Terrorism and International Air Travel: A Gravity Approach *

Transcription

1 Terrorism and International Air Travel: A Gravity Approach * Devashish Mitra Ω Cong S. Pham Subhayu Bandyopadhyay ɸ May 2, 2018 Abstract. We present a theoretical model (adapted from the structural gravity model by Anderson and van Wincoop, 2003) to capture the effects of terrorism on air passenger traffic between nations affected by terrorism. We then use equations derived from this model, in conjunction with alternative functional forms for trade costs, to estimate the effects of terrorism on bilateral air passenger flows from 58 source countries to 26 destination countries during An additional small-scale terrorist incident in the origin country and destination country together results in a reduction in bilateral air passenger travel by, at least, 1.3% and 0.81% respectively for pairs of countries located 1000 km and 2000 km or less apart. The adverse impact of transnational terrorism is approximately five times larger. Terrorism adversely impacts bilateral air passenger travel both by reducing national output and especially by increasing psychological distress. Last but not the least, international air passenger travel is found to be extremely sensitive to fatal terrorist attacks and terrorist attacks on targets such as airports, travel or tourists. Keywords: air passengers, airline industry, gravity equation, international trade, terrorism JEL: F1, F14, L93 * The authors would like to thank two anonymous referees for their very valuable comments on a previous version of the paper. Thanks are also due to Javed Younas for providing us with the data used in this paper. Our paper has also benefitted from comments and suggestions by seminar participants at Deakin University s Trade and Development Workshop and the 13 th Asia Pacific Trade Seminars at the Foreign Trade University in Hanoi, Vietnam. This study was initiated when Cong S. Pham spent his sabbatical leave visiting Syracuse University between mid-august 2015 and mid-september He would like to thank the Department of Economics for its hospitality. Ω Department of Economics, Maxwell School of Citizenship and Public Affairs, Syracuse University. Tel: (315) dmitra@maxwell.syr.edu. Department of Economics, Deakin Business School, Deakin University, Victoria, Australia. Tel: cpham@deakin.edu.au. ɸ Federal Reserve Bank of St. Louis, USA. Tel: (314) Subhayu.Bandyopadhyay@stls.frb.org. Disclaimer: The views expressed are those of the authors and do not necessarily represent official positions of the Federal Reserve Bank of St. Louis or the Federal Reserve System. I. Introduction 1

2 A central factor distinguishing terrorism from other major types of violence or intimidation is that it is usually focused by the perpetrators on specific high value targets. An attack that is likely to involve targets of several nationalities is especially attractive for the terrorists because of the international press attention it is likely to garner. 1 Such attention is exactly what the terrorist organizations seek because of a variety of reasons, including raising security concerns in multiple nations, and motivating potential future recruits in different nations who may be sympathetic to the terrorist organization s beliefs. A prime sector offering such a target is international air passenger traffic, which by its very nature, is likely to involve citizens of several nations. Accordingly, an analysis of the effects of terrorism on the international air passenger is important due to its centrality in terrorists targeting. Terrorism adversely impacts international passenger flows for several reasons. First, terrorism generates a strong psychological negative effect on travellers because airlines have been frequent and the deadliest targets of terrorism for a long time. Targeting of popular tourist destinations like Istanbul also creates a strong deterrent for potential travellers. These effects are magnified by a number of factors. First, recent terrorist attacks have become more deadly because terrorists target densely populated and popular locations in major cities to maximize the impact of their attacks. Second, media coverage of terrorist incidents on a regular and repetitive basis reinforces the psychological trauma that terrorism seeks to create. 2 Finally, terrorism may also reduce international travel via general-equilibrium impacts on an economy. For example, terrorism can reduce income levels through macroeconomic disruptions 1 Enders and Sandler (2012) define terrorism as the premeditated use of or threat to use violence by individuals or subnational groups to obtain a political or social objective through the intimidation of a large audience beyond that of the immediate victims. This suggests that advertisement of attacks to a wide audience is of utmost importance to terrorist organizations. 2 It has been shown that people, who spent more time watching the coverage of the September 11, 2001 attacks, had more substantial stress reactions than those who watched less of that coverage (Schuster et al. (2001)). While terrorist attacks are found to have mental health effects on communities geographically distant from the attacks (Whalley and Brewin (2007)), it is also well documented that the negative psychological effects are likely to be decreasing with the distance from the location of the attacks. For example, Schuster et al. (2001) find such a correlation in the case of the 9/11 attacks. 2

