public good with residence taxation turns into an efficient benefit tax. And so work from home can be welfare improving. Why? It's decoupled the l and the n, and I now can have optimized both of these things separately to satisfy all three of those optimality conditions.

The last thing you might wonder is states change their tax code all the time. Now that we have work from home. Suppose I have a marginal increase in th. How's that gonna affect the comparative status? Well, under the old regime, population and employment in state age need to move opposite of their tax rates.

Higher tax rates repel high- income workers, and these two things move in lockstep. But now, depending on if you have the source principle, the migration response is either entirely in terms of residence population when you have the residence tax regime, or just jobs when you have the employment tax regime.

And so what do we notice? All the literature that focused on the pre-pandemic era on residents' mobility responses, estimating the residents elasticity is no longer sufficient to measure the spatial distortion from taxes. If you have some states that operate the source principle, we now need to study where the jobs or the employers are potentially moving.

And it's an empirical question as to then whether the population or the employment elasticity are going to be larger or smaller than each other. If I add public goods to this, then we get even a wider range, potentially even opposite signs for different groups. So with that, let me just conclude and say that the incidence and the mobility responses are going to be different before and after work from home.

They're going to differ depending upon the tax rules. And as I've said, we have a variety of tax rules in place that we can now empirically exploit to estimate both the mobility and the employment elasticities. But to do this, we need kind of really precise data sources to capture the location of work, which is not necessarily equal to the place of residence.

And ideally, you would want to know the split. If an individual is working a certain number of days in one state and in another state, you might want to know that, because the states actually might apportion taxes depending upon the duty days that you spent. And we kind of need a database that kind of allows us to partition based on income.

Why? Because we know that these tax effects are most prominent when the tax differentials become really important. So with that, I'll kind of conclude and take any comments or questions or suggestions for great avenues for empirical work.

>> Audience: Thanks. Super interesting. And I think you've hit on a really important policy issue, so keep working on it.

A couple comments. First, I think the source-based taxation will ultimately devolve into effective residence-based taxation because the arbitrage incentives are so strong. And so the way that happens, and you kind of hinted at this in your remark, if Massachusetts tries to impose source-based taxation on people who live entirely in New Hampshire, then at least the larger companies are just going to open up a unit in New Hampshire and say, okay, well, now we're going to do source-based taxation.

These people report to an office in New Hampshire. And the financial incentives to do that are enormous. So I'm sure Massachusetts and other states that had large numbers of inward commuters before the pandemic are going to try to resist that in a lawyerly manner with all kinds of rules, but ultimately that's a losing game, I think.

So that's kind of point 1. Point 2 is, do we see any evidence of this already? So there's the Census Bureau has this new business applications data which Halteringer, in particular has helped develop and push, though, you know what I'm talking about, the new applications for EIN, and those are available at the state and industry level.

Maybe they're publicly available at the state by industry level. So question, do you see in industries with, say, lots of potential teleworkers, all of a sudden, there's a bunch of new business applications in New Hampshire that, relative to Massachusetts, that's a way to just evaluate these. Third thing, the whole picture is actually more complicated than you suggested, because once you go, because there's big questions about, okay, who's responsible to which state's unemployment insurance system, and which state's workers compensation rules is the worker attached to?

And that becomes especially an issue when you move towards residence based taxation. So great work. Keep it up.

>> David: Yeah, thanks. Thanks. So I'll just respond a little bit quickly. So on the first one, you know that the source is kind of potentially a losing battle. I agree with you in some spirit, there could potentially be some offsetting forces, in particular, some of these regulatory differences across states.

So if you establish in another state now, you might trigger nexus there as the firm, which might trigger corporate tax liabilities. And so there could be complications. Now, that being said, there's probably a place that's a haven in terms of everything out there. And so that's one aspect.

In terms of the history of the income tax, originally we were like for physical workers like residence base, and New York moved first and they said, hey, we want to tax all these in border commuters. And now that kind of got us to where we are. So they were, so the question is, like these big source employment states, do they have enough power to do that?

That might be one thing in terms of do we see anything in the data yet? So we've explored a little bit initially to try and look at this both in terms of the business side and also on the population decide. I should say there's one major complication in the first, about year and a half after the pandemic, many of the states, what they did is they said that, hey, we just have no idea how to deal with this.

And they kind of put restrictions on their tax laws. So there was a period of kind of uncertainty over what actually the tax law was. And so when we first looked at this in the first year, which was the data that we had, it's kind of hard to notice something.

And I think part of it is just that uncertainty. But now we've kind of settled to a new equilibrium, and I'm going through and I'm cataloging and I can actually, you know, figure out every state's tax law. But I couldn't do that, you know, in the first part of the pandemic.

And, yeah, I certainly agree with you that, you know, there's a lot of complications that I've swept under the rug here.

>> Audience: Two thoughts. So super interesting. And so one question and one comment. So I was wondering whether there's also beyond tax, a question around benefits and, you know, whether there's arbitrage opportunities for benefits across states.

Just a simple question about what is the cost for a firm to open a payroll across 50 states. Is it a small cost? If you can just educate us on what is the cost of just doing it. Widespread 50 states?

>> David: Yeah, I think actually that the cost is quite large, especially for small or medium-sized businesses.

And one place that we're seeing this is even just thinking about having a work from home employee, even not opening another place, that triggers understanding of the tax laws, potentially of where that worker is located, even if you haven't established there. And many companies view those costs as really quite large.

And in particular, we've seen some places now that get around this. What they're doing is instead of issuing a W2, they might be issuing a 1099. I've heard anecdotal evidence of that. And then they're altering the wage and they're saying, employee, you go figure out all of these state law issues.

We're going to give you this 1099 with a higher wage to have you deal with that. And so I don't really know the kind of the cost of then opening the business, but it seems that which suggests that they kind of view that as costly. Right. Because they haven't impartially done that yet.

And so But certainly larger companies have some right to do this. And then, yeah, I agree that there's certainly also kind of benefits here. So potentially the cross-state issues triggers what are the unemployment insurance aspects that you're eligible for? Obviously the public services that are matter there. And so the decision for many people is not going to be just tax.

And at the end of the day, really it's a net benefits, which is like a dollar change in tax is net of the utility change in marginal benefits. Which is what we talk about more generally in the paper and that I didn't talk about today.

>> Audience: Just one footnote of that discussion, it is a high cost per for a single worker.

But there's other ways around the arbitrage, which is you lease employees from a company that has established a nexus in another state. And then, you're not having to deal with all this. Sometimes some jobs, at least employees work well and some jobs they don't. But there's many avenues for these arbitrage possibilities which will evolve over time, I think.

 

>> David: And I think this is really a great way, then why just kind of segue into why I think this is a really interesting empirically area to work out. Cuz it's not just about the mobility of the workers, but also then thinking about the tax incentives that these firms have.

How is this changing their business models, right, and how are these state tax deferentials doing that? And so like the data sets, ideally we would want to do this. Might be ones that shed light on those questions as well.

>> Audience: A two part thing. One, to Raj's comment, yes, I've seen organizations that are fully distributed saying, okay, you can move from state number one to state number two because it's lower personal taxation.

However, you cannot move to state number three, because we do not have a legal entity there and we do not want to deal with figuring out the tax filings for that state just because one human moved there. So you can move to these states, but not these states.

And if you do insist on moving there for personal reasons, you have to switch to being a 1099, just like so. I've seen multiple cases of that. Personally, I can vouch for that. I've also seen corporations saying, well, we'll take our entire, we have a big fixed asset like a airplane manufacturing or car manufacturing factory.

We're gonna move the nominal headquarters and a bunch of the, quote, information workers to a different jurisdiction. And we're gonna leave the fixed asset, whether it's a ship building, airplane building or car, that stays there and all of the higher taxable things go elsewhere. This kind of game is going to be going on.

This is, by the way, fantastic. I'm just going to add that.

>> David: What we know from the corporate tax literature, especially with respect to digital things, is that this is huge arbitrage. And so what telework does is it brings a lot of those things that we previously knew were happening for the corporate tax to now the personal income tax side.

 

>> Audience: And I'd also just add and say this is also happening internationally too. This is all domestic, but-

>> David: When I use the word state here, you could equally substitute locality, but also country. And then the tax treaties, the bilateral tax treaties between them, they're either source or residence or some combination of the two.

 

>> Speaker 3: Great, I think we should-

>> Audience: Let me just ask, make a co author point here, which is that some of you may be wondering, well, wait a minute, how does this apply to the work from home that we're mostly talking about at this conference, which is hybrid work from home?

