How Big Is The Media Multiplier? Evidence From Dyadic News Data? | A Workshop Series On Using Text As Data In Policy Analysis

>> Steven Davis: Welcome, everyone, to the Hoover Institution workshop on using text as data and policy analysis. My name is Steven Davis, Justin Grimmer and I select speakers and moderate the workshop, Tara Mahan is our master engineer and Cecilia Chen keeps us organized. Today's guest is Thiemo Fetzer of Warwick University, he will present his paper titled how big is the media multiplier?

Evidence from dyadic news data, It's a paper co authored by Timothy Besley and Hannes Mueller. Here's our format, Thiemo will take about 30 minutes to present, then we'll turn to a discussion. If you have a question or comment, put it into the Q&A box, and depending on the flow of these comments and questions, we may paraphrase them, combine them, or ask you to state them directly.

We'll run for about 60 minutes, then we'll turn the recording off and we'll have another more informal session for anyone who wants to stick around that's offline, that won't go into the videotape session, okay? With that, Thiemo, the floor is yours.

>> Thiemo Fetzer: Thank you so much for having me, it's a great pleasure to present this work.

This work actually started in 2014, so it's been a while, it was just recently accepted for publication, so this is joint with Tim and Hannes. What we do in this paper is essentially trying to quantify the extent to which selective media coverage can distort the economic effect potentially or exacerbate potentially risk perceptions that are associated with relatively rare events.

One of the motivating figures that we have for this work is from our world in data, where if we look at what the media is reporting on vis a visa, what people are, in this case, dying off, there's a vast difference in reporting skew. So if we look at the number of news articles that are in the Guardian or the New York Times that cover, for example, deaths that are related due to terrorism or homicides.

They take up a disproportionate amount of media real estate, so to say. Whereas actually, in terms of the actual causes of death, the silent killers like cancer and heart disease obviously matter materially much more. We study this phenomena in the context of, obviously this is relevant for economic decision making because factors like heart disease have a much higher human death toll.

But humans seem to be drawn to kind of gory images, if it bleeds, it leads. And nowadays, where media has moved more into being in the marketplace of attention rather than in the marketplace for information. And in this paper, we study essentially the distortion that this can induce, essentially amplifying economic cycles or potentially economic shocks.

So I don't wanna motivate this much further, I think I've done a lot of motivation in this one slide already. What we do in the paper is to model the economic effects of negative new shock, and we study the extent to which they are amplified by media coverage.

And through that, we try to quantify what we call the media multiplier as sort of approach that is grounded through a formal model, a data generating process that we posit. And that we then bring to bear on data that allows us to both address causal identification challenges, as well as also provide us with a setting where I think we can do quite decent quantification.

So let me just motivate the story a little bit, or the specific context. This is a type of violent event that we're studying in the paper, which is the SUS attack in Tunisia that took place in 2015. When about several gunmen started shooting up tourists that were just visiting a beach.

And that has resulted in 38 fatalities in this specific case, the vast majority of victims in this particular instance were British nationals. And this event dominated the news cycle for about a week, and even then resulted in the government, the UK government, to repatriate its tourists while German nationals were not evacuated.

So just imagine, this is a very strong and pronounced shock and had significant implications for policymaking, and obviously also had economic implications. We can study this through a granular card data, payment card data, that it took us, I think, three years to secure. Actually, four it started in 2014, which basically provides us a measure of service sector trade, which is what tourism is, tourism as a form of service sector trade.

And here what I'm plotting is the activity, the card activity that's associated with British issued, British bank issued master cards, it's all aggregated and non-disclosive data that we're working with here. British cards issued that are used in Tunisia over time, and German cards that are used in Tunisia over time.

And what you see is around the month, it's monthly data, we see that there is a significant kind of decline relative to after having residualized, removing some lever shifters, as well as trends. From the data, we see that there's a significant decline in the sort of amount of spending that can be attributed to british issued cards in Tunisia.

And that effect has been much more pronounced relative to german issued cards in Tunisia. Ultimately, that is sort of the differential that we're trying to explain to what extent this type of differential. Of course, this is a very specific context, a very specific event that was very high intensity.

But essentially, what we're trying to do is explain to what extent this differential response can be attributed to the different intensity of media coverage. This is not the archetype event that we see in our data, but it's sort of one where it's just very plain obvious in the raw data to visualize and motivate the exercise that we're doing in this paper.

So we're exploiting dyadic data in two ways, It's dyadic data because we have network data, both from the side of the card payment cards that are being used. So we see that in five prominent tourist destinations, we've gotten data on cart activity, both the number of active cards, the number of transactions, as well as the spend.

That arises from about 128 card issued countries. So it's granular, high frequency, monthly service sector tourism data covering the time period from 2010 to 2016. We got the data in, I think, October 2017, after searching for three or four years of finding a partner that was willing to share that with us.