3 (discussed in more detail in section 2 of this paper), which, in turn, are likely to reduce international travel (Czinkota, Knight, and Liesch, 2010). We first present a model where terrorism is conceptualized as adding to trade costs. The analysis yields equations (based on various alternative functional forms for trade costs) which we use to investigate the effects of terrorism on international air passenger flows using a sample of 58 source (origin) countries and 26 destination countries for the period of 2000 to Specifically, we estimate the combined effects of terrorism occurring in the pair of source country and destination country or, alternatively, the separate effects of terrorism occurring in the source and the destination country on their bilateral air passenger flows. To be clear, we empirically evaluate the effects of terrorism on bilateral flows of air passengers between countries where the terrorist attacks occurred. These countries may be or may not be the countries of residence of the terrorist organizations or the countries of residence of most of the victims (Mirza and Verdier (2008)). We believe that our paper contributes to the literature in a few ways. Firstly, ours is the first study that estimates the effect of terrorism on international air travel using a large sample of 1222 country pairs over the recent period Importantly, it is during this recent period that the world has experienced rapid growth in terrorism. Other studies either look at the impact of the 9/11 attacks on the US airline industry or look at the impact of tourism flows within a small set of countries for earlier and shorter time spans. Our study is also the only one in this literature that addresses endogeneity issues through the inclusion of source-country-year and destination-country-year fixed effects (as well as military spending as an additional variable). Unlike papers in the existing related literature, we investigate potential channels via which terrorism impacts air passenger travel. And very importantly, we believe that ours is the only study of the impact of terrorism on any kind of passenger transport (air, rail, water etc), that is closely guided by a fully and carefully specified structural-gravity model, namely the 3

4 approach to trade in services, pioneered by Anderson, Milot and Yotov (2014), itself an extension of the path-breaking paper by Anderson and van Wincoop (2003). 3 In addition, our empirical investigation is important because air travel has been and will remain a vital industry in the global economy for decades to come. According to estimates of The Air Travel Action Group (ATAG) in 2015 the aviation industry travelled close to 3.57 billion passengers and supported nearly 63 million jobs worldwide, either directly or through related tourism (9.9 million people work directly in the aviation industry). The aviation industry directly generates about $664 billion worth of value added per year, which is comparable to being ranked as the 21 st country in the world in terms of GDP. 4 It is estimated that aviation industry s global economic impact (direct and indirect) in 2008 was USD 3,560 billion, equivalent to 7.5% of that year s world GDP (ATAG, 2008). Air travel has been and is expected to remain a fast and sustainably growing industry in years to come. According to the forecast by the International Civil Aviation Organization (ICAO) the annual growth in international air traffic for the period is about 5% globally. 5 The air travel industry s most important contribution, arguably, is its role in facilitating the growth of other industries. For example, air passenger travel is of critical importance for the growth of tourism, because more than 40% of international tourists travel by air (ATAG 2008). Air travel also has positive supply-side spillovers on investment, innovation and productivity. Travel links are an important factor that determines companies investment, while travel services, including air passenger travel services, help in penetrating distant markets. For example, air passenger travel facilitates face-to-face networking and collaboration between 3 As discussed in the literature review there exist some related studies that look into the effects of terrorism on the airline industry (Lee et al. (2005), Ito and Lee (2005) and IATA (2011)) or on tourism (Llorca-Vivero (2007)). Yet, these studies either focus on the impact of the 9/11 attacks using US airline industry data or on the impact of terrorism on tourists travelling not only by air but also by road, river and sea. 4 More details on the importance of the aviation industry can be found from Facts & Figures of The Air Transport Action Group (ATAG), which are available online at: and It is estimated that aviation indirectly currently helps sustain $2.4 trillion in economic activity and these figures will more than double in 20 years time (IATA (2015)). 5 See Gillen (2009) for more details. 4

5 businesses and companies located in different places around the world, which also enhances firm productivity. It is estimated that for the European Union as a whole, the impact of overall air travel usage over the decade raised the level of underlying productivity by 2.0% and is predicted to raise it by another 0.6% by 2025 (Cooper and Smith (2005)). 6 Finally, since human beings are directly involved in air passenger travel, this sector is most likely to attract terrorists. Accordingly, use of bilateral air passenger flows is most pertinent to the identification of the effects of terrorism on the global airline industry. 7 We find that terrorism has a particularly strong negative effect on bilateral air passenger travel especially over relatively short distances. Specifically, an additional terrorist incident (usually of small scale and nonfatal) in the source country and the destination country combined results in a reduction in bilateral air passenger travel by at least 1.3% or 0.84 million USD for pairs of source and destination countries with their bilateral distance of 1000km or less. The effect of an additional transnational terrorist incident is on average approximately five times larger. It is important to note that the above magnitudes describe the effect a terrorist incident in a country can have on air travel between this country and one other country at the specified range of distance. However, this incident will additionally affect air travel in all the other bilateral pairs this country forms with other countries through connecting by air, so the overall effect of this incident will be many folds higher. In addition, there will be effects on the rest of the economy through reduced business travel as well as tourism, which will be many further folds higher. Going back to a single bilateral pair (and, again, without considering linkage effects), if we consider bilateral travel between France and UK, our estimates mean 6 Air transport also generates negative environmental impacts such as low air quality, noise and congestion in the vicinity of airports. Yet, these negative effects of air transport are considered to be largely outweighed by its positive spillover effects. Cooper and Smith (2005) referred to these positive effects as the catalytic effects of air transport industry. 7 According to a study by IATA (2011) on the effects of the 9/11 attacks on global aviation, the industry suffered a 6% drop of $22 billion in its revenue between 2000 and While the 9/11 tragedy was unique and much more severe in terms of its consequences relative to most other terrorist incidents, these estimates clearly reveal the extent to which aviation in general and air passenger transport in particular can be adversely influenced by terrorism. 5