The answer is it doesn't. This paper is about fully remote work where people are living, working in different places. Now in the paper we have a little bit of a fake hybrid model, but it doesn't quite exactly translate into what we see in reality. In order for the model to really work in a hybrid world, is the jurisdictions, the suburban and central city jurisdictions have to have their own income taxes.

Then it sort of looks like the model we're talking about. So to the extent that fully remote work is really a minor part of the whole picture, our model applies to just a minor part of the picture. But we think that it's conceptually really important to get this sorted out, and that's what the paper does.

>> Speaker 3: I think we should take our break now, but continue to discuss and we'll be back at 10:30.

Part 8:

>> Moderator: We're gonna be hearing from Michael Koelle about work from home in 20 different countries.

>> Michael Koelle: All right, thank you very much and let me start this last session of the conference by thanking the organizer for putting together this wonderful event that we've had the privilege to attend for the last three days has really been nice.

And of course, thank you for including this paper. So this academic paper that I'm gonna present is actually the outcome of a policy collaboration that we had between what was back then, our team at the OECD. And indeed way back in 2021 during the pandemic, we were interested in monitoring labor market trends and remote work across countries.

And so my co authors are my colleagues Gabriele and Cyril from the OECDE, I'm Michael and three co authors from the indeed Hiring Lab, which is the research arm of the job site indeed. Alexandre, Pavel and Tara has since moved to the United States government. So before I say anything else, I should note that anything I'm going to say is only position of the authors.

And should not be attributed to, indeed, the OECD, the OECD member countries, the us government, or the US Department of the Treasury. A long disclaimer in this case. Okay, so the basic question of this paper is, how is the pandemic catalyzed, the surgeon working from home? Of course, especially to this audience, it's clear that something has happened.

We're not back at stage one where we were three years ago, but then we also know that from some data sources, we get the impression working from home is peaked. Maybe it's going back, some employers are recalling their workers to the offices and moreover, a lot what we know about working from home, we know from the United States.

I think the discussions that the conference again attest to this. However, the pandemic was global, so it affected labor markets in many countries and this is what we're going to shed light on for 20 countries. And we use data from the job site indeed, so indeed is the largest job site in many markets.

We're covering these 20 countries, and I extracted data from about a billion and a half job month postings on indeed. The data is not only close country and high frequency at a monthly level, but it's also classified into economic sectors, basically occupations, mostly. That allows them to do difference type analysis, where we exploit these three margins of variation.

So across countries, across time, and across occupations with high and low working from home potential. To then estimate the cause effect of the pandemic on advertised working from home, distinguishing between periods of pandemic easing and pandemic tightening. And that will matter as you see. So this graph, this descriptive chart here basically tells the main story and shows the main findings.

So what you see here in blue is our time series of advertised working from home. So that's in the average country. The share of job ads on, indeed that mentioned the possibility of working from home. And as you see this increased from 2.5% to 3% just before the pandemic to 11.5% in March.

That's the latest data point on this chart. I now, 2 hours ago, my co author updated our publicly available data series up to August and so the data point for August is 11.2%. All right, so basically in the last half year, this has stayed ever up at close to 11.5%.

And then secondly, advertised working from home went up, you can also see here, we haven't read here one measure of pandemic stringency and you see the difference of like lockdowns. And every time it's a lockdown, basically the work from home share inches up. This is across countries of course, we're gonna do that in immaculate analysis at a much more fine grid level.

But basically the story is the same whenever there's a new lockdown, working from home goes up. When there's easing of pandemic restrictions, this doesn't go down again. Ok, so a word about the related literature. So as I said, our measure is advertised working from home. So basically we interpret that as a commitment by the firm to make working from home available to workers.

So also you should note that in many jurisdictions, what you mentioned, a job ad, has legal implications. All right, so it's important enough that the firm feels it is worth mentioning to a potential worker in a job ad. So contrast to survey measures, this doesn't capture the sort of like ad hoc adoption of working from home in the early stages of the pandemic, when everybody was sent home to work there.

Now, a few papers and we have to expand this discussion here has used online job postings in single countries. One paper by Nick and Steve and co authors also looks at across countries, across five anglophone countries. And I think these two papers are closely related, not only because of the cross country perspective, but also because they take two different, but in the end, complementary approaches to measuring online job postings.

Measuring working from home from online job postings as you will see, this is not as straightforward as one might think. There are some complications attached to that, so these need to be dealt with. And in two different ways, I think we arrived at basically the same conclusions. Okay, so let me first talk about the data and about measurement and I spent some time on there because I thought this might be of interest.

So as I said, we have 20 countries, so indeed has job portals, I'll show you a screenshot in a minute. Country by country, and these are 20 countries, a bunch of countries in Europe, all countries in North America, Israel, Japan, Australia, New Zealand. So if you think about the pandemic, these are countries that the pandemic unfolded in very, very different ways, right?

They locked down harsher or earlier, or some could open up earlier, had to lock down again, etc. So you have all these different varieties of pandemic experiences and also we cover eleven languages, right? So you need to find a way to deal with multiple languages. Basically we identify working from home using keywords in the job title, description or location, all of these languages, right?

And in the latest version of the paper, we also include hybrid keywords. So that's kind of like the examples on the second row here that basically a specific language to identify sort of like hybrid work arrangements. We test it quite extensively, sort of like to figure out what words exactly to use in all the different languages.

And the main variable then we construct is the share of doc postings that advertise working from home at the country occupation month level. This is the dataset we do analysis on. This is also the data set that's publicly available now until August on our website. Okay, so indeed, their business model is basically for job seekers to go on their website and find jobs and apply through that and for firms to want to advertise because the job seekers are there.

Okay, so this results in a very high quality of every single job posting because the business model is not sort of like analyzing aggregate trends, but, you know, having a good quality job posting for every single job. So when you go indeed and there's a country specific website, this is an example from the UK, you type in a job title, you type in location, you click on find.

Then you basically gonna get a list of different jobs and also to help job search for the job seekers, the pointer here. Yeah, you have sort of like these fields here that allow the job seeker to narrow down. This is metadata that's extracted from the posting or indeed when an employer posts a job that the employer basically provides.

Okay, and so how does a job posting look like? This is one job posting during the pandemic. So you have job title, you have the location, you have some other fields and basically have a long sort of like text of job description. And here I want to draw your attention to the part at the bottom of the job description.

And this is data that basically does employee provided metadata that indeed then adds sort of like to the text. So this is, this is data that sort of like helps you to narrow down your job search, but it's also added and for completed information to the text. That includes the job type, the contract type benefits here, in this case, a specific requirement about the driving license.

And then crucially, in the last Two rows, it says work remotely, column no. And this is a field that indeed added during a pandemic, basically to help job seekers navigate the labor market. But that, from the perspective of a researcher, poses some issues that we have to deal with.

So if we just do the keyword search as I described until so far, we get the following picture, and this is, again, across countries. So what you see here, basically, when indeed ads is employer provided metadata, both in the positive and in a negative. If you only search for keywords such as work remotely or remote or whatever it is, you're gonna get basically a huge jump that's gonna pick up all of these false negatives, right?

At some point this goes down, that corresponds to a time period when indeed changed the wording of this location question. So it's not work remotely, no, anymore, it's not differently. This was rolled out across countries in a staggered way, so you have this gradual decline, okay? But what you have basically, before anything else is applied, is this picture, to me, it reminds me of the little prince, the book.

There's a drawing of the snake that ate the elephant, this to me is like the elephant in the snake broth. This is the typical thing, and if you do that, say in burning glass, sorry, light cast data, you basically get the same thing, okay? So how do we correct for this?

Well, we use the information that I just described you, how this metadata comes in the job postings in the first place, right? So we know that indeed added work remotely called a no in the United Kingdom and different things in other countries. You basically use that information to undo, in a sense, these false positives, okay?

Then we get basically the blue line, which is, in a sense, the line I showed you earlier vintage. But the line that I showed you in the first class, okay, I just want to be clear, is that this is how we've always done. But in the first versions of our papers, we didn't explain all the full details.

Mostly thinking that this would be something that's behind the scenes happening in the indeed proprietary database. But, of course, we learn things from people in this room. This is something actually an issue that probably through, indeed is inherited by lots of other online job postings, databases, right? Everybody who works with other and job postings in a way, has to confront this.

Okay, so this is one thing about measurement. Another issue that is also raised in the literature is what I call false friends. The fact that basically, especially in the English language, you use very casual words, working from home, remote work, working remotely to describe these arrangements. Of course, you can come up with examples of strings of text where you have these words in a different meaning, right?