The second, and that's also novelty in this type of work, is that we've built a dyadic news corpus dataset. So we're capturing data on how others are reported on each other's news, which ultimately, is what we want to study here, which is. How, for example, a violent event that is taking place in Tunisia is affecting the media reporting or the media representation of Tunisia in a set of potential tourist origin countries.

So think of this as like a directed graph. But of course, since the five destinations are predominantly service sector exporters, meaning they import tourists, they have tourists arriving at them. And the focus is the service sector trade from the perspective of mostly western European and so on countries into these destinations.

I think when we started this, I think we're probably one of the first to build such a dyadic dataset on the news data side as well as on the card side, and combine that together in that paper. Let me just talk a little bit more about the data description and diagnostic, how we sort of leverage the corpus, how we try to filter the news articles that are diet specific.

And that are capturing violence using, I mean, I think nowadays probably pretty redundant NLP techniques. And how we then take that to bear on the data building reduced form exercise, but also then building what we call a data generating process representation that we then let bear on the data.

So we've obtained this confidential, aggregated data from Mastercard, capturing these three things from 2010 to 2016. It's a monthly data set, so they wouldn't part with more granular data for confidentiality reasons or trust reasons or whatever. At the time, this was fairly new for them to actually build these type of partnerships.

And I think we're probably one of the first ones that actually were benefiting from such a data donation. The first time I actually, it's a nice anecdote. When I spoke with some representatives of Mastercard's main competitor, they were actually quoting, I think, a single time series of the data that we're working with.

The quote was about $10,000. And I said, well, this is some business you're in, but maybe you get your business wrong because you should be probably in the type of consulting business rather than the data merchant business. And I think some of the firms in this sector are now picking up on this quite significantly, but so do other players that are entering that realm.

So we construct a diet level panel data set. We're focusing where we have data for 140 credit card issuing countries, it's credit and debit card, and we focus on a subset where we have a balanced panel just to make things a bit more, more. It's just a bit more elegant, but it doesn't actually change anything fundamentally.

The second data source that we leverage is also quite painful to collect because we don't have APIs and whatnot. Or we didn't have API access was content data from LexisNexis and Factiva that essentially through which we try to identify the population of all destination specific news coverage. So what we did is for every tourist originating country, we've tried to identify a large consistent source that is available over the whole time period.

Actually with one year lag as well, that we query essentially in terms of whether the country name or prominent city location names appear in that article. That would define the population of articles that are capturing a destination. So it's Tunisia, Morocco, Egypt, Israel and Turkey that we're working with.

And this essentially sort of provides us like the corpus of all the articles that are covering said destination. We actually literally had to download these piece by piece. We had some RAs, like they were downloading the data in batches of 100, which was a very, very unhappy job because at the time we just couldn't afford or it was not possible to get API access to either LexisNexis or Factiva.

In total, we have about half a million articles across 20 different languages that we then, for convenience, translated to English. So what we need to do naturally is identify among this corpus the articles that are likely indicative or capturing violent events. That we think are what drives the underlying mechanism of how people who might selectively read news might respond to in terms of their consumption or travel decisions.

Just to give you a sense of what this looks like, and there's usually always some feedback on this, which countries are being covered and what is the source, origin, origin, and is it appropriate? Is the New York Times an appropriate representation of a newspaper in the US? When I originally designed this paper, and there's a dual paper of this where we look at reporting on NATO casualties in Afghanistan.

The idea was to find both a left leaning, a centrist, and a right winning newspaper for especially the democratic countries that were troop contributing nations. It is simply for the content aggregators that we had access to. It was hard enough to identify a consistent source that goes back to 2009 over the time period.

But that was originally the idea, to look at the different way that the same story might be slanted or spun across different types of sources. As a usual question I would come up. We also flag up that in some instances, the only source that we had consistently available are news agencies.

So for China, for example, Xinhua, and that obviously is not a representative source. But all the results that we work with are robust to the exclusion of news agencies or agency sources. Because, of course, there's two factors that matter. Whether something gets reported on and whether that something actually is appearing, let's say, in the daily news diet that the average household or consumer might be subjected to.

Yes, so this is sort of just a comment here on the sources. In terms of the countries that we cover, the dark gray countries are all the countries for which we have both data on the underlying sources. So we have both a news corpus as well as. So we basically cover, pretty much on a substantive terms, much of the world economy and all of the G20 countries.

So it definitely captures most of the countries that would be service sector importers via sending out tourists. In terms of the machine learning, I mean, I think this is, I mean, now with LLMs and so on, so forth, this is actually pretty simple. What we do is we use human coders to classify a stratified random subset of the articles.

This half-million articles where the stratification was done because obviously violent events thankfully, are sufficiently rare. But it also causes a problem because obviously any classifier would have a tendency to just classify all articles as being non-violent or capturing reporting on nonviolence if 95% or 96 98% of the articles are covering nonviolence.