6 that a (usually) small-scale terrorist incident and a transnational terrorist incident cause passenger travel between the two countries to decrease by 12 million USD and 80 million USD, respectively. The nonlinear adverse effect we find is consistent with the intuition that terrorism is more likely to cause air passengers to shift to other modes of travel (for example, road, rail or water transport) for relatively close destinations. We also find robust evidence that terrorism in the destination country adversely impacts bilateral air passenger travel via their negative effects on national output, increased psychological distress, higher costs of bilateral trade and through counter-terrorism related costs. Furthermore, we show that the responsiveness of international air passenger travel to terrorism critically depends on the type of terror that hits a country. In particular, we find that bilateral air travel is extremely sensitive to fatal terrorist attacks and to terrorist attacks of targets such as airports, transportation or tourists. In the next section we provide a review of the related literature. Section III provides a theoretical framework for our empirical analysis. Section IV discusses the data. We discuss our econometric method and analyse the regression results in Section V. Section VI concludes. II. Related Existing Literature Studies that are closest to our paper are those that look into the effects of terrorism on tourism. Using data for Greece, Italy and Austria, Enders et al. (1992) find that terrorists deter tourism in these countries, and an incident in one nation acts to reduce tourism in neighbouring nations. Using a consumer-choice model Drakos and Kutan (2003) find that, for the period , the tourism industry in Greece and especially Turkey and Israel was sensitive to terrorism and that there were also significant regional contagion effects of terrorism. Arana and Leon (2007) relied on a discrete choice experimental approach and survey data to evaluate the impact of 9/11 attacks on the welfare of tourists. They find that the 9/11 attacks caused a significant 6

7 decrease in tourists utility for those deciding upon travel plans for a set of Mediterranean destinations and Canary Islands. They argue that this decrease in utility might be explained by the state of anxiety surrounding the tourism industry. Using the gravity model, Llorca-Vivero (2007) estimates the impact of terrorist activity on international tourist flows from G-7 countries to 134 destinations over the period and find evidence of a damaging impact on international tourism flows. Yaya (2008), who examines the effect of small-scale terrorism in Turkey for sustained periods of time, documents that the lag length of a small negative effect of terrorism is one year. Over a period of 9 years this effect caused a reduction of six million foreign tourists. In a recent study Neumayer and Plumper (2016) analyse spatial spillover effects in international tourism as a consequence of transnational terrorist attacks executed in Islamic countries on citizens from Western countries. They find that terrorist attacks on tourist destinations in a country reduce tourist flows to the targeted tourist destinations, and also reduce tourist flows from nations whose citizens have been attacked. In addition, tourism flows from other similar source countries to the same destination countries are negatively affected. Bassil et al. (2017) evaluates the impact of domestic and transnational terrorism on demand for tourism to three countries: Lebanon, Turkey and Israel. They found that terrorism in one country has a spillover effect on other countries. Another strand of literature which consists of studies such as Ito and Lee (2005) and Lee et al. (2005) estimates the effects of terrorism on the airline industry. These studies however focus on the impact of the 9/11 terrorist attacks on the US air passenger transport demand. They find that there was a significant reduction of US air passenger transport demand following the 9/11 tragedy. Our paper differs from the afore-mentioned studies on tourism in several ways. First, it is the first study that estimates the effect of terrorism on international air travel, using a large sample of 1222 country pairs over the recent period Importantly, it is during this 7

8 recent period that the world has experienced rapid growth in terrorism. As mentioned above, both Lee et al. (2005) and Ito and Lee (2005) focus on estimating the effects of one major set of attacks, namely the 9/11 terrorist attacks, that were extremely large scale with a high death tool and, consequently, are very different from the majority of smaller scale terrorist events in the Global Terrorism Database that are the focus of our study. Also, both Ito and Lee (2005) and Lee et al. (2005) look at air travel for a very restricted sample of US air passenger transport only. Llorca-Vivero (2007) estimates the impact of terrorism on tourism flows that occur not just via air but through the combination of air, roads, river and sea, and his study is restricted to G7 countries for the period Secondly, in terms of methodology, our study also puts emphasis on addressing the endogeneity issue in the relationship between terrorism and tourism or air passenger transport. Specifically, we address the following source of endogeneity: countries, which are important tourism destinations, have a greater incentive to increase their counter-terrorism effort. Note that this source of endogeneity, if not controlled for, can cause an upward bias in the estimated effect of terrorism. As we explain more in detail in the section on empirical methodology and analysis we rely on gravity specifications that better address sources of endogeneity, through the inclusion of both source-country-year fixed effects (FEs) and destination-year FEs (or through the inclusion of source-country-year FEs and destination-country military spending as an additional variable). Thirdly, and very importantly, in our study we not only estimate the impact of terrorism on air passenger travel but also investigate potential channels via which terrorism impacts air passenger travel. To the best of our knowledge, the existing literature has not studied the 8