I would argue that this is less of a concern in many other languages. Partly because many languages use dedicated words to describe sort of like working from home arrangements. So you have the first two examples of German and Italians that use what linguists called pseudo-Anglicism's. These are English words that are borrowed by that language in a meaning that an English speaker wouldn't even recognize, right?

So in Germany, does home office, everybody knows that this is working from home. In Italy, this is smart working. And it's basically these words in these languages that describe these kinds of arrangements, and there's no sort of room for misinterpretation. Furthermore, many other languages basically use this artificial word telework and basically have tele plus their own word for work together.

And this describes these arrangements for some reasons in the English language, although we now know that this has existed for 50 years, but it hasn't really stuck to people's usage. Okay, so I want to be clear also that our methodology just not sort of like allowed to correct for this false fence problem.

But again, we view this as less of an issue than in many of the other languages. Basically, we do a robustness checks where we're gonna include English-speaking countries and the results fully go through. Okay, so after this part about the measurement, let me then go to the geometric strategy that we have.

So we're gonna marry a difference type of approach with something that's very common in the applied macro literature, which is called the local projection method. So this is a way of identifying the impulse response of a shock in a setting where you basically have many shocks happening at many different times.

So you don't have a nice, so like events before and after, but many shocks. You still want to see what the sort of like the unit effect of a shock. And so on the left hand side, what you're gonna have is the k-period change in the share of job postings, advertising, working from home, in country i, occupation, j, and time, t, and k is months.

And basically this window goes up over time, is separate regressions for each. But that basically allows you then to plot the coefficients as an impulse response function, yeah. On the right hand side, we're gonna have the variables that denote increases and decreases in pandemic severity at the country-level, country month-level.

Interacted with a predetermined occupation-level, working from home potential that we take from literature, from Jingle and Niemann and crosswalk it to our occupation categories. So this is basically the three types of differences that we exploit. And then we're gonna include two types of fixed effects. First of all, taking advantage of the fact that we're at a sub-country at the occupation-level, we're gonna put in country-time fixed effects.

So these vary at the same level as the pandemic measure, right? So they pick up anything that's unobserved at the country-level about the pandemic that might be correlated with the pandemic. Think about seasonality, think about there are many other things that happen at the country-level, okay? And basically we're gonna then compare occupations with low and high working from home potential in the same country.

And in a sense, occupations with low working from home potential serve as a counterfactual to what would have happened in occupations with high working from home potential in absence of the pandemic, okay? This counterfactual, we can condition it in one dimension because we are left hand side remembers the difference.

We put in occupation country fixed effects, this control for trends. So we're gonna control for linear occupation-country trend. So we allow that basically before the pandemic, say, IT jobs in the US were on a different trend to it jobs in France or to cleaning jobs in Belgium as an example, right?

That's the identification assumption is a different diff, so we can provide evidence of pre-trends. Actually, the evidence points to a stricter version of the assumption probably being true parallel trends. This is a standard parallel trends graph. You basically have the outcome on level-fixed effects and then month dummies.

You see basically before the pandemic, there was no clear pre-trend. As soon as the pandemic hits, basically the high worker from home potential occupations hit off. Okay, so then the results, and as I said, we present them in these impulse response functions. So first on the left hand side, what you see here is the differential effect of a change.

The one standard deviation change of pandemic severity measured by Google Mobility on advertised working from home. Between occupations with a high and a low potential, there's a relative effect, not an absolute effect. What we see basically a month after impact, you have a significant effect that goes up and stabilizes around 4 or 5 months after.

And the effect size basically for unit standard deviation changes is about 0.8 percentage points. To put that into perspective, the first lockdown between February and April 2020 on average across country corresponds to six standard deviations increases in severity, okay? That's just sort of to give you a ballpark of what that estimate means in size.

Then, by contrast, if you look at decreases in pandemic severity basically, we find nothing, okay? So the pandemic catalyzed the increase in working from home, but didn't undo it when the pandemic went back. This is a result, and basically we do a bunch of robustness, variants of the same result that don't change anything meaningfully.

And we have alternative measures for different variables, we can use them all. Changes in a lab structure, that's what people do with this kind of methodology. This,

>> Michael Koelle: Then we exclude health occupations in the baseline. We can include them. We can control for health outcomes. We exclude English speaking countries.

What I said before, we have different ways of clustering. We run a placebo test on 2019. So we present a pandemic happened in 2019 in the same months and basically see whether we find anything that we don't. So I, for example, was worried about seasonality possibly playing a role.

Apparently it doesn't seem to be a major driver. You do separate analysis before and after the vaccination campaign. Maybe expectations of people might have been different. Doesn't really any difference that shows up here. Use different vintages of our data. Basically the results all go through. Okay, so to conclude, we documented this paper a fourfold increase in the share of honor job postings across 20 countries that advertise working from home.

And that has not been reversed to date. And to date, basically meaning end of last month. This suggests that it triggered some kind of path dependency through different mechanisms. And we're at a conclusion here. So we're not gonna provide evidence about this mechanism. I think the data really doesn't allow us to say a lot about this.

So we can only speculate about them. I think especially the last paper in the program, we talk about that. So the implication, of course, is that working from home is here to stay. There are many implications for firms, workers, cities in a macroeconomy, as we have seen. And we continue adopting to that.

With that, thanks very much and looking forward to questions. That was great.

>> Speaker 3: Obviously, no one liked the paper. One question I'm going to ask somebody, I can't remember who asked it on the first day about, there's a bunch of, we have the same issue with Hansen. Everyone has the same issue, which is there are some things that obviously allow work from home, but don't say it in the job Act.

Like assistant professors, whereby, but I don't know how to deal with it, actually. And what I was interested in is particularly the early period, you had the work from home allowed, yes. No, it's quite useful actually that covers all jobs. And I guess was the share. Does every job have that field?

Because that's forces you to answer it. The problem is when it's nothing spoken, because with assistant professors, if you look at job openings for economists, clearly it would be yes if it had it actually. I don't know what would have said. It's interesting, if Stanford put out a job posting, what would we have said?

We probably said no cuz we mean that you have to come in at least to teach, but you're. Yeah, I don't know. That was one question. The other question was on what you saw in the international variation, but that was more just a quick data question. But on the first, I don't know if you had thoughts on when you had that field, how to deal with it.

We have the same issue. It's actually really hard to think about how it crashes up.

>> Michael Koelle: Yeah, no thanks is a very good point. So this field, we have that for all the job postings that are posted actively by employers. Which is not all the job postings, indeed, because they're sourced from a variety of sources that share various.

Across the pandemic, I think at the height point is about 50%. It's a bit lower in other instances. And this is actually. This is the source also of the false positives only what's posted by employers. So that's the first answer to the question. So we have that for a subset, it's not sort of bio occupational or by anything like that.

But for a subset of dog postings, we have 55 occupational categories. They don't correspond to any standards of occupational classification and they're detailed in the paper, but they vary a bit. So on some occupations, they're very detailed. IT occupations, engineers, for example, and some others have a course.

So I think we couldn't identify assistant professors or maybe even academics.

>> Speaker 3: It would be great to just redo it just for ones where they have the field, actually, just to see what it looks like for that subset, because for that subset, they're forced to either say yes, no.

The problem is when they don't say the way, it's slightly opaque, actually. So you could redo it for the subset, but they're forced to say yes or no. It would be good to see what share that was.

>> Michael Koelle: Yeah, thanks. As far as I remember, basically both of these subsets have similar trends, so there doesn't seem to be a big difference across somebody.

It's a very good suggestion. Thank you.

>> Ben: So, yeah, hi. I heard that indeed provide access to the applicant data. Maybe that's just a rumor, but someone told me that. So I wondered if you ever looked at all or done any work and if there's any trainings, if you have, and who's applying to these jobs and locations in particular?

 

>> Michael Koelle: Yes, no, short answer? No, we're not looking at applicant data at all in this project.

>> John Taylor: So your measure of pandemic severity is just mobility, is that right?

>> Michael Koelle: So what I presented here is mobility. The alternative measure is coded government restrictions, but they're different.

>> John Taylor: Mobility restrictions, laws that are coded up.

Have you made an effort to parse out the partial effect of government mandates, restrictions on mobility and activity, apart from what you've got is just the Google mobility measure. Which is a reflection of what people do voluntarily and businesses do voluntarily, plus what they're mandated to do by the government.

But simple approach would simply be to add into that current specification the restrictions on commercial and social activity. Which I think you have for all or almost all of these countries, and then ask whether there's a separate marginal effect of that and whether it looks different.

>> Michael Koelle: So, yeah, so what I just put up, this is the equivalent graph using basically the government restrictions index.

