In 2016, terrorism in the US caused fewer than 0.01% of the nation’s deaths but was covered by its newspapers more than any other cause.1 Cancer and heart disease caused more than 60% of all deaths but were represented in only 15% of news. This kind of sensationalist bias can influence perceptions of the causes of death when individuals use ‘availability heuristics’, i.e. when they judge the probability of a cause by the ease with which they can recall it (Tversky and Kahneman 1973). In psychology these kinds of biases have received a lot attention, and recent research reviewed by Zhu et al. (2020) suggests that they occur because we use Bayesian sampling intuitively in our minds. These kinds of biased beliefs can affect behaviour and they lower expected welfare if, for example, one avoids public spaces but continues smoking.
Biases of this sort might also affect the macroeconomy, trade, and even economic integration. This is most apparent for developing countries, which are rarely covered by international news unless something terrible happens. If it is not discounted by consumers, this sort of news bias can shape the international image of countries – or even entire regions – as dangerous and miserable, affecting travel to these places and isolating them economically. But quantifying these effects is close to impossible, as the effect of events and reporting on them cannot typically be studied separately.
Tourism offers a unique chance to study the effect of news, as tourists from different origins evaluate the same destination in order to decide on a holiday. If different origin countries report differently on the same events, then tourists from different countries may react differently to the same events. In a new paper, we use this separation to quantify the economic consequences of bad news (Besley et al. 2020).
To capture the news environment, we build a corpus of nearly half a million news articles from different origin countries. Our corpus consists of news articles from 57 origin countries covering five destinations – Egypt, Israel, Morocco, Tunisia, and Turkey – from 2009 to 2016. We auto-translate all articles and identify a subset of articles that report on fatal violence or violence against tourists. Given the size of the corpus, we use methods from computational linguistics and supervised machine learning to develop an automated classifier of reports on violence. This software automatically flags any article that reports on fatal violence or tourists being attacked. In this way we can compare, for example, how news is reported in South Korea and Germany on the same events occurring in Egypt, Israel, Morocco, Tunisia, and Turkey.
Figure 1 shows the aggregate reporting share of our definitions of ‘bad news’. The first measure is reports on fatalities, which we plot as blue lines. The other definition of bad news in the context of tourism is whether tourists were targeted. We report this measure as red dashed lines. In both cases, we report the share of bad news relative to all news. This relative measure increases when either the number of bad news reports increase or reporting on other news decreases. In other words, our relative measure not only allows us to capture the effect of bad news but also the absence of other news. Figure 1 shows that this is a particularly big problem for Tunisia – a country which is reported on very little in international news. Reporting on tourists being targeted captured more than 40% of all reporting on the country in the month of the 2015 Sousse attack. A country of 11 million inhabitants was, to a large degree, defined by a series of terror attacks in the eyes of an international readership. For Turkey, a country that experienced terrible terror attacks a year later, the relative news share of bad news never increased beyond 8%.
Figure 1 Aggregate reporting shares of ‘bad news’
In order to gauge the economic effects of reporting, we leverage aggregated and anonymised monthly card spending data proxying tourism activity of MasterCard holders from our 57 origin countries in our five destinations countries – a pairwise data structure very similar to data measuring international trade flows. This data structure allows us to study whether tourists react to events at their destination or to reporting at home. We can thereby learn whether the intensity of bad news is at least partly responsible for changes in card spending. Do countries that report most intensely on an event experience larger drops in spending? How large is this effect compared to the effect of the event itself?
We document a robust relationship between the intensity of reporting on violence and subsequent drops in tourism spending and visits. The effects we find are sizeable: we estimate that if reporting on a specific dyad switches from reporting on topics unrelated to violence to covering only stories about tourists being targeted, tourism spending drops by about 56% a month later. This effect is fairly persistent in the first few months but then dissipates, lasting for about nine months following the negative coverage.
News biases and availability heuristics: A damaging combination
Our results show that some destinations receive much more intense relative coverage of bad news, driven to a large degree by the absence of other news presented on those countries. South Korean news outlets, for example, rarely report on Northern Africa. When they do, it is typically to report that South Koreans have been kidnaped or that other tourists have been killed. This leaves audiences at home with the impression that Northern Africa is extremely dangerous. Tourists from Spain see more other news from their neighbours south of the Mediterranean and are therefore more impervious to bad news. Most importantly, the same is true for different destinations: Tunisia suffers by far the strongest effect of bad news items on spending because the country is reported on so little otherwise.
To fully estimate this relative news effect, we develop a model to separate the effect of the event from the effect of news. To capture the role of news in the model, we distinguish between two types of potential travellers. We call the first type ‘sophisticated’ tourists. These tourists are assumed to observe actual measures of violence and build their perception of the risk of danger based on these data. We build a model of the best possible risk evaluation that ‘sophisticated’ tourists would have made regarding any of the destination countries. This risk is not origin specific because it is driven by events at the destination. The other type of tourist is ‘naïve’ in the sense that they only respond to the information they observe from watching the news. The worldview of these tourists behaves as if they were using availability heuristics. Even with constant objective risks, the perceived risk for these tourists can change dramatically with news about violence.2 These theoretical types represent two extremes; actual card spending will be driven by a mix of both behaviours. However, by separating types of behaviours in this way, we can use a grid search to estimate what mixture best describes the aggregate data. We find that a weight of more than 50% for tourists using availability heuristics best describes the data.