9 channels via which terrorism impacts international trade, tourism or air passenger travel. The investigation of these channels is largely missing from the existing literature. 8 Finally, and very importantly, we believe that ours is the only study of the impact of terrorism on any kind of passenger transport (air, rail, water etc), that is closely guided by a fully and carefully specified structural-gravity model. Using the structural-gravity approach to trade in services, pioneered by Anderson, Milot and Yotov (2014), itself an extension of Anderson and van Wincoop (2003), we attempt to fill this gap, and complement the extant literature on the economics of terrorism. III. Theory and Methodology Theoretical Framework The theory is adapted from Anderson and van Wincoop (2003), Anderson and Yotov (2010), and Anderson, Milot and Yotov (2014) to the case of passenger air travel services, with trade cost depending on the incidence of terrorism. While the main goal of the paper is to empirically estimate the impact of terrorism on international air travel, the value added of this section is to guide our empirical work in a relatively structural fashion. We arrive at our theoretical predictions (that are empirically investigated in later sections) by performing comparative static exercises on international passenger air travel with respect to terrorist activities in the origin and destination countries. Denoting passenger travel by the superscript A (since all international passenger travel in this model is going to be by air, even though domestic travel will be a mix of air, rail and road), let XX AA iiii be the value of passenger travel from country i to country j, i.e., how much passengers spend on traveling from country i to country j. In the case of international passenger 8 Pham and Doucouliagos (2017) is an exception in this regard. The study looks not only into the spillover effects of terrorism on international trade but also the channels through which these effects operate. 9

10 travel, i.e., for i j, we assume that all of it is air travel. However, when i = j, which is domestic passenger travel, a large part is road and rail. While these assumptions affect our results, they are also very realistic. Let EE AA ii be the total expenditure by passengers on passenger travel from country i to all destinations of the world and YY AA jj be the total expenditure by all passengers in the world on travel to all destinations in country j. Note that both of these include internal travel within countries i and j respectively. In addition, let σ be the elasticity of substitution between traveling to any two destination countries. Then our CES sub-utility function for travel from and within country i is given by σσ σσ 1 UU AA ii = ββ AA jj 1 σσ σσ σσ 1 cc AA σσ jj iiii (3.1) where cc iiii AA is the consumption of travel service in the form of traveling from country i to country j. 9 We denote the price of the passenger travel service bringing people to a destination country (adjusting or controlling for, i.e., net of, additional costs related to distance, security AA and safety and other kinds of frictions, passed on to passengers) by pp jj, while we capture those additional unit costs related to distance, security and safety and other frictions by a factor tt AA iiii. Following Anderson and van Wincoop and Anderson and Yotov, we assume an iceberg form for these costs. Then from (3.1), based on the maximization of this sub-utility subject to given expenditure on passenger travel and its prices, we have XX iiii AA = ββ jj pp jj AA tt iiii AA /PP ii AA 1 σσ EE ii AA (3.2) 9 Alternatively, this function can be viewed as a production function for the aggregate output of an economy (or for the aggregate of a class of goods), with the arguments of the function being the various kinds of inputs, including travel for business to various destinations. Our analysis and the results coming out of it will not change with this modification. 10

11 where ββ jj s, as we see above, are share parameters in our utility function and PP ii AA = ββ jj pp AA jj tt AA iiii 1 σσ jj 1/(1 σσ) is the CES price index of passenger travel faced by passengers traveling from country j to all world destinations. As mentioned above, this includes passenger travel to other countries (all of which is assumed to be air travel) and passenger travel within the country (that, in addition to air travel, has sizeable portions of road and rail). Market clearance, followed by appropriate manipulations of expressions, gives us XX iiii AA = EE ii AA YY jj AA YY AA tt AA 1 σσ iiii PP AA ii Π jj AA (3.3) where PP ii AA and Π jj AA are the origin and destination multilateral resistance terms respectively, as explained in Anderson and van Wincoop. Multilateral resistance terms capture the impact of relative trade costs in gravity models of trade. It is important for our analysis to note that Anderson and van Wincoop show that the same gravity equation and multilateral resistance terms can be derived if the trade costs are AA non-pecuniary and such costs are directly reflected in the utility function where cc iiii terms in the utility function are replaced by cc iiii AA tt iiii AA. This is especially important in our specific context since many of the costs related to terrorism in the case of passenger travel can be psychic (or psychological). So far the model has been identical to the general framework used in Anderson and van Wincoop, Anderson and Yotov and Anderson, Milot and Yotov, except that we have a specific AA interpretation of tt iiii which can enter the utility function to reflect psychological costs. We now introduce some more aspects that are specifically related to the question we are studying. Let Τi and Τj be measures of the extent of terrorism in countries i and j respectively, where tt iiii AA Τ ii > 0 and tt AA iiii > 0. To simplify our analysis, let s assume that Τ jj 1 tt AA iiii tt AA = γγ iiii Τ ii > 0 for i ii 11