So this is now not Google mobility. This is the other measure. So basically, politely looks the same. I think we think of both of these measures as being highly correlated, in a sense, Google mobility.

>> John Taylor: They are correlated, of course, but I want to know whether there's a marginal effect you can anticipate.

Why might there be a separate marginal effect? Because if the government is imposing restrictions on activity above and beyond what people are voluntarily doing. That might have implications for future government behavior that would lead firms to make investments or decisions to adopt this new technology. So it's a simple expansion of your existing specification, and I wonder whether there's enough power in the data to sort these out.

 

>> Michael Koelle: Yeah, I want to say the wrong thing. I think we tried, and I think there's not a lot of power, but definitely something we can look at again.

>> John Taylor: It's worth saying that, too.

>> Michael Koelle: What I can tell you is we can control for health outcomes, so we worry.

Basically, it's sort of the health situation as such. Sort of also the role so that we can partial out a control for in addition to either good mobility or this measure. And basically, that doesn't change anything.

>> John Taylor: Yeah, well, that's consistent with what I've seen in other settings.

The direct health measures don't take you very far.

>> Speaker 6: One on the metadata and then follow up on. So for federal employment in the United States, there's a, you have to go through. Not to, indeed, but you go through another site called USA Jobs. And USA jobs as of summer of 2022 now has a remote as a setting.

So when you're posting a job, you must say, is it location specific or telework negotiable or remote? Okay, and so we can search on that for exactly the stuff you're talking about, just to something Nick was saying earlier. That's an option, and then to Ben's comment earlier, I think there's an interesting thing here.

I think I heard you say you don't do this, but I think there's an interesting thing to be able to use the same information and talk to whoever was. I don't know if she's here today. The person who did the digital nomad piece to track. And see does a person with a profile on indeed change from position in Location 1 to a position in Location 2, going through different postings in indeed, there could be some interesting trending things there.

 

>> Michael Koelle: Yeah, thank you.

>> Speaker 7: So building on that. So I thought Ben's suggestion was very helpful because in terms of mechanisms, I wonder if you can test some demand side mechanisms. So conditional on the same wage or other observable characteristics of the job, do you see more applications for these remote job postings?

And that could just shed some light on why the firms are posting more, right?

>> Michael Koelle: Yeah, thank you.

>> Moderator: Guess we can moved on to the next one if there aren't additional questions.

>> Michael Koelle: All right, thank you very much.

>> Moderator: Thank you. Okay, we're gonna have Victoria Vernon talking about remote work, wages, and hours.

 

>> Victoria Vernon: Hi, this is joint work with Sabrina Pabilonia from the US Bureau of Labor Statistics. And none of this represents the view of Bureau of Labor Statistics. All mistakes are ours. We'll talk about remote work, wages and hours worked in the United States. This work is based on American community survey.

Our research questions are, do remote workers earn higher or lower wages than onsite workers? Do they work longer or shorter hours? How do remote workers differentials? How did it change over time throughout the pandemic? How did hours of work of remote workers change compared to on site workers?

Do trends vary by gender or by some other characteristics? And whether the wages of remote workers grown slower or faster than the wages of onsite workers and how that relates to the inflation? So, based on the American Community Survey, on a sample of full time full year employees, they share, those who work primarily from home has increased steadily from 2010 over to 2019 and then shot up.

So our shares are based on the commute question answer. So how did the person usually get to work last week? So if the person worked primarily from home, even being a remote worker, not necessarily a full time remote worker, but hybrid worker, a couple of times a week at home, no, a couple of times a week at work onsite, that would still be considered a home-based worker.

So work from home. So we see the share of women is larger than the share of men who worked at home, but on average, it's about 20%. We compare it to the American time-use survey, and I've seen that the trends are largely the same, even though the time-use survey doesn't measure the share of workers, but measures the share of work days by full time full year workers.

So it's based on the work days slightly larger share, but the trends are the same so back to the american community survey. If you split respondents by occupation, then you'll see that the most teleworkable occupations are no surprise, computer and mathematical. So from 2019 to 2021, the share of remote workers has increased to more than half of workers, the report being remote.

So in business and financial operations, arts and design, legal, the increase is tremendous. And more than 30% of workers consider themselves are home based. We call those white collar occupations, those who earn more than 10% of workers in 2021 are remote and the rest are either health care or blue collar occupation.

Theoretically, how can the wages of remote worker differ from those of onsite workers? Well, there could be a penalty because of the compensating wage differential, because being working from home is an amenity and therefore people are willing to pay for it and give up some wages. We can also all hypothesized that less productive workers may be selected into work from home.

Some of us have done more research on that. Then workers are less productive because there might be children running around. However, there might be an opposite story at play and there might be a premium for work from home people. Let's say, because work from home increased workers productivity due to reduced commute, maybe there's a better work environment now and people are better able to do their job.

The cost for employers also decreased and now the employers may be willing to share some of the through cost reduction with workers, their cost for working from home has increased for employees. Therefore they should be compensated for that a little bit through higher wages. There might be finally social isolation impact of work from home and therefore it's a disamenity, not an amenity.

So what has been done before? We know that Oettinger and White both wrote paper on similar topics, and Oettinger looked at workers from 1980 and over up to 2014. Prior to the 21st century, there was a penalty for working from home which changed to a small premium by 2014.

The Barrero have shown that wage growth is lower in high work from home jobs, which may lead to lower inflation. Our sample looks at workers aged 25 to 64, full time, non-farm sector, and we dropped some observations with wages below $3. Nominal wage trends and real wage trends show that on average, real wages for remote workers have increased over time, whereas for non-remote workers, slightly kind of always the same, but went slightly down over the last couple of years.

Wage premiums unadjusted. This is mean wages. Main differences in wages between remote and onsite workers have increased in all, but one occupation, only healthcare supports. Still had a wage penalty in 2021. But overall, 22 out of 23 occupations have a wage premium on average, and especially large in sales.

And it actually used to be, even before the pandemic sales had a very large wage premium. I'm assuming these are real estate agents and each other. Actually, I was gonna point out legal occupations used to have a penalty and now have a pretty large premium or remote work, usual weekly hours of work.

So these are reported usual weekly hours of work. They converged over time. So it used to be that remote workers were two hours longer per week than on site workers, and it's no longer the case they work the same number of hours. We estimate the usual wage regressions.

So, for each year, we estimate the effect of being a remote worker with the usual set of controls for age, quadratic and age, number of children, number of older children, teenage children. Number of adult family members, indicators for education, race, Hispanic heritage, marriage, cohabiting status, disability, living with a partner with a disability or having a parent with a disability.

Being a government employer, and then controlling for 21 occupations, 19 industry and state fixed effects. So a full set of controls, and all we're interested in that first row of results, where we'll present it in the form of the thick lines on the graph year by year, and then the thin line on the graph, the blurry line, will be our attempt to estimate Oster beta.

So what that is, is a measure that allows us to attempt to control for the unobservable in the regressions, a lot of unobservables in the regression. So the regular wage regression does not allow us to control for a lot of things, like the size of the firm, the size of the city, workers motivation, workers tech skills, a lot of things, even simple workers tenure is not available in the data.

So if we were to assume that the selection on unobservables is the same size as the selection of unobservables, we can estimate the Oster beta and consider it kind of a lower bound on our coefficient. So the estimates are robust to omitted variable bias if the bound is not including zero, essentially.

So for men and women, what we could see is that the simple wage regression show that the premium for remote workers have increased both for men and women, and have become, in fact, very large over the last couple of years during the pandemic. And there's probably evidence that there's more selection among men as you see, the Oster beta is negative, but less selection among women.

Let's move on. So we do the same analysis by groups, so separately for various occupations. And what we see is that the remote work wage premium is highest in sales, about 20% then in management, production, arts, design, and it's almost non-existent in, let's say, building grounds maintenance. And actually there is a penalty in healthcare support, still by occupation.

But overall they wage premiums about 14%, so on average. And these are the difference by occupation. So by white collar and blue collar occupation, you could see the difference, 14% in 2021 among white collar occupation, but very small, about 4% among blue collar and healthcare occupations. We use the composition similar to Oaxaca and the kind that Ottinger also uses in his paper to figure out whether the total change in a remote employment share is due to within or between occupation changes.

And it turns out it's due to changes in remote employment shares within occupations. And then the second decomposition asks whether the total change in mean log wage gap between remote and onsite workers is due to various factors. And we find that it's mostly due to the mean observed skill gap between remote workers and on site workers, as well as due to remote wage premium within occupations.

So we don't find that there is any occupational shift overall in the economy, but within occupations. Biparental status also where we might expect that mothers could be paying wage penalty, they are not. We don't see any evidence of that. Mothers and fathers are all groups which have experienced an increase in wage premier, these are white collar workers too.