So what we're doing. Doing is, we use some conflict event data sets, a broad menu of conflict event data sets that are ultimately we consider to be the superset of all the, let's say news about the actual factual events. We take this as an anchor and then oversample articles that around like a one week window or three day window around an event, so that we address the inherent class imbalance problem through that, through the training data selection.

So we use this supervised machine learning where we classify articles in two ways, whether a text is indicative of whether there was any violence, or whether a text is indicative that tourists were the target of violence. These are not nested, but you can think of them as being supersets.

But in terms of the tourist targeting, obviously this is a much more narrow, much more specific type of article. We used a very simple naive Bayes classifier and two sets of random forest classifiers that we sort of process a little bit iteratively, where we sort of recode post the first run of it to identify essentially to see how many articles we might miss, or the classifier has a low performance in some area domains.

Since the analysis is all done with a monthly dyadic card spending data, we have to essentially turn that daily data set or article level data set into a monthly aggregate. And our preferred measure of violent news reporting is expressed as a share where, and we talk about this in the modeling section.

The numerator is the article, the denominator is the number of articles where we have to add plus one in case a country never appears in the news of let's say Tunisia never appears in the German news, which has very interesting non linearities and features that we try to represent in the model.

Since it's a short presentation, I can sort of just quickly rush through this, I think I've said most of this. We use an ensemble agreement to classify an individual article aggregating. So the linear modeling that we get from the naive Bayes with the nonlinearities to try to sort of combine some of the different features where random four seven, tendency to overfit, whereas naive basis is linear.

We're trying to trade off the two in a very agnostic way, but by giving each of them a third overweight, we could have optimized this as well. We're not using the Bayes-optimal cutoff but we're using sort of something that because since we aggregate the data, the aggregation helps get rid of some noise in the process.

We work with slightly different cutoffs but all the results are robust to using different cutoffs. Just to look at the class imbalance, out of the 450,000 articles, 16,000 are covering, or 17,000 are covering anything related to violence, and about 1000 articles are covering violence with a specific mention of tourists being targeted.

This is just the headlines of some example articles with tourists being with the general violence topic. And there we see we're picking up quite a bit of let's say, kinetic force or military type engagements. So like clashes of PKK rebels with the Turkish military or sort of violence on the Egyptian Sinai, which might not directly be relevant for tourists, but it's obviously a country specific risk that matters.

This is the other outcome or the other classification or the other label, so to say, which is tourists being the target of violence. And again, if we look at the articles, the classifier seems to perform reasonably well. Let me now just walk quickly through the reduced form evidence and then the model, and the model fit.

There's a whole lot of exercise that just tried to situate this work within the existing literature on this. We look at violence data, violence separately from news is there content in the news reporting that seems to explain variation above and beyond what simple violence level controls would do, which is what the first generation papers in the early two thousands did on, like regressing violence on some economic outcome.

So we have this sort of news measure which captures the share of news on a specific origin like home country, destination country pair. We lack this by one month in this very simple reduced form setting. And so we have a whole exercise justifying the use of that share measure, which is sort of backed up by again our modeling exercise that we do.

So if we just look at the reduced form, what this is saying is that, well, if you move, if a country pair moves in its reporting from 0 to 100%, you see a near 80% decline in tourist activity. And that is quite robust to multiple different ways of accounting for linearities or nonlinearities, different time trends, origin, destination, specific linear, nonlinear time trends, and a whole lot of other exercises that are now sitting in various appendices never to be, yes.

 

>> Steven Davis: Just a clarifying question. An increase in violent news share to 100% means all articles about that destination country and that origin country or about violence. Okay, so that's a huge shift, what would be something more within the or maybe that is within.

>> Thiemo Fetzer: So that's exactly part of the story which I'll try to highlight in the model.

 

>> Steven Davis: Okay.

>> Thiemo Fetzer: I'll get back to this, if I don't please remind.

>> Steven Davis: Okay.

>> Thiemo Fetzer: We also do some iv that exploits because of course you might be concerned that terrorists might target specific tourist nationalities which there is some evidence on tourists terrorist targeting. So the timing might not be exogenous.

I think what we do here is we exploit variation in the casualty distribution, the nationalities of casualties for some events where we're able to identify the casualties, the country of origin of the casualties, to then look at whether we can document this treatment effect through spillovers. So the idea is very simple.

We're saying is, well, what is the effect of a German casualty in Tunisia on the spending of Austrians in Tunisia? And we ignore the full direct treated diets, which of course highlights that we will likely have a violation of the Sutva assumption because there's information blowers, because the German cavity is more likely to be reported on in the Austrian news because we share a common language.

So this is just a cute extension which sort of highlights. I mean, this is a super powered instrument, but we can identify these treatment effects through these. These information spillovers, which address a whole lot of endogeneity concerns. There's a dynamic to this, but let me maybe jump straight to the modeling section because the reduced form is cool, but also a little bit limiting.