12 j and 1 tt AA iiii tt AA = θθ iiii Τ ii > 0 for i = j where θθ ii < γγ ii. The reason for this differential impact is that, ii unlike international travel, which is mainly by air, domestic travel is a combination of air, road and rail. It is realistic to expect both the pecuniary and non-pecuniary costs to rise with terrorist activities proportionally less for road and rail travel than air travel.. These assumptions ensure that, with σσ > 1, we have XX iiii AA Τ ii < 0 and XX AA iiii < 0 for given EE AA Τ ii, YY AA jj and YY AA, since unit costs jj of domestic travel will not rise in the same proportion, resulting in a substitution into it away from international travel. We can define the entire utility function as a function of the various sub-utility functions ρρ ρρ 1 (one of them being passenger travel) as follows VV ii = (BB KK ) 1 ρρ ρρ (UU KK ) ρρ 1 ρρ KK, ρρ > 1. In the case of passenger travel K = A. In this case, treating EE ii AA as endogenous, the impact AA ddxx iiii ddτ ii < 0 is even stronger (greater in magnitude) since people switch between travel services and other goods and services. Similarly, the value of passenger travel into and within a country decreases with terrorism in the country, as people all over the world move into other goods and services and decide to travel elsewhere. Testable Hypotheses and the Basic Regression Equation Based on the above comparative statics, we have the following hypothesis. Hypothesis 1: Holding other things constant, the volume of international air travel between two countries is negatively related to the extent of terrorism in the source as well as the destination country. 12

13 Given that one of the channels through which terrorism affects air travel is the destruction of output and, therefore, incomes, we have the following hypothesis. Hypothesis 2: Controlling for GDP reduces the magnitude of the destructive effect of source-country and destination-country terrorism on international air travel. Finally, since substitution possibilities play a key role in our analysis, it is easy to see that when distances are short, there might be a bigger percentage impact of source- and destination-country terrorism on air travel. The reason is that in such cases there are greater substitution possibilities into road, rail and, sometimes, even sea transport. In addition, developing country governments do not have the resources (or sometimes even the skill or the will) to take the requisite actions to respond appropriately to terrorist incidents. Thus we have the following hypothesis. Hypothesis 3: The negative impact of terrorism on international air travel is greater in magnitude when the distance between a pair of countries is small enough to allow for substitution possibilities into other modes of transport and when at least one country in the pair under consideration is a developing country. Taking natural logarithms on both sides of equation (3.3), we have llllxx iiii AA = llllee ii AA + (σσ 1)llllPP ii AA + llllyy jj AA + (σσ 1)llllΠ jj AA (σσ 1)lllltt iiii AA (3.4) Substituting origin and destination country fixed effects, OO ii = llllee ii AA + (σσ 1)llllPP ii AA and DD jj = llllyy jj AA + (σσ 1)llllΠ jj AA, we have llllxx iiii AA = OO ii + DD jj (σσ 1)lllltt iiii AA, (3.5) where we can let 13

14 lllltt iiii AA = bb 0 + bb 1 ln (DDDDDDDDDDDDDDDD) iiii +bb 2 (BBBBBBBBBBBB) iiii + bb 3 (CCCCCCCCCCCC) iiii +bb 4 (LLLLLLLLLLLLLLLL) iiii +bb 5 (CCCCCCCCCCCCCCCC) iiii + φφ Τ ii, Τ jj. Based on the equations above, using a panel dataset on bilateral passenger air travel expenditures for several country pairs and over a few years, we can run a regression with timevarying origin and destination country fixed effects, in which case mainly country-pair-specific variables can be used. We will have time subscripts as follows. llllxx iiiiii AA = OO iiii + DD jjjj (σσ 1)lllltt iiiiii AA Also in that case φφ(τ iiii, Τ jjjj ) cannot be additively separable in Τ iiii and Τ jjjj since additive separability will make them perfectly correlated with the fixed effects, OO iiii and DD jjjj representing the multilateral resistance terms. For our purpose we use the following function that is not additively separable: φφ Τ iiii, Τ jjjj = aa + bbln (Τ iiii + Τ jjjj ). We can also go somewhat astructural where in place of the time-varying country fixed effects we have time-invariant origin and destination country fixed effects and the GDPs of both countries that vary over time. In such a case, we can actually have φφ Τ iiii, Τ jjjj = aa + bb ln(τ iiii ) + ccln (Τ jjjj ) or alternatively, φφ Τ iiii, Τ jjjj = aa + bbτ iiii + ccτ jjjj. Regressions incorporating these functional forms will shed light on Hypothesis 1. We do several other things in our estimation along these lines to make sure our results are robust. We can also drop the GDP terms to see the effect of terrorism through the destruction of national output. This will allow us to investigate Hypothesis 2. Also, based on Hypothesis 3, we will add interactions with distance as well as an indicator for being a developing country. These additional regression equations are presented in later sections while we present our results. 14

15 IV. Data Data on bilateral air passenger travel between 58 source countries and 26 destination countries for the period are from the UN Service Trade Database. Those are the source and destination countries for which the data on air passenger travel and control variables of the gravity equation is available. Specifically, the data on bilateral air passenger travel correspond to code 211 in the 2010 Extended Balance of Payment Services Classification (EBOPS). Note that EBOPS code 211 of Air Passenger Travel belongs to the EBOPS services that are exchanged between residents and non-residents through Mode 1 (cross-border supply). 10 Note that not all 58 source countries and 26 destination countries report the data for the whole period and that a source country s passengers usually travel to some but not all destination countries. As a result, we have an unbalanced panel. In order to avoid measurement error we only consider bilateral air passenger flow with value of US dollars and more. We have a final sample size of 8,416 observations for our OLS regressions and second-stage regressions of our Heckman model, with 13,277 observations for its first stage. Note that these are the sample sizes when all the data on the dependent variable and explanatory variables are jointly available. Data on terrorism are constructed based on the Global Terrorism Database (GTD) of the University of Maryland. 11 The database includes systematic data on domestic as well as transnational and international terrorist incidents that have occurred during the period 1970 to For our purpose we use all non-state terrorist attacks that the GTD can classify without uncertainty as terrorist incidents in order construct the measure of terrorism as our main explanatory variable of interest. 12 This choice of terrorist incidents minimizes the overlap 10 For more details on the classifications of different categories of service trade see WTO (2010). 11 The link to the database is the following: 12 Note that only 272 incidents in the GTD database are associated with state actors, representing less than 0.2%.The use of all terrorist incidents from the GTD as our measure of terrorism yields very similar results. Later in our analysis we also use transnational terrorist attacks as our measure of terrorism. 15