Over time, then among white collar workers, we further split the sample into, let's say, fathers of younger children, mothers of younger children, older children, college, no college, black, Hispanic, Asian, no disability, disability private, and public sector. And we do see some important heterogeneities here. For example, black workers versus white receive a lower wage premium while working from home or remote work.

Then, workers with disability versus no disability also receive a lower wage premium. Government employees, surprisingly or not, also receive lower wage premium. But yes, we do notice that some groups that could be called marginalized, let's say maybe even you could put mothers of younger children into this category, but maybe we don't because the 12% wage premium, now this is the regression coefficient.

The premium itself would be what? The exponent minus 1 times 100%, or almost a little bit higher than this number, right? If you convert it. But all of this This category still receive a wage premium, not a penalty. So, hours differential, we do a similar analysis and notice that the hours have gone down, but these changes in hours are small, relatively small.

If the changes in wages are huge, the changes in hours are small. And we could see that, again, some occupations, the hours actually increased, personal care services, healthcare support. But in some occupations have decreased like legal protective services, agriculture and all. So, white-collar occupations, I guess, the increase was pretty strong in the blue-collar occupation too, but we see it in both.

And it affected both groups of people by parental status, especially fathers more than mothers. So, we don't see that much of an increase in hours, well, for mothers, the hours differential for fathers is stronger, right? So fathers who work remotely, they work less, whereas mothers don't. Then we do some analysis by subsamples and see that government employees actually reduce their hours when fathers of young children have reduced their hours more than other groups, and fathers of not so young children also.

Our final part of the analysis is to use occupational level data rather than personal level. We aggregate by occupation and by remote, non-remote groups, and use 2019 and 2021 data to ask the question whether wages grew faster or slower for remote workers relative to office-based workers between occupations.

And these are occupation level regression, where we have a control for remote status and interacted with 2021. And in some regressions we don't add mean demographic control, mean industry control, and some we do. And we find that the growth and remote wages in 2021 has been positive. So the wages grew faster in 2021 remote wages, it's about 4% faster than non remote.

We also look at the wage growth and how it relates to the share of remote work within occupation in 2021. There is a tiny, tiny relationship, so it's almost like a straight line, and the equation is on the graph. So what we see that the average percent of remote workers across occupations increased by about 15% during the pandemic.

So, we conclude that the rise in remote work is associated with about less than half of a percentage point increase in occupation level wage growth. So, very small contribution to occupation level wage growth. So, the key takeaways is that it was a substantial jump in the wage premium for remote workers during the pandemic.

And primarily remote workers earned 14% more than office-based workers. So the average real wages grew 4.4% faster for remote workers than for office-based workers within detailed occupation groups. The percentage of remote workers during the pandemic is positively correlated with the growth in occupation level wages. Usual weekly hours offer remote workers decrease steadily, and the data supports productivity effects more than the compensating wage differential story, okay?

Your questions?

>> Speaker 9: Yeah, great work. One piece of corroborating evidence, I feel behind your productivity story might come from the BEA data, right? So, they break down productivity by industry, and I understand industry probably doesn't map perfectly into your categories, but for some of them it probably does, like legal and so forth.

So, if you observe that productivity across industries kind of lines up with the fraction of workers in that industry that are remote. So, if industries that have more remote workers also see higher productivity growth, then I feel that would justify even further the wage evidence that you see.

One maybe another sort of competing story might be that there's unreported hours. So, employees are basically just being paid for working more total, so their productivity per hour isn't increasing, but there's like hidden hours. And I wonder if maybe American time use survey or other data sources might help rule that story out.

 

>> Emma: I was really encouraged to see your findings, this is so helpful, by the way, thank you very much. I was encouraged to see your findings that mothers of young children are seeing a similar wage premium to remote work compared to other groups. But then I saw again that the number of hours worked per week were not going down for mothers as much as for other groups.

So, it seems like maybe they're working more for the same premium, which is especially frustrating given in my project, we find that women with children, are perceived to be less likely to work extra hours. And I wonder if you can speak to that potential contradiction there.

>> Victoria Vernon: Yeah, we should also take into account the fact that many mothers or young children have exited labor force, right, during the pandemic, and we're only looking at full time, full year employees, so we're not taking into account that impact.

 

>> Speaker 11: Couple great work, a couple observations. One, Emma, on your comment, Kelly Yost has some data that, like pre-pandemic, the majority of remote workers were more often men than women in the first place. And I'm wondering how much this is a mix shift issue also, because if it was primarily a male dominated set of workers versus now you've got more gender mix.

What you may be observing is men's hours work dropped off partly because they're exposed to, God forbid, doing some housework and childcare during the pandemic. But that may be part of the mix that's going on. But my other question is from a mixed perspective, are you able to adjust for within occupation and industry levels of experience?

Cuz at least anecdotally I know a lot of Wall street firms are wrestling with the fact that the people that they're convincing to come in three and four days a week are the younger, less experienced workers who are afraid of losing their jobs. It's their managers who it's the more experienced people?

People that are staying home more often.

>> Victoria Vernon: We are adjusting for occupations where management is one of the major occupations and for age, which is the approximation of experience.

>> Speaker 11: Thank you.

>> Speaker 12: Many people here are familiar with the ACS data, but I just would like to remind that the way that the ACS asks the question confuses full time remote with hybrid work.

They ask you about how do you usually commute to work? Which confuses both and working with other data sets. One thing that I noticed is that it seems to be that the share of hybrid used to be that the share of hybrids, as a share of all types of remote work seemed to be lower before the pandemic than it is now.

So your results may mix both the change in attitudes towards work from home and the change in share of hybrid.

>> Victoria Vernon: Yeah, absolutely. We also don't know how people really, whether they respond as hybrid or as remote when they go to work twice a week or three times a week.

 

>> Speaker 3: Yeah, I was gonna follow up. So there's great to Shelby here as well. So in stuff with Shelby, exactly on this point, if you look at the ACS question, it actually has two sentences. So the first thing is, where do you usually work? Which you showed the problem is it has then has a second sentence, which is, please report the mode of transport used for most mileage.

And so the problem is, even if you work from home four days a week, presumably that's zero mileage, and then if you commute in one day. So when we look through Shelby, it looks like it's a measure of fully remote, cuz the numbers on this, it looks pretty much like a fully remote measure, which is great in some sense.

But one question will be interesting if there's any other analysis, when we've looked at it, you look at the levels. It lines up very well with fully remote measures, and it depends which sentence you read. So the problem is that the question isn't that well designed cuz the first sentence suggests include hybrid and the second suggests don't.

And I don't really know, but I think it's mostly fully remote either. I mean, Shelby's here, so we could all see.

>> Shelby: Thank you, at least compared to the SWAY survey, the ACS lines up mainly with fully remote, whereas the American time use survey does seem to include more hybrid workers.

So if you look more at us instead of the ACS, you may pick up more hybrid.

>> Victoria Vernon: Yes, our next project is on the time use survey.

>> John Taylor: So, yeah, the ACS has got the interpretation problems that we just discussed, but it is a big data set, which is its advantage and I don't.

You won't be able to do things at such a granular level with the AT US. But I wanted to first, just a clarifying question. The measure of the wage here is what? Hourly wage, weekly wage, annual salary.

>> Victoria Vernon: So wages, salary, income divided by number of weeks times usual hours of work.

 

>> John Taylor: Okay, so it's an hourly measure.

>> Victoria Vernon: One measure refers to 12 months before, and the other measure of hours refers to a weekly usual. I always work last week.

>> John Taylor: There's measurement error.

>> Victoria Vernon: But there's a little bit of that, of course.

>> John Taylor: But it's a, it's an hourly wage measurement, basically.

I didn't, maybe I missed it, you showed a lot of stuff. I didn't see any clean quantification that might allow us to get here how big the productivity effects are. And by the way, there's compensating differentials, there's productivity effects and there's wage markdown effects, which you didn't mention.

They could also be in the mix, I think we know from many other studies that it's not a compensate. The evidence you showed us is not a compensating differential story because all the evidence says work. The vast majority of workers like to work remotely. The only way to interpret your evidence through a compensating differentials is that they dislike the work remote.

I think we can set that off, set that aside. And so it's some combination of productivity effects and shrinkage of. It could potentially be a shrinkage of wage markdowns. Okay, right, that's another possibility. But what I wanted to see, I don't think I saw, but what I wanted to see is, say an occupation level analysis, the change from 2019 to 2021 or 2022 in the occupational wage.