And for the model, what I like to do is present this as sort of like, we had this idea of like, can we represent this in, that looks a little bit like plate notation. And so this is like a very rough attempt of late notation. We posit that there is essentially countries, there's a latent state that's unobservable, a country can be safe or dangerous.

That's sort of evolving according to like a Markov chain process. And that Markov chain is governed by some transition probabilities and persistence parameters. What is observable is the actual violence time series in a specific destination, and the violence process is governed by this latent state. And then if the country is in the violent state, you draw from violence realizations.

In this case, we assume the simple normal distribution with a state specific normal distribution with mean and variance. The other bit that's observable is the bad news and all the other news. So that's in the bottom handed to notes, slightly gray shaded. So we have bad news that it's origin, destination specific, and all other news, which again is governed, the evolution of that is governed by the latent state evolution.

Now, we posit that there's two types of consumers. One of them I recall, and there's no normative notions to these labels, there's the sophisticated and non sophisticated consumers. The sophisticated consumers are the ones that essentially consume the superset of all the news, which should cover just the raw facts.

So think of this as like the superset of all potential articles, which means it's our best estimate of what the ground truth violence data is in a specific destination. So this is capital PI DT. The other set of consumers, we call them naive. Again, there's no normative notion here to this labels, but they predominantly consume news that's available to them through the media reporting.

So they form their beliefs about the latent state and the evolution of the latent state based on consuming media reporting and of course, bad news and all other news. This is essentially the inputs to our model that we then bring to bear on the data, we incorporate a lag structure.

And obviously we allow there to be different weights of both, let's say naive versus sophisticated tourists or potential travelers to be in there. The lag structure we embed to incorporate the delays or the forward looking nature of some planning travel decisions. Because, yeah, I have about three minutes, so I'm not gonna, if that's okay, not bore you with this in terms of the formalization.

Our sophisticated tourists, as I said, they observe the objective violence data when forming beliefs about the underlying state, whereas the naive tourists mostly rely on the news reporting. The evolution of the violent state is governed by a Markov process that we then sort of bring to bear on the data.

And this is sort of the output of that Markov switching models. That is the probability of the state being dangerous. That's just imputed from the extractor from the country level violence time series data, so to say. So you see the switching and not switching. And for the naive tourists, we model the arrival of bad news and all other news as being governed by a negative binomial distribution that's conditional on the latent state.

Where we calibrate the arrival rate of news, bad news, based on just the overall share of the probability of news being of an article being violent, basically. Through this, we can sort of apply Bayes rule to study the evolution of both the news based belief as well as the sophisticated beliefs.

And get essentially some of the nonlinearities that the Markov switching model implies into the model that the reduced form modeling cannot directly incorporate. This is a distribution of the beliefs for naive tourists about a country being safe versus dangerous. Again, because we do this for both latent states, when a place is dangerous versus a place is safe.

And we then can estimate the median multiply is essentially down to the mixture of consumers in a representative consumers that form these beliefs based on naive versus sophisticated belief. This is sort of our conceptualization, essentially, of the media multiplier. What is the share of sophisticated versus non sophisticated agents in that economy, which we calibrate based on a grid search?

So we're trying to kind of optimally fit essentially the parameter distribution. We do an exercise to just show that they signal. The statistical processing of the underlying data induces some signal and non linearities, which are not sort of, there is the statistical processing of the data, refines the data, the raw data, in a way that preserves a signal.

I just want to finish off based on sort of some calibrations of what happens to a new shock. And this is where, the point that you made, Steve met us. The media multiplier, the media multiplier's size is a function of the numerator as well as the denominator. So if you are Tunisia and your country is hardly ever in the news, one bad news induces, of course, you to go bang bang on the reporting.

But because of the inertia and the lack of data of other reporting that allows consumers to update their belief, it results in a degree of stickiness that makes the whole envelope of the recovery out of that new shock much more costly. And so the other news arrival matters, which of course is something that we could use very much to understand lots and lots of phenomena that are happening around the world that I've been somewhat working on.

So I'll just leave out here, we can calibrate this. Also, we do some speculation out of sample validation around the size of the media multiplier across different contexts using different corpora. But I'll just leave it here. Thank you so much for listening to me.

>> Steven Davis: Great, thank you, Timo.

That was really interesting. I wanna ask you, I'm struggling to just interpret exactly what the multiplier means conceptually, and maybe you can help me here. One thing I'm struggling with is the ground level truth that the sophisticated consumers know in your setting. As I understand that that's something that's constructed from multiple sources, and I think largely after the fact.

So I'm not sure how anyone would know that in real time. So that's the thing that's being multiplied upon. But your thought experiment is that there's some way you know these things in real time that would inform the tourism spending decisions and travel decisions of the sophisticated agents.

The other agents are just responding to what they see in the news. But you can see my struggle here. I don't really know what the base that's being multiplied on, is it's really a conceptual object. As I understand it, it's an as if object, not a real time thing you could look at.