16 between terrorism and other forms of crime and political violence. 13 It is important to note that approximately half of terrorist incidents can be considered to be of small scale and have no human casualties. Data for transnational terrorist incidents for the period 2000 to 2012 are provided by Javed Younas. Data on some of the standard gravity variables are available from CEPII s gravity dataset. 14 Data on whether the pair of trading partners has a common currency or a common free trade agreement are available from De Sousa s database. 15 Finally, data on GDP, GDP per capita and military spending of countries in the sample as percentage of their GDPs are available from the World Bank Development Indicator Database. Data on air crashes from 2000 to 2014 have been collected by the authors from the Aviation Safety Network (ASN) database of Flight Safety Foundation (FSF). 16 The ASN database contains detailed descriptions of some 14,000 incidents, hijacking and accidents since For our analytical purpose, we only consider terrorism-unrelated air crashes and use them as an additional control variable in our gravity models. Table 1 provides the summary statistics of all variables used in the paper. It shows that we have 8,146 observations for all our regressions except for those using transnational terrorism (denoted by a T superscript), where the number of observations is smaller. Table 2 provides the detailed distribution of observations over the 15 years of the sample span. Each year starting 2004 has about 7-9.5% of the observations. The first four years together have less than 10% of the observations. V. Effects of Terrorism on the International Air Passenger Travel 13 See the GTD s Codebook for details on how an incident can be classified as a terrorist one. 14 The dataset is available from the following link of CEPII, a French Center of Research on the World Economy: Based on a method similar to the method CEPII s researchers used we however need to compute bilateral distances between Serbia and Russia and other countries for which the data is not available from the CEPII s database. For this purpose we use the information on bilateral distance between cities in the world from the following link: 15 The link to de Sousa s gravity data on common currency and free trade agreements is the following: 16 The database can be accessed via the following link: 16

17 The gravity specification of bilateral air passenger flows For our purpose, we rely on the following modified gravity model of bilateral air passenger flows between source country i and destination country j: 17 LLLLLL AAAAAAAAAAAAAAAAAAAAAAAA iiiiii = αα 0 + αα 1 BBBBBBBBBBBB iiii + αα 2 LLLLLLLLLLLLLLLL iiii + αα 3 CCCCCCCCCCCC iiii +αα 4 LLLLLL(DDDDDDDDDDDDDDDD iiii ) + αα 5 FFFFFF iiiiii + αα 6 CCCCCCCCCCCCCCCC iiiiii + αα 7 LLLLLL AAAAAAAAAAAAAAheeee iiiiii ) + αα 8 LLLLLL(IIIIIIIIIIIIIIIIII iiiiii + αα jjjj + αα iiii + εε 1iiiiii (5.1) where our explanatory variable of interest is constructed as follows: LLLLLL IIIIIIIIIIIIIIIIII iiiiii = LLLLLL(IIIIIIIIIIIIIIIIII iiii + IIIIIIIIIIIIIIIIII jjjj + 1) (5.2) For gravity specification (5.1), we use the country-pair-specific terror variable as defined by (5.2), which is the logarithm of the number of terrorist attacks in the source country and in the destination country combined (plus 1 to make sure this variable is not undefined when there is no terrorist attack). Taking the logs makes sure that this country-pair-by-year variable is not colinear with origin and destination country year fixed effects. However, when we look at the impact of origin country or destination country measure of terror one at a time (by eliminating the respective fixed effects by year), then we do not need to take logs. 18 As for the dependent and other explanatory variables of gravity equations (5.1) they are defined as follows: AirPassengersijt: the value of bilateral air passenger travel from source country i to destination country j. 17 Empirically, the gravity model has been found to be successful in explaining the volume of bilateral trade in services in numerous studies. (Kimura & Lee, 2006 & Anderson et al., 2014). Specifically, Kimura and Lee (2006) rely on the gravity equation to investigate what determines bilateral goods and services trade between 10 OECD member countries and other economies for the years 1999 and 2000, while Anderson et al. (2014) estimate geographical barriers to trade in nine service categories for Canada's provinces from 1997 to An alternative gravity specification is to include the year dummies and both measures of terrorism in the source (origin) country and in the destination country. Egger and Gassebner (2015) use this specification to investigate trade flows in their paper. We later use a modified version of this alternative gravity specification in order to investigate the direct and indirect effect of terrorism on the international air passenger transport. 18 An important feature of the data on terrorist incidents from the Global Terrorism Database (GTD) is that 49.64% or of the total of terrorist attacks reported by the GTD have zero death toll. 17