The mean occupational wage related to scatter plot related to the occupation level change in the extent of remote work for those same occupations. And then controlling, using the microdata for potential changes in the composition of the occupational level workforce over time. I don't think you showed anything quite like that.

But that it seems like it's the slope coefficient in that scatter plot, that is telling me something I'm gonna set up aside, the I'm gonna assume it's not compensating differential story that would actually go the other way. The slope coefficient in that scatter plot is informing me about some combination of productivity effects and wage mark.

Shrinkage of wage markdown, increased productivity or shrinkage of wage markdown. That's kind of what I wanted to see, the end of the day, and I didn't see that. Maybe I missed it, maybe it's in your regression somewhere, but I don't think you had quite that exercise.

>> Victoria Vernon: Okay, we'll check some more.

 

>> Moderator: So Jose is gonna finish us off with why working from home will stick.

>> Jose Maria Barrero: Great, so thanks, everyone, for sticking around to the end. So you may be wondering why I'm presenting this, given this is a paper that we wrote almost two and a half years ago.

We put it out as a working paper. And so I promise you I'm gonna try and show you a lot of new stuff and basically trying to add to the evidence that we got earlier on in the pandemic to think about some of the consequences of the big shift to working from home that we document in this paper.

And that we are, you are going to stick based on the evidence that we originally documented in this paper. So kind of, these are the original research questions that I think we were after in this work. So how much working from home will there be as the pandemic ends?

I think we have a pretty good answer to that. Along the way, we collected a lot of evidence on the mechanisms that support a persistent shift to working from home. And so today I'm gonna focus on this third part, which is, what are the consequences of this persistent shift for workers and for productivity?

Basically, what I'm gonna do, is I'm gonna use data from our survey of working arrangements and attitudes, in particular data from late 2022 and early 2023. To present some key facts about how workers self assess their productivity while working from home, and in particular, how that relates to actual working from home outcomes.

I'm gonna use those facts to motivate a model for how workers and firms optimally choose working from home in 2023. And basically, I use the structure of the model to infer the relative productivity of working from home from those observed choices via basically revealed preference arguments. And then kind of with that model that we calibrate to 2022 and 2023, we can basically conduct counterfactual exercises.

What if we impose a return to office to the levels of 2019? What if we impose kind of lockdown levels of working, lockdown levels, working from home as in 2020? How does that change aggregate outcomes in this model economy? All right, so what I wanna do is, again, go take you through the facts.

So the key results are that, on average, workers say that they are more efficient while working from home. I think the keen sort of new result here is that Is that these self assessments of relative productivity among people who have working from home experience are highly correlated with the amount of working from home that they're actually doing right now.

This basically tells me that there's a strong revealed preference argument, and we can talk more about what this means. But I'm going to use this to basically, as a condition to discipline the model a bit in order to conduct these counterfactuals. And so kind of the key counterfactual, I think, is if we force the economy to return to 2019 levels of working from home in this economy, that would lower GDP by a bit under a percentage point.

Productivity would be aggregate productivity. This is, would be lower by 0.3 to 1.3%, depending on how you measure it. Welfare for the average worker would kind of. Here we have a range of estimates, so it can be as small as 0.1% or 2% cut to welfare productivity, cut to welfare of the average worker.

So let me start out with the facts about working from home productivity and preferences. Again, I'm going to be using SWAA data from late 2022 and 2023. I think I'm going to skip over the details about the SWAA. I think most of the people in this room are familiar enough about the survey.

In the background notice that the, working from home has stuck at least up until the period that I'm gonna be looking at. So I'm gonna be taking data from late 2022, from about October until June of this year, when we basically see working from home stabilizing at about 28% of full paid working days in the US.

And so during this period, workers are reporting higher relative efficiency of working from home and according to their own self assessment, compared to working in the office. When we ask them, how does your efficiency while working from home compare to your efficiency while working on business premises there's a big group that says that they're about equally efficient.

But on balance, more people tell us that they're more efficient at home, and there's a lot of dispersion. People's experiences differ dramatically. And yes, on average, people tell us that they're about 7.8% more efficient while at home. But this is masking a huge amount of heterogeneity, which is going to be one of the messages that I want you to take away from this.

Here's I think my new result is that if we look at how much working from home people are doing in 2022, and 2023 that's on the horizontal axis here. That is very strongly related to the average self assessment among each of these groups for how much better or worse they are while working from home.

People who are working from home basically full time say that they're much more productive while doing so. People who say that they who are not working from home right now are a lot less optimistic. And so I want you to take away more the gradient in these numbers than the overall level because it could still be that these self assessments are over optimistic.

Really what I'm going to use is basically this correlation. Finally. So to motivate the model, let me argue that what people like about working from home has a lot to do with how much time they save by doing so. This and the same goes for their self assessment.

Here on the vertical axis, I have again their self assessed relative efficiency of working from home. And on the horizontal axis I have their two way commute time plus whatever extra grooming they do on days when they work from home, clearly a positive relationship. Part of what makes me efficient while working from home is if I have a two hour, two way commute, that makes it really painful and makes my day kind of very inefficient.

If I have a ten minute commute, then I'm a lot more positive about working in the office compared to working from home. And we could see a very similar relationship when we think about worker preferences. Here the vertical axis is not people's self assessed relative efficiency. It's how many days, on average they would like to work from home in the long run kind of their desired level that if in an ideal world.

And then that is again, very much positively related to how long you commute. The model that I'm going to about, that I'm about to show you is going to have this element in terms of worker preferences, kind of commutes are going to be painful to people and that's why they're going to like to work for me.

Here's the recap of the facts. On average, workers report higher efficiency while working from home. The self-assessed efficiency rises with working from home frequency. Both preferences and self-assessments are strongly, positively related to commuting and grooming time. And these facts are going to be embedded in the model that I'm about to show you.

Here's the model sketch. And I'm going to try to minimize the equations as much as possible, but I couldn't manage to take them all out of the presentation, so bear with me. But basically, the model is going to be a static economy with several groups of people. So there's going to be workers that are going to supply a number of hours per week.

They're going to work a fraction of delta of their days from home, and they're going to receive a wage that can depend on the number of hours and on how much working from home they're doing, and they're going to consume the output good of this economy. There's going to be an intermediate firm that is going to hire these workers for a given number of hours per week.

Again, a fraction of those hours are going to be work from home hours. I'm going to make the assumption that the intermediate firm is a price taker. They're going to take wages as given, and these firms are going to produce efficiency units of labor from these raw hours of work, from a given work.

There's going to be a final goods firm that basically takes all of the efficiency units in the economy and transforms that to aggregate output. If you're a macroeconomist, you might like that part. I'm not going to say anything more about this final goods firm. It's just to close the model.

The equilibrium concept that I'm thinking about here is a static equilibrium in which, again, the firms are price takers, but the workers are going to internalize the impact on their pay of choosing a different amount of hours and choosing a different amount of working from home, especially kind of in the more complex versions of the model.

That's going to be important. I think if there's one thing in this presentation that I'm not sure the final version is going to look like is the equilibrium concept. I think we're still experimenting there, and I can talk more about that if people are interested. Here's basically what workers look like in this economy.

They get utility from consumption, and so they like making money and they like working a little bit more because that allows them to consume more. But they dislike spending hours on work and on commuting. So kind of their total hours devoted to work, which they have this utility from, include the measured hours that they officially get paid for, they dislike, as well as kind of the baseline grooming that they do every day, and their commuting costs, which include kind of any extra grooming that you do when you go to the office.

I don't know. It might take longer to put, put on makeup, to shave, put on a nice jacket to come into work, but you only have to incur those costs on days when you commute, not on days when you work from home. The firm that hires these workers has a production function that looks like this.

Each worker has an overall productivity shifter that is basically accounting for the human capital of an individual worker relative to others. Then this firm is going to generate efficiency units of labor by combining in person hours with working from home hours. Kind of the new stuff here is in the working from home hours, where b is going to shift the relative productivity of those.

Remote hours. And Alpha, this parameter over here is going to govern the substitutability of working from home and in person work. And so we've had a lot of discussions over the past couple of days about how some tasks are more amenable to remote work, some are not. And so we're going to make the assumption, and in our calibration, it's going to turn out that alpha is going to be less than one.

So there's effectively decreasing returns to having remote hours. So basically, eventually, if you try to do everything remotely, kind of, that last hour of remote has less returns than the first hour of remote work, because there are some tasks that are really poorly suited for doing remotely. Okay.

And as I said, this firm, the assumption that I'm making today is they're going to be a price taker. And I should add that basically. So in the data, there is a very strong evidence that when people don't commute, they actually work more hours. So some of the time that they save, when they don't commute, they devote to their job.