Let me just make an analogy. If we were doing infectious diseases and the country-specific risks of infectious diseases, then a sophisticated traveler may just go to the World Health Organization site and see what the latest advisories are. So you'd have a very clear concept of what the ground level truth was in real time.

That seems missing in this context, or have I misunderstood?

>> Thiemo Fetzer: So I think, again, we just posit that there's these two types of agents, right? And you can very much think of the sophisticated agent as the one that actually looks at life. However, the travel advice is obviously incredibly, potentially incredibly sticky.

So travel advice hardly ever gets updated or gets updated with significant lag, even though there might be no material risk, right? Because the decision to adjust travel advice is also something that can be politicized and is actually used in bilateral kind of debates, right?

>> Steven Davis: But your version of ground level truth in this application is not travel advisories, as I understand it.

It's some-

>> Thiemo Fetzer: It's the history of violence.

>> Steven Davis: It's the history of violence which constructed after the fact. And I'm guessing that, often, constructed from news sources, but I don't know that.

>> Thiemo Fetzer: Yes, exactly, it's constructed from news sources. But again, we look at this, that these news sources are not country specific.

So whether a tourist got killed in Tunisia is not specific to each of the countries that might be sending tourists there. So it's sort of like it's an objective risk of becoming the victim of violence, irrespective of nationality, so to say, I mean.

>> Steven Davis: Got it, got it, that's a good way to put it.

But again, correct me if wrong, it's an objective measure of risk, but not a measure that was available in real time. It's one that's been constructed after the fact. That's what I stand.

>> Thiemo Fetzer: That is true. That is true. But that is sort of like any. We can make, of course, an out of sample prediction based on that, what the probability is that a place would be dangerous at any given point in time.

Because there's persistence in the Marco switching model, right? It's not like the places go bang bang all the time because there's baked and persistence. And so at some level, we think of this as a latent indicator of risk that I think has some signal. We tried to ground truth this with country travel advice data, but getting country travel advice data dynamically over a long period of time.

Yeah, we didn't wanna go down that road after trying to websites.

>> Steven Davis: I understand that. It's just one last comment. The conceptual model that you've set forth that you use to estimate the media multiplier, it just strikes me as more apt for the information environment that surrounds infectious diseases than violence.

For the reason that there is a fairly reliable, as I understand, and maybe I'm naive, a fairly reliable source you can go to to get travel advisors with respect to infectious diseases in near real time.

>> Thiemo Fetzer: Exactly, no, I'm totally with you. And it's also one example that we highlight in the introduction that it mimics very much that setting.

There was just no infectious disease environment that we could exploit over that time period in these five destinations that would give us. Because obviously, there's a causal identification part as well, right? That we need to address.

>> Steven Davis: Okay, thank you. Let Justin jump in here. I suspect he's got some comments.

 

>> Justin: Yeah, so this is great. And I had a lot of fun thinking about this project. I think there's a lot of facets to it that are deeply fascinating. I have sort of like two broad sets of questions. One is sort of about the, it sort of builds on what Steve was talking about, the information environment available to travelers.

And so part of that is the news. Part of that I was trying to puzzle through a little bit is the travel advisories or travel restrictions that sometimes one country will place on traveling to another that may not be based on a sort of sophisticated model. That also could be based on reactions to news coverage or sort of latent concerns that a country has about sort of risks.

But also that information environment could include positive statements from a country that's sort of advertising itself as safe, or you could imagine safe in particular ways, despite the certain kinds of things you may have heard. And that can interact, I think, in some pretty interesting ways with when we start thinking about the consumer model, even if we just think that they're making the sort of dyadic decision.

Do I go to Tunisia or not? Lots of other things are going on other than the news coverage.

>> Thiemo Fetzer: Yes, it's one of the exercises that we've now had to put into the appendix because you have to fill it to 45 pages. But we looked at press freedom as one of the features, the extent to which a country has a free press when reporting on violence.

And you see that there's interesting heterogeneity. Let's say Russian tourists traveling to Egypt do not respond in the way that European tourists would travel, because, well, reporting on violence in Egypt is suppressed in Russia. And there might be geopolitical alignment or incentives to keep that out of the news in the same way.

And so there is some interesting heterogeneity that is playing around with the consumer side, but also with obviously the producer side, right? As in, the tourist destinations might have incentives to suppress a certain negative news coverage, the same way this mechanism or the model here. Again, I look at this very much from the perspective of projecting soft power or constructing a soft power index, which is something that I want to train my ERC if I ever get to it.

Because obviously there is two things going on. Any type of news coverage might help you see that this is part of countries sovereign strategies that often gets labeled as greenwashing. Or potentially sports washing and whatnot, to improve the external image that the countries have, which might be not legitimate.

This is not for me to judge, also bringing in sports influences into their respective economies. The effect is not on the own economy. The effect is on the origin, where the respective David Beckhams are for coming from. That's the audience that are being played when we think about this phenomena.