18 Distancesij: the bilateral distance between the source country of the air passenger flows and the destination country of the air passenger flows. Borderij: the dummy on whether the source country and the destination country share a common border. Colonyij: the dummy on whether the source country and the destination country share a common colonial relationship. Languageij: the dummy on whether the source country and the destination country speak the same language. FTAijt: the dummy on whether the source country and the destination country are members of a free trade agreement in year t. Currencyijt: the dummy on whether the source country and the destination country are members of a common currency area in year t. Milspendingit and Milspendingjt: the military spending as percentage of GDP of the source country i and the destination country j. AirCrashesijt: the number of terrorism-unrelated air crashes in logarithms associated with flights from, to or within the source country and the destination country. Some discussion of gravity equation (5.1) is warranted. First, in the literature on the gravity model, αit and αjt account respectively for the multilateral resistances associated with source country i and destination country j (Feenstra (2004)). The inclusion of these terms reduces the omitted variable bias to the extent that destination-year specific factors or sourcecountry-year specific factors may correlate with the explanatory variables of interest, i.e. the terrorism in the source country and terrorism in the destination country. As will be explained later, αit and αjt also take care of some of the endogeneity problems. Second, since it is the paper s goal to identify the impact of terrorism we use variable AirCrashesijt to control for normal (terrorism-unrelated) plane crashes. Terrorism-unrelated air 18

19 crashes may be an important determinant of air travel. Increasing frequency of terrorismunrelated air crashes potentially may raise security concerns and cause travellers to downsize or abandon their travel plans. Finally, in gravity specification (5.1) the effect of terrorism is implicitly assumed to be linear in bilateral distance. This assumption may not hold in reality because air travel is more easily substitutable by other modes of travel (for example, road, rail or water transport) for relatively close destinations. 19 In other words, the impact of terrorism is likely to go down with distance. In order to account for this possibility we rely on a gravity specification that includes as an added explanatory variable an interaction between terrorism and bilateral distance. Specifically, we use the following gravity specification, which is more general and more flexible than gravity equation (5.1) and is also our preferred gravity specification: LLLLLL AAAAAAAAAAAAAAAAAAAAAAAA iiiiii = αα 0 + αα 1 BBBBBBBBBBBB iiii + αα 2 LLLLLLLLLLLLLLLL iiii + αα 3 CCCCCCCCCCCC iiii +αα 4 LLLLLL(DDDDDDDDDDDDDDDD iiii ) + αα 5 FFFFFF iiiiii + αα 6 CCCCCCCCCCCCCCCC iiiiii + αα 7 LLLLLL(CCCCCCCCheeee iiiiii ) + αα 8 LLLLLL(IIIIIIIIIIIIIIIIII iiiiii ) + αα 9 LLLLLL(IIIIIIIIIIIIIIIIII iiiiii ) LLLLLL(DDDDDDDDDDDDDDDD iiii ) + αα jjjj + αα iiii + εε 2iiiiii (5.3) Note that if air travel for close destinations are more easily substitutable by other modes of travel we expect αα 9, the coefficient estimate of the interaction between the measure of terrorism and bilateral distance to be positive. This, of course, should happen in the presence of negative estimate of αα 8, the coefficient of the measure of terrorism by itself. These signs of estimated coefficients are expected from Hypotheses 1 and 3. Flow and Stock Measures of Terrorism 19 The dummy variable on sharing a common border only partially controls for the fact that passengers are more likely to substitute air transport for other types of transport when the destinations are neighbouring countries. 19

20 For our purpose, we use both the flow measure and the stock measure of terrorist incidents taking place in the source country of and the destination country for air passenger flows. The flow measure corresponds to the total number of terrorist attacks or the number of terrorismrelated human casualties in year t. When making their travel plans air travellers are likely not only to take into account today s terrorism but also what has taken place in the past. However, as one goes back more and more into the past, the impact keep diminishing. In other words, the impact of terrorism on air travel depreciates over time. That past terrorist attacks matter has been shown by Yaya (2008) for Turkey. The stock measure is the accumulated terrorist incidents or the accumulated terrorism-related human casualties in year (t-5) to year (t-1). Specifically, we construct flow and stock measures of terrorism using the number of terrorist incidents (Incidents). We use the following formula to compute the stock measures of terrorism in the source country and in the destination country for the international air passenger flows, respectively: SSSSSSSSSS_IIIIIIIIIIIIIIIIII vv,(tt 1) = ii=5 ii=1 ( ii)iiiiiiiiiiiiiiiiii vv,(tt ii) where v denotes either the source country or the destination country. Thus, terrorist incidents of years (t-5) to year (t-1) are accumulated in the stock measure of terrorism with a discount rate of 20%, meaning that the impact of terrorism in a particular year fully dies over the next five years. It is reasonable to argue here that the accumulated past terrorist incidents up to year (t-1), taking place in the neighbors of a country, have effects on its bilateral travel that are likely to be psychological. The reason is that people usually have short memories and that the media broadcast more recent terrorist incidents than older ones, and hence it is expected that recent terrorist incidents have a larger adverse impact on bilateral air passenger flows. 20 As a 20 As a robustness check we later compute the stock measures of terrorism by accumulating terrorist incidents of the last ten years with a discount rate of 10%. The regression results remain essentially the same. 20