And so I'm not showing you in the equations that I'm putting up on here how that enters because that really makes the math more complicated. But in all of the final numbers that I'm going to show you today, I'm going to take into account that basically when workers don't commute, they work a little bit longer.

And I'm going to, and that's going to change a bunch of the numbers that I'm going to show you. Okay? So the nice thing about this model, and again, these are the equations that come out of the very simplest version, is that based on the equilibrium conditions and on what firms and workers are optimally choosing, I can infer a lot of these unobservables.

So I can infer what the relative productivity of a given worker is, what the relative. So given alpha, I can infer the relative productivity of working from home, B, for every individual worker in our data. And basically all I need to do is to calibrate alpha. And so here's where I'm going to use the fact that self assessments are predictive of how much workers are actually working from home.

So I'm going to try to minimize, so I'm going to try choose an alpha that minimizes the distance between these Bs that I infer from the model and the self-assessments that I get just from the raw survey data. And that's going to give me an alpha of 0.97.

That tells me that basically, remote hours are very close to perfectly substitutable with working with in-person hours, but not perfectly substitute. So here's basically the distribution of productivity shifters that I get. So this is the vector of Bs for a bunch of these people. The relative productivity working from home seems to be zero because we see so many people in 2022 and 2023 basically not doing any remote work at all.

The model is basically inferring that these people are terrible at their jobs when they're remote, and that applies to about 60% of the sample here. On the other hand, there's a lot of people for whom there is a relative productivity of working from home that is positive for many of those that b is greater than one and b is greater than one.

B being greater than 1 means that there are lots of returns to doing remote work for these people. Okay? And if you want. So, just as a sanity check. So I calculated the average B. So this average relative productivity shifter across industries, and you see this makes a lot of sense.

So in information, finance and insurance, we see the average b is quite high. In transportation, warehousing, hospitality and food services, it's a lot lower. And basically, what's going on in the background is what's really determining these b's through the structure of the model is how much people are actually working from home in our data, and our measures of working from home are giving exactly these patterns.

Let me move on to the counterfactuals. What are the consequences of changing working from home levels, from changing working-from-home levels to what we saw in 2019? So, basically, the thought experiment that I'm doing is I'm going to exogenously set working arrangements, and then solve for the new equilibrium in this counterfactual economy, and compare aggregate outcomes in that counterfactual to the aggregate outcomes in the model based calibrated to 2022 and 2023.

And so, 2019, working from home levels here are going to mean that basically, you only work from home about 10% of the time if you're currently doing it to any extent. So anybody who is not working from home now won't in 2019, and basically anybody who is in hybrid or fully remote now is going to be scaled down to a level of working from home, where they come into the office about once every two, sorry, where they stay at home about once every two weeks.

The rest of the time they're in the office. And so here's what we get in terms of impact for the aggregate economy. So GDP goes down by almost a percentage almost 1% measured productivity in the sense of output per measured hour goes down by 0.3%. But I think the important, or one of the important lessons of this exercise is if we really think that basically, in a world where you have to commute, you should really count those inputs into producing market output.

If I then add into my hours of input, into my measure of hours worked, all of the commuting and grooming time, this effect on productivity basically multiplies by 4. There are also potentially some big impacts on welfare. So if we hold prices constant, we get that workers on average are worse off by an equivalent of 1.6% of consumption.

In equilibrium, that sort of washes out to a much smaller number. But I think the bigger point when I think about worker welfare is that there's a lot of heterogeneity. And so here, this is the distribution of changes in welfare for individual workers from going from 2022, 2023 to the counterfactual.

And, yeah, so I can give you an average number, but there is so much heterogeneity here that it makes me think that the average number is almost meaningless. And kind of across all the counterfactuals that I've calculated, the amount of dispersion is so huge that that it is meaningless to think about the average number.

All right, so let me very quickly, so going to 2019 has some negative consequences on the economy. Productivity goes down. GDP goes down. Let me just do two more thought experiments to give you an idea of the sorts of things that we can do with this model that I think are a big contribution of this exercise.

So what happens to an economy when we impose 2020 levels of working home? So, basically, in this case, I'm thinking about imposing working from home at 100% for anybody who remotely can and not if you obviously cannot. So kind of what happens in this world? Well, it turns out this is also very costly.

So kind of, even if I try to make things not as serious as, even if I try to find a calibration of the model that is not as negative as it could be, I get that the cost of going to lockdown levels of working from home is about 11% of GDP, 11% or 10% of productive in terms of productivity.

So lockdown is very, very costly. And I can do the same thing with worker desires. So in this way, I think one of our more successful questions that we've been asking more or less continuously since 2020 is, is it was this question of how much would you like to work from home after the pandemic.

And so with this model, I can make a small modification so that instead. Instead of people liking working from home because they avoid commuting, maybe they have a preferred level of working from home, and they experience a loss if they're not at that level. And I can be a little bit agnostic as to where these preferences are coming from, and I can do a counterfactual where I give everybody their preferred level of working from home.

So what comes out of this exercise is, again, this is quite costly to the economy. Workers failed to internalize kind of the, failed to internalize the productivity benefits of being in the office. And so if we give all of them exactly what they want in terms of working from home, the model is saying that GDP would drop by about 5%, measured productivity by 4%, and kind of our expanded measure of productivity by a similar amount.

So again, this is very costly. So let me leave it there. But basically, what I want you to get out of this is, I think this is in some sense, an advertisement for the sway, in the sense that we have many, many data points and we've asked many different questions that can tell us about how workers and employers are choosing working from home.

And if we impose some structure on these data by thinking about a model like the one that I showed you, we get that we can do these counterfactual experiments that tell us things like forced return to 2019 leads to a 0.3 to 1.3% loss in productivity. It matters for work or welfare, and there's a lot of heterogeneity there.

And looking forward to see what people think, this is absolutely fresh, even though this is a two and a half year old paper. So thanks for, yeah, so yeah, thanks for listening and hope to hear what you wanna say.

>> Speaker 7: So, first of all, this is a paper I guess all of us have read many times.

So these are really cool new results, and you've done enough for one paper. So I don't know if these suggestions are super helpful, but I was thinking about the intermediate firm and whether you can also add a matching model there, because I imagine that firms which offer more work from home would attract, there'll be some selection in terms of better quality workers.

And the reason I say that is in your counterfactual analysis, instead of scaling up and down uniformly for all firms, you could scale up and down heterogeneously. And then you could even tell us the predictions about what would firms that scale down less or more, what kind of productivity impacts would they see.

 

>> Jose Maria Barrero: Yeah, that's a great suggestion. And that's the sort of counterfactual that I think we can, we can think about structuring in this environment.

>> Speaker 12: Yeah, this is fantastic. As somebody who spent a lot of time trying to model work from home, I would love to see that you have a model.

By the way, in our work, we use a very, very similar structure of the model. And one of the pushbacks that we receive all the time is, what about firms in your setting? Workers decide entirely how much they want to work from home, and firms just take whatever output workers produce and pay for it.

Whereas what we see in practice is that very often, firms are very proactively deciding work-from-home policies. And if anything, they choose how much workers work from home, or at least in some kind of collaboration with the worker, right? And I'm just curious to hear your thoughts, how the model could be amended to take into consideration what firms want to do.

 

>> Jose Maria Barrero: Yeah, so I think that's a great point. And I think conceptually, knowing what is the mechanism at which firms and workers have arrived at the current level of working from home is difficult. And so I took one stand in the presentation. I've solved a different version of the model where the worker takes as given what the firm wants to do, and the equilibrium is set that way.

And so these are the numbers from that version, you get something fairly similar. So I'd say kind of for the 2019 numbers, they're a little bit bigger. So it's 1.9 instead of 0.9, it's 0.8 and 2.1. But I think kind of the key lessons are, in some sense, the same.

That there are costs of productivity and output, if we force a return to 2019. And if we take into account that commuting hours are potentially an input into producing market activity. Again, kind of my productivity number in this case multiplies by more than 2.

>> Michael Bordo: CEO's are arguing that people need to come back to the office.

So are they right or wrong? First line says that they're wrong. The last line says that they're right. So which one is it?

>> Jose Maria Barrero: So, I mean, so let me be very clear. Here I'm assuming that basically, the equilibrium in 2022 and 2023 is set to optimize what workers are doing and in some sense, what firms are doing.

So any deviation from that should generate a loss, the question is how much? And I think so people who want a full return to office, these are the potential costs that you should be willing to take. I think, yeah, telling the workers, you can just do whatever you want, that's also gonna be very costly.