And so again, as Steve, as you highlighted, this very much applies to the context, to a pandemic, where we have this sort of objective source is much clearer, right? But there's lots of other factors, again, that in the broader context, the probability of dying in a car accident in Egypt is much higher than the probability of dying from a violent event.

And is this tendency, of course, of the media to potentially skew information in a way that selects and attracts people's attention, that can be both weaponized. But also is something that can lead to over and under-reaction, producing volatility, that might adversely affect development outcomes. And I would very much like to say that this also has, I think, relevance for foreign direct investment decisions, or in general investment decisions, especially of retail investors.

Which in the context of attacking the climate crisis is particularly important because we need to facilitate capital flows from rich countries to poor countries. And that involves mobilizing retail investors. And of course, if there is such negative news bias and a history of bad news, I mean, I always gave the example of, I was born in 86, so I grew up with images of starving east African children.

That just sticks, that type of is a cohort, almost a generational cohort effect. If somebody says Ethiopia, to me, that's the first image that comes into my mind when in fact it might be very devoid from the economic reality at the time. But it might deprive that region of economic activity or capital flows that could help it develop its domestic economy.

 

>> Justin: Okay, I have a narrow question and then a bigger question. So I'll ask the narrow one first and then I'll get to the bigger one. The narrow one is violence is sort of a stand in, I think, for unpleasant experience while abroad. And the next sort of proximate unpleasant experience I was trying to think about was bad interaction with the government in the other place.

And the most salient example of that is you occasionally hear these stories of someone from Australia, goes to Thailand and they brought in drugs or something like that, and then they end up in prison for 30 years or something like that. Any sense of how prevalent those sorts of stories are?

Do they work in similar ways, different ways? Do people react to those?

>> Thiemo Fetzer: I would love to study it. We only had four countries, and I think they are the risk of, I mean, it's highly heterogeneous, but I would love to study that.

>> Justin: Yeah, okay, so the bigger question.

So again, thinking about the decision of the consumer. So I'm not only making an inference about the safety of a country, I'm thinking about going to, I'm assessing the relative safety of all the countries. And if I'm making a decision well, I wanna get out of my cold country or cold place that I am and I want to go to someplace warm.

And perhaps you have a preference to go to the perceived safest place. Well, you could imagine that if your sort of structural model is correct, those things are moving around quite a bit. And so I didn't think this is in the structural model, but maybe it is. I infer that a country is becoming less safe.

Does that increase my probability of traveling to a different country? Is there sort of like a reservation, well, I just view the world as being too dangerous to leave and I'll just tolerate the cold for a little bit.

>> Thiemo Fetzer: Yeah, I would again, would love to study this because I think this is super relevant, especially nowadays.

Think about the general equilibrium effects, both sort of the trade risk, so to say, there's a exchange rate and relative risk, and there might be an absolute risk, right? Or absolute risk perception. I think in order to study this, we would need to have broader data, which, however, has not been, well, it's been just difficult to access because again, some of this data.

As I described at the beginning, this data was being sold commercially, high prices, making it basically unaffordable, which raises bigger questions around building knowledge, public goods and limitations there too. But you're spot on. I think both margins matter, the extensive margin as well as the relative margin. But I think most consumers probably do, in this case, sequential decision making.

But that's something that, again, people, it would be great to test some of these mechanisms with microdata. You see some of the work from Leo Burstein and so on, essentially doing exercises like this with experiments, which of course is great. But it's very nice to see this actually bearing out, trying to map that to data that is, let's say, stuff that moves GDP, so to say.

 

>> Steven Davis: So I want to ask another question on the more technical side. You've got, in both your reduced form exercises and in your model based estimates of the media multiplier, you've got extensive sets of controls for fixed effects. You've got, as I understand, you've got dyadic fixed effects.

You've got origin by time fixed effects, destination by time fixed effects. But I don't see how you adjusted maybe I just missed it, how you adjust for the fact that the baseline level of news coverage is very different across your destination countries, even for a given origin. So there's gonna be more coverage of Israel and Egypt than Tunisia.

So when there is an event in Tunisia that involves coverage of violence or crime against tourists, that's gonna move things a lot, the proportion measure. Whereas if it was the same violent effect, the same implied increase in violence, the baseline level of news coverage about Israel is so much higher, I'm guessing.

That it would only move your truth measure, excuse me, your news coverage based measure, a small amount. So that suggests there's not gonna be anything approaching homogeneous responses to the violent events.

>> Thiemo Fetzer: Yeah.

>> Steven Davis: Across countries, especially across the destinations. Is there some way your model is dealing with that, or are you just imposing homogeneity?

 

>> Thiemo Fetzer: So, I guess on the reduced form, I mean, we obviously check whether the results are driven by any one destination. And that is not the case. The model, essentially, because the Markov switching model, the filtering essentially ensures that basically what moves the needle in Israel is different than what moves the needle in Tunisia.