The resulting model of tourist behaviour allows us to simulate a combination of a violent event at a destination with a news shock at the origin. The result is shown in Figure 2. We assume that a single, unrepeated event happens. This makes sophisticated tourists change their travel plans and, as tourists book travel in advance, this leads to an effect on tourists’ spending, which only slowly fades out. The resulting reaction is shown as a dashed line. Tourism falls by four percentage points and recovers slowly after nine months. Overall, an isolated event like this would make a country lose close to 15% of one month’s tourist revenues.
Figure 2 Simulated news effect
Figure 2 also shows the simulated news effect. If one news item covers the event, the potential loss increases considerably. Losses without other reporting drops by close to eight percentage points. This is the kind of undampened effect that many countries experienced as a reaction to events in Tunisia. The combination of media bias and tourists’ availability heuristic drove up their assessment of the risk of travel to the country, doubling the country’s loss from the original event. Figure 2 also shows the effect on tourists who receive a lot of other news on the destination country. When a hundred other news items are seen together with the bad news, the overall effect of the event is close to the one felt by sophisticated tourists.
For a country like Tunisia, which receives little international coverage, this leads to dramatic effects on aggregate spending. We use the estimated model to show, for example, that in 2015 the news effect alone was responsible for a spending decline of about 15%. How large was the overall loss for all five countries? The World Bank reports that tourism revenues in 2010 were $3.48 billion in Tunisia, $5.6 billion in Israel, $13.63 billion in Egypt, and $26.3 billion in Turkey. Based on the estimates in Besley et al. (2020), total losses between 2011 and 2016 could have exceeded $35 billion due to violence, with in excess of $10 billion due to negative news reporting. Our simulations also indicate that Egypt and Tunisia are predicted to have recovered their losses towards the end of the sample period.
Broader implications of these findings? How important is negative news as a phenomenon? How balanced is reporting across the world? Figure 3 summarises the impression that readers of BBC Monitoring, The Economist, The New York Times and The Washington Post will have from different regions of the World (Mueller and Rauh 2019). It plots the average number of words appearing on the respective region per quarter in all of these outlets on the horizontal axis and compares this with the share of news about conflict on the vertical axis.
Figure 3 Reporting across regions of the world – the vertical axis measures the share of news written on conflict vis-à-vis the average number of words per quarter on the horizontal axis
A clear pattern emerges. Whereas reporting per quarter on developed regions is very intense, reporting on Africa and Latin America is considerably lower. At the same time, there is much more reporting on conflict in these countries. There are more conflicts in these regions, but the particularly worrying pattern here is the low level of reporting in the absence of conflict. Even if violence is contained, single isolated violent events may still erode perceptions and trust in the relative safety of a country, undermining the economic development that is usually fostered by trade more broadly (e.g. Frankel and Romer 2008, Donaldson 2018).
Azeredo da Silveira, R and M Woodford (2019), "Noisy Memory and Over-Reaction to News", American Economic Association Papers and Proceedings 109: 557–561.
Besley, T, T Fetzer and H Mueller (2020), “Terror and Tourism: The Economic Consequences of Media Coverage”, CEPR Working Paper.
Bordalo, P, K Coffman, N Gennaioli, and A Shleifer (2016), "Stereotypes", The Quarterly Journal of Economics October: 1753–1794.
Combs, B and P Slovic (1979), “Newspaper Coverage of Causes of Death”, Journalism Quarterly 56(4): 837-849.
Donaldson, D (2018), “Railroads of the Raj: Estimating the Impact of Transportation Infrastructure”, American Economic Review 108(4-5): 899–934.
Frankel, J A and D Romer (2008), “Does Trade Cause Growth?”, American Economic Review 89(3): 379–399.
Shen, O, H Al-Jamaly, M Siemers and N Stone (2018), “Death: Reality vs Reported”, https://owenshen24.github.io/charting-death/
Tversky, A and D Kahneman (1973), “Availability: A heuristic for judging frequency and probability”, Cognitive Psychology 5(2): 207–232.
Zhu, J, A N Sanborn and N Chater (2020), "The Bayesian sampler : generic Bayesian inference causes incoherence in human probability", Psychological Review.
1 Following Combs and Slovic (1979), Owen Shen et al. (2018) compares causes of deaths in 2016 to news reporting on different causes of death by The New York Times and other news outlets. Our World in Data compiled this striking summary based on their efforts: https://ourworldindata.org/uploads/2019/05/Causes-of-death-in-USA-vs.-media-coverage.png.
2 Of course, in our aggregate data we cannot pin down the explanation for why we observe strong reactions to news stories. For alternative explanations, see Azeredo da Silveira and Woodford (2019) and Bordalo et al. (2018).