It's going to be more costly, but I don't know that it's in the cards to give every single worker, white collar firms exactly the amount of working from home that they want. I don't think anybody would realistically give that. I think it might be easier for me to buy the people who are realistically thinking that we need to come back to 2019 levels of working from home.

And then so that's why I think that's the focal counterfactual.

>> Speaker 16: So, interesting. Just a couple of questions. One is the main measurement of time and the commute. Does that mean that you can expect those countries and cities with smaller geographies and different topographies to, in fact, show different figures?

And the second question is the other forms of measurement, which is, I believe, without the data to back it up. But I believe very strongly that there are other non commute reasons why people are reluctant to go back to work in the old ways, not least the toxic workplace, bad management.

How can you begin to start to measure that, if at all?

>> Jose Maria Barrero: Yeah, so both great questions. I'd say yeah, so I think this model is telling us that the commuting costs are a big component. And to the extent that commuting is less painful because transit systems work more appropriately, traffic is, is better, I think that's gonna definitely lower the costs of going back in person.

Because people are gonna be happy to do it. They're going to cut back on their hours less when you force them to commute. And so, I mean, definitely we can say more on that, but I think that's the direction that it would go. In terms of preferences, I think.

So the first version of this model that I did didn't really incorporate any preference heterogeneity, but then I realized, hey, actually, we have a very good measure of preference heterogeneity that we've been collecting since 2020. So in the sense that we know how much each person in our data would like to, ideally, work from home.

And so it took me a while to find the exact specification of the model that would work. But I can give you. But we can very much quantify kind of what is the value to these people of doing exactly what they want. And we can definitely do more on that and look at patterns as to how this varies by observables like gender, age, etc, as well as unobservables.

 

>> John Taylor: Elaborate on two sets of remarks. Just first, what Jose was saying, because we can either use the data directly or we can use it as. Filtered through the model, it is telling us something about how much people have a desire to work from home, conditional on their commuting and their productivity characteristics.

We can go back to the individual level data and relate the average values of these gamma i's to a whole set of observable characteristics.

>> Jose Maria Barrero: Exactly.

>> John Taylor: Gender, education, marital status, kids in the household, and so on. So even though we haven't done it yet, we have a data set and a framework that's very well suited to addressing the issues you raised.

So, second thing, back on this question about, well, two more things. Are the CEOs right or the workers right? And that last line says the CEO is a right from a productivity perspective, doesn't mean they're right from a welfare perspective. Okay, so the people are making these productivity wage reducing choices, but they're nonetheless better off, okay?

 

>> Jose Maria Barrero: That's right. So workers are, on average better off if we give them what they want. That's totally right.

>> John Taylor: So I just wanted to make that clear. The third thing back on this issue that you raised about employers differ. Employers are offering different contracts. So just two things.

First, Jose went over this part quickly, but the employers in this model are offering a wage, which is a function of both your hours and the workers choice of working arrangements, that thing is determined in equilibrium, okay? They're competitive firms, but it's determined in equilibrium. So that wage working arrangement bundle, so to speak, is an equilibrium object in this model.

This version of the model has one intermediate firm. There's nothing to inhibit us from taking this model and have two different kinds of intermediate firms that have different technologies for transforming hours and working arrangements into labor services. They can both be operating in a competitive market and competing with each other.

What does that allow us to do? Within basically the same kind of competitive equilibrium structure, we can have two or even n firms offering different kinds of wage working arrangement bundles in the market. To capture some of this evident heterogeneity that we see in the real world, where the Elon Musk Jamie Dimons of the world, they seem to be operating with a different production technology than firms that are quite willing to accommodate flexible working arrangements.

So that's not in the current version of the model, but it's well within our scope to do that in this framework. And I don't think it's analytically any harder than what we've already done. It's just another layer of heterogeneity.

>> Speaker 17: So if I've understood your model, the equilibrium is always optimal from a productivity standpoint.

And so the counterfactuals that you're talking about here are not counterfactuals that you'd ever really see, because they're counterfactuals that would reduce productivity. And so I'm wondering if you'd considered calculating counterfactuals that involve changing some of the parameters in the model that I think of as really exogenous.

So I guess that would maybe be the gamma i's. Because those are counterfactuals that we might actually see in the world and lead to a new optimal conditional on the gamma i's, a new equilibrium. Maybe the way to do that is to do what you just said, which is project the gammas onto observables.

And then that would allow you to ask questions like, if people's preferences were different, what would the level of work from home be? What would the productivity implications be if men and women had more similar types of preferences, if commute times were lower, those types of questions.

>> Jose Maria Barrero: Yeah, no, so I think that's an excellent suggestion.

I mean, maybe I would think about it slightly differently. So I'd say kind of the productivity shifters for working from home that we're inferring here are basically embedding any changes in technology and human capital, basically embedding all of the learning and changes and investments that we've experienced over the past three years that enable working from home now in its current form.

And, yeah, and I think of a very. So maybe the thought experiment that is coming to my mind right now is basically how much would we have to shift these B's to get workers and firms to choose to go back to the office, as in 2019? Yeah, that's the counterfactual that I'm thinking about now, or the thought experiment that I'm thinking about.

 

>> Speaker 11: This is really interesting, it's making my head spin a little bit.

>> Speaker 11: Going back to some what Steve was saying in response to Julia and Raj's earlier conversation. It'd be really interesting to me to actually find a way that you can actually model out different behaviors of firms and what impact that has or how this actually works across demographics.

If I'm wrapping my head around it, right, the answer to the CEO question is it's worse because the answer to the employee question, if the CEO gets what they want, that's not what we are doing today. So therefore it's gonna drag us back in a different direction. But also, if the employee gets what they want, things get worse.

Because we're starting off with an assumption that we're at an optimal in 22 2023. So neither of those people are actually right. Ends up being the answer. If you just look at it on that face value. I think it'd be much more interesting to look at this in some way in which you could model out what if firms behave differently and that that attracted labor towards them.

Because that's what's happening in the market, we're seeing it in the real world, right? Firms are actually applying. Within every single industry, there are examples of people that are applying different models and they're doing it on the basis of to attract talent, not because they necessarily think they're gonna save on real estate.

That's what I think would be really interesting out of this, if there's a way to do it.

>> Jose Maria Barrero: Yeah, I think I need to think about how to operationalize that, but it's definitely an interesting thought experiment.

>> Moderator: This is super cool. I think probably somewhat motivated by my work and Charlie's work.

I think it would be interesting to try to think about spillovers across workers. I understand that's probably gonna be tough to do, but if I recall correctly, you have some question about, like preferences about coworkers coming in.

>> Jose Maria Barrero: Yeah, and we haven't used any of that.

>> Moderator: Yeah.

Like, I feel like on either putting it in as a preference shifter or as a productivity shifter, it would be interesting to think about, like, if we're thinking of a coordination problem within the firm, like, how would you put that into this model?

>> Jose Maria Barrero: Yeah, no, that's a great point.

So off the top of my head, I think Charlie's better suited to thinking about and modeling spillovers than I am, given my skillset. But, yeah, I think those are super important questions and maybe there is a way that we can use. Exactly. Again, this question that we've asked in the survey and think about, okay, how does that show up in the model and incorporate it?

 

>> Moderator: Have you asked anything about how co workers coming in impacts their productivity?

>> Jose Maria Barrero: Yeah, so we asked about it about mentoring, and we asked it about preferences, but not about their self-assessed productivity while at work or in the office with colocation, no.

>> Speaker 18: I'm wondering how you're thinking about the nature of the pandemic shock to working from home.

When I look at the model, it looks as if maybe you're thinking that it involved investment in new technologies that made it easier to work from home, and that's where all the action is coming from. And now we're in a new state. But it also seems to me there's likely some amount of, given how well some people managed to work in 2020.

It seems there's also likely some amount of firms were poorly informed about what was possible. How do we think about that difference? There was a forced experiment that nobody had been willing to take before that taught us about technology versus we invested in technology to improve things.

>> Jose Maria Barrero: Yeah, so the best I can say is kind of, so the parts of the paper that I haven't shown you are using data that we collected earlier in the pandemic to make exactly these arguments.

That there was a forced experiment, that people were forced to learn that on average what they thought before the pandemic was overly pessimistic, that they made investments into working from home. And so how I think of the model is, again, I'm calibrating it to data in late 2022 and 2023, likely after all these dynamics have subsided.

And so now what we observe people doing late in 2022 and 2023 incorporates those investments, incorporates that learning, incorporates those changes in human capital that are enabling working from home to persist to the levels that we see now. And I taking the counterfactuals, basically, from that baseline.

>> Moderator: Thanks so much.