21 robustness check, we later accumulate the terrorist incidents of the past 10 years with a discount rate of 10%. Terrorism in the World from 2000 to 2014 Figure 1 presents the evolution of terrorism from 2000 to 2014 in two categories: the total number of terrorist incidents and the total number of people killed in the world. Note that for our purpose, the table reports the number of non-state terrorist incidents in the GTD. A noticeable pattern is that from the start of our sample, 2000, and especially from 2010, terrorism has become much more frequent. Specifically, the number of terrorist incidents has increased approximately by a factor of five from 2003 to The number of people killed in terrorism related incidents also has been increasing exponentially. Table 3 shows that in 2014 the world recorded more than terrorist attacks, which was by far the highest number of terrorist attacks for the period 1970 to In 2014 more than people were killed in terrorism related attacks, which is 3 times (or more) the global death toll in any year during the period Figure 2 shows that terrorism s average death toll per terrorist incident generally has become stable over time at 2 to 5 human casualties. The number of people killed per terrorist attack has declined very slightly recently. The analysis of the geographical distribution of terrorism also reveals that regions in which terrorism has been active the most are the Middle East, India, Pakistan as well as the United States and Russia. Effects of Terrorism on Bilateral Flows of Air Passengers: Analysis of the Regression Results Our regression results of the gravity specification (5.1) are presented in columns 1 to 4 of Table 4. The first two columns report results using the flow measures of terrorism while the last two columns present the results using the stock measures of terrorism. The results show that the gravity specification generally performs well. Factors such as sharing a common border and 21

22 having a common colonial relationship have a positive and statistically significant effect on bilateral passenger travel while bilateral distance significantly reduces it. The magnitude of the coefficient of distance on bilateral travel is very comparable to other studies that investigate the determinants of bilateral trade in services using the gravity model (Kimura and Lee (2006)). 21 Factors such as common language, FTA membership and common currency membership are not found to impact bilateral air travel. The coefficient estimates on those variables are not statistically significant. Importantly, terrorism-unrelated air crashes related to the source country and/or in the destination country are found to have no impact on air passenger travel. This result is however not surprising to the extent that terrorism-unrelated air crashes have found to be significantly decreasing over the time span of our sample (See Appendix A4). Columns 1 and 3, which present the results of gravity specification (5.1), show that terrorism adversely impacts bilateral passenger air travel and the effect is statistically significant at the 5% level when the stock measure of terrorism is used. Results in columns (2) and (4) which correspond to gravity specification (5.3), show that the effect of terrorism on bilateral air travel is not linear but decreases with bilateral distance. Comparing results from gravity specification (5.1) to those from gravity equation (5.3) provides strong evidence that the adverse impact of terrorism on air passenger travel interacts with bilateral distance and is especially significant for those travellers of short distances. Next, we run regressions of our preferred gravity specification (5.3) using a slightly different measure of terrorism. Specifically, we first use as our main explanatory variable the lagged value of our flow measure of terrorism in year (t-1) or year (t-2). It is highly plausible that terrorism may have a lagged effect on air passenger travel. The new set of results from 21 For example, in Kimura and Lee (2006) the coefficient estimate of bilateral distance ranges between to (See Table 3). 22

23 gravity specification (5.3), presented in Table 5, provides robust evidence of a strong adverse effect of terrorism on bilateral air passenger travel. In all regressions terrorism is found to significantly and adversely impact air passenger travel at 10% or 5% level. The negative impact of terrorism on air passenger travel is decreasing in magnitude with bilateral distance. This finding is in line with the argument that air passenger travel is more easily substitutable for closer destinations by inland modes of travel or rivers. So far our construction of the measure of terrorism makes no distinction between domestic and transnational terrorist incidents. Since data on air passenger travel used in our study relate to international travel, it is important to investigate the impact of transnational terrorism. It is our expectation that transnational terrorism is likely to more adversely impact bilateral flows of air passenger travel, and transnational terrorists are more likely to target air passenger travel from one country to another. For our purpose, we now use the number of transnational terrorist attacks constructed by Enders et al. (2011) as our measure of terrorism. 22 The new set of results is presented in Table 6. Note that the use of transnational terrorism as our main explanatory variable results in a reduction of the sample size. All the results show that both the flow measure and the stock measure of transnational terrorism have a strong adverse effect on air passenger travel and this adverse effect decreases in bilateral distance. Compared to results in Tables 3 and 4 where a more general measure of terrorism is used, transnational terrorism has a much larger negative impact on bilateral air passenger travel. Overall, the results show that terrorism adversely and significantly impacts bilateral air passenger travel. What is the economic significance of our estimates? According to the results of column 2 in Table 3 a 10% increase in the number of terrorist incidents in the source country and the destination country results in a reduction in bilateral air passenger travel at least 22 Gaibulloev, Piazza and Sandler (2017) and Younas and Sandler (2017) adopt the measure of transnational terrorism constructed by Enders et al. (2011). They also derive an alternative measure of transnational terrorism measure from the International Terrorism: Attributes of Terrorist Events (ITERATE) database. 23