And so I think that that addresses this heterogeneity that you have in mind here. So on the reduced form side, as I said, we can show that results are not driven by any one destination. And through the modeling of the parameter space in a way that's country-specific. And we address, again, this concern in a way that allows us to model this with a uniform preprocessing.

 

>> Steven Davis: Okay, but then the media multipliers that you estimate, the number three, is that an average across all dyad pairs?

>> Thiemo Fetzer: The interesting heterogeneity is at the tourist origin level, yes.

>> Steven Davis: At the origin, okay. And so it's not-

>> Thiemo Fetzer: Which is exactly the example, what I highlighted.

Russia versus Germany. If there's a violent event that is targeting a tourist in Egypt, the media coverage that this gets in Russia is different media coverage that this gets in Germany.

>> Steven Davis: Yeah, but I'm still struggling. There's two things going on. There's the extent of coverage in Russia versus the UK about violence that happens in Egypt.

I was actually asking about something different, which is the baseline level of news coverage about Egypt and Israel will be much greater than about Tunisia. How are you capturing, certainly the reduced form analysis you showed us quickly, it seemed like you just imposed homogeneity of the slope the response coefficients.

 

>> Thiemo Fetzer: Yes, yeah. Again, on the reduced form exercise, you can just drop each country and get similar coefficients. So that's not relevant heterogeneity. I think, in the context of the model, I think what we do is we have a destination specific arrival rate of news. And so that takes into account the fact that places can be different size in terms of some places just have a higher arrival rate or a lower arrival rate.

 

>> Steven Davis: I see.

>> Thiemo Fetzer: I think that tackles this in terms of the preprocessing. So it's a combination. I mean, my answer would be, again, through the reduced form, it doesn't seem to matter that much. And again, because we allow for heterogeneity in two ways, for the arrival rate of news and through the parameters that govern the Markov chain, that these are destination specific.

Because we only have Morocco, basically nothing is happening, which is interesting in itself. But obviously, because we only had four destinations, data only for four destinations, it was very difficult to explore this further. So what we did was explore the heterogeneity in the tourist sending countries, so to say.

And this is where the observation that press freedom matters is, I think, quite insightful and quite revealing, which, I mean, in terms of the broader discussion of all of this work. I very much tied this to, for example, the work that I did on the trade war, because there is a question to what extent countries that are not subject to, well, that do not have a free press are able to interact and engage, let's say, in geopolitical domains, very different.

Because they're subject to different domestic constraints, which is also something that the media effectively imposes, a constraint, which is also something that comes through in the paper that I did on the coalition casualties in Afghanistan. What is just a strong effect that we're losing a soldier has a multiplier effect.

And, of course, that opens the door for malign actors to, you know, offer bounties, which is, I think, exactly what happened.

>> Steven Davis: Okay, so we got some questions in the QA. First one's very easy to address. I'll just take it first. Will this paper become available after the presentation?

Well, it's already available, I think, on the website for this workshop. There's a link to the paper there. But maybe you want to just tell us you've got a website. Where can people find the paper if they want to find it?

>> Thiemo Fetzer: Yes, you can find it on my website.

So trfetzer.com, but you can also find it pretty much, I mean, just title. If you look for the title of the paper, the accepted version of the paper, which is not the cut down version that we have to now submit, is available open access. University of Warwick has an open access repository for the accepted version of papers.

So, yeah, you should be able to find it.

>> Steven Davis: Okay, great. And a second question is, it says, does this multiplier idea work, even if it comes in the form of good news? So can positive news about a country generate a multiplier effect if it's on tourism spending?

 

>> Thiemo Fetzer: My hunch is it works for the denominator, not the numerator, just through a drowning out effect. So that's ultimately, I think, the main mechanism. What we can't and what we did not discuss in this paper, which I think is really important, is the new selection function, because there is reporting and then there's reporting that gets put in the evening news.

So to say, if you think traditionally of a traditional news diet. And I think the selection function is actually really interesting. There's a nice paper, I think, that looks at the new selection function across sectors. I forgot. I think it's an AR paper that looks at heterogeneity in sort of sector level business cycles and the new selection of sector relevant news that tries to quantify as the relevance of the new selection function.

So I think it does apply to positive news, but it operates through essentially the drowning out of potentially bad news. And of course, there's a dynamic component to it, the build up of a reputation, which is we can also consider this to be a country's brand or soft power.

So one application I still want to do at some point is to bring this to bear on Brexit in the UK, because I have a bit of an issue with that.

>> Steven Davis: Okay.

>> Thiemo Fetzer: With Brexit.

>> Steven Davis: Okay, you can elaborate on that if you want, but why don't we give the, Hans Loiters has a longer question.

Why don't we let Hans articulate his question? Can we let Hans have the mic for a second?

>> Hans: Awesome. Yeah. Thank you. Can you hear me?

>> Steven Davis: Yes, we can. Thank you.

>> Hans: Awesome. Well, first of all, thank you, this was a really interesting talk. I really enjoyed your presentation.

So, I'm trying to understand kind of how to interpret this finding from the reduced form where you show that there's a decline in spending after this negative coverage. And I think there's two mechanisms. One is suggested by this, second finding that you have that kind of, the number of active credit cards goes down, which is just that fewer people are visiting the place.

But I'm wondering if there's a complementary second mechanism whereby the behavior of visitors in the country changes as well. So for example, you have tourists that might just decide to spend more time in their hotels rather than going outside and spending money doing souvenir shopping and so forth.

And I wonder if you have, first of all, I'm not sure how plausible that is, so I would love to hear your thoughts on that. And second, do you have data on the total number of transactions to maybe get into the second mechanism as well?

>> Thiemo Fetzer: Yeah, great question.

It's funny, I remember when I started working on this in 2014, I built this sort of booking.com scraper. That was going via the wayback machine to extract, essentially hotel price data, to look at some of these questions. Ultimately, the data is to allow us to look at substitution effects or behavioral changes within the country to distinguish.

We cannot reject the null hypothesis that the difference in the average transaction is different from the effect sizes that we document on the intensive and extensive margin. So to me, these are all really relevant questions, but with the data that we have, we cannot study them. Of course, I would think that this mechanism that we have in mind here applies very much to within country variation as well.

Think about behavioral changes around, another good example comes to mind. But any sort of news shock that might have heterogeneous impacts on media coverage, that interacts with some people's individual news filters. Or just a news diet, where they observe or where they obtain their information from, can produce heterogeneous responses.

And I think this is where I think this very much generalizes beyond the specific case. It's just that for the of course, having an environment where we can segregate the information sphere a little bit better because there's mechanic dividing lines. Simply language or distance and whatnot, can give us bite on this dyadic muse measure, which I think makes this quite unique from an identification perspective.

 

>> Steven Davis: Great, thanks so much for this, Timo, I'll give you a chance to make a few last words, but it's very interesting work. And I've struck, listening to you talk, by how complex the information environment is, both for consumers trying to update their beliefs about where things might be dangerous or safe.

There's no reason to expect, that I can see, to think that the actual dynamics of the evolution from safe to dangerous is the same across all these countries. So the Markov chain may differ greatly across countries. And of course, that's all being filtered through this very imperfect reporting that is the source of information for many or most of the consumers.

So the inference problem facing consumers is extremely complex and challenging. And then you're trying to, on top of that, infer what they are inferring and how that's affecting their behavior. So it's really quite a complex problem when you think about it that way. And as we've discussed, there's many other potential applications of your basic approach.

And again, the health advisory one, the infectious disease advisory, seems to me like the most natural application of your framework and your methods. I don't know if you plan to do that or.

>> Thiemo Fetzer: I think it would be great if somebody-

>> Steven Davis: You want somebody else to do it.

 

>> Thiemo Fetzer: Yeah, I think that would be great. I mean, to me, I think some very relevant questions arise around just geopolitics of stories. Because ultimately, when we're moving into a more fragmented global order and a lot of things are happening that are very difficult to comprehend. But ultimately, within this framework, we can think of stories as being weapons of war in the service sector, trade escalation.

And just like the decisions for countries to remove travel warnings are somethings that country leaders, that Turkish president talks to about when he visits the German chancellor, for example. So these are materially relevant from a geopolitical standpoint. And this huge heterogeneity, as I said, this type of framework of thinking is just very interesting to measure.

For example, soft power, other countries representation in each other's news. And here's just a great example that I just put together for my year grant application. This is three countries, Argentina, China and Germany, what is the share of the top five news sources that are represented? So top five foreign countries, and that's done with Imperfect Factiva data.

But we see that, for example, in Argentina, not surprisingly, Brazil takes up a much bigger chunk of its reporting because of geographic proximity. But then there's Spain, and there's obviously a language dimension that matters. If I think of the spread of, let's say artists, musicians, and so on, where they have market potential, where they can tap in, language is one of the factors that matters, right?

And so we see that there's a lot of latent dimensions that I think are quite relevant to explain the topography of a news coverage or media coverage, how other countries are representing each other's news. And of course, as technology has evolved, we're moving to a more global information sphere.

And the media consumption, how we consume news, is oftentimes generational, cohort-specific. This can create economic phenomena as well as political phenomena that I think worthy of studying this a lot more, yeah.

>> Steven Davis: Okay, great, we should sign off from the recording, but we're gonna continue the conversation for anyone who wants to stick around.

So, thank you, Timo, and thanks to the audience, this was a lot of fun, and it's been a pleasure to think about your work. And we'll see everybody next month, bye-bye.

>> Thiemo Fetzer: Thank you so much, bye-bye.