The negative impact of terrorist attacks on tourism demand has frequently been the focus of attention of the academic literature and the media. Just recently, the media reported that the terrorist attacks conducted by the Islamic State have harmed the tourism industry in such places as Turkey, Egypt, Paris or London. However, the scholarly literature on the impact of terrorism on tourism or the terrorism-tourism nexus, as I call it here, was mainly dominated by qualitative analysis of specific case studies and anectodal evidence.

This study, by contrast, is supposed to go far beyond the scope of most academic studies. In fact, the analysis will focus on the general effects of terrorist attacks on tourism by using quantitative methodology. More specifically, the main purpose of this study is to assess the causal effect of terrorist attacks on tourist arrivals, how strong this effect is, and whether or not terrorist attacks have a lagged effect on the tourism demand one year after an attack had occured.

In so doing, I will conduct a difference in differences analysis (DID) based on panel data between 1994 and 2014 combined with cross-sectional data of 150 countries. The results of my analysis will show that the causal impact of terrorist attacks on tourism demand is not as straight forward as one might believe. There are, of course, single cases, in which terrorism leads to a collapse of the tourism industry. In most cases, however, there is a complex set of variables that determine tourism demand, many of which diminish or mitigate the impact of terrorist attacks.

The analysis will start with a review of the scholarly literature on the relationship between terrorism and tourism. The discussion of the academic literature will be followed by a sound description of the theoretical framework, in which I will outline my research question and a set of hypothesis. In a next step, the study provides a detailed chapter on the data and the statistical model developed for the purposes of this study. After that, the study goes on with exploring the main findings of the empirical analysis. In the last chapter I will discuss the results within the context of terrorism and tourism research and will summarize a couple of future research questions.

### 1 Literature review

The conventional view on the linkages between terrorism and tourism is pretty straight forward. Terrorist attacks, as the argument goes, decreases tourism demand by changing the behavior of tourism consumers. From a theoretical perspective, this thesis is often explained by what risk perception theory predicts, arguing that terrorist attacks lead to decreasing tourism demand as they spread anxiety, affect tourists’ risk perception and determine their choice of destination (Baker 2014: 58-67).

In a qualitative analysis Walter and Sandler (1991: 49-58) found evidence that the terrorist attacks in Spain had a negative impact on the Spanish tourism industry. One of the most interesting findings of their study is that the terrorist attacks on Spanish soil turned out to have a lagged effect on tourism demand, which lasted up to nine months. The reason for this is, as they argue, that tourism agencies had already finalized travel contracts and tourists are barely willing to cancel their bookings. As already mentioned in the introduction, this hypothesis will be tested in the following analysis as well.

One of the few quantitative studies was conducted by Neumayer (2004: 259-281), who found statistical evidence for a negative impact of terrorism on the number of tourists. As he concluded, however, there a variety of other factors influencing tourism demand such as human rights abuses, political instability or the overall conflict intensity. Another quantitative study by Martin and Gu (1992: 3-15) investigates passenger numbers at Florida International Airport and concludes that airline hijackings conducted by terrorists have led to a change in the behavior of tourism consumers, in so far that destinations in the Middle East were substituted by destinations in North America.

In addition to this, a variety of researchers argue that the effects of terrorism on tourism go beyond national borders and even influence the tourism demand at a regional and global level (Arana et al 2008; Ahlfeldt et al 2014: 3-21). The 9/11 attacks, for instance, increased anxiety of tourists at a global scale and led to a widespread change in consumer preferences. A new study by Neumayer and Plümper (2016: 195-206) concluded that terrorist incidents in Muslim countries such as kidnapping do not only decrease the tourism demand in the affected country but also lead to a spill over effect to other Muslim majority countries. The reason why the impact of terrorism on tourism appears to be transnational, as they argue, lies in the transnational nature of Islamist terrorism.

Moreover, studies emphasize the multi-dimensional effects between terrorism and tourism, of which the former does not only negatively influence the latter but is also negatively affected by the loss of tourists. Hussein Gure Diriye (2015), for instance, argues that the terrorist attacks in Kenya led to higher unemployment and lower productivity, which can lead, in return, to more terrorism.

As described in this section, a variety of studies indicate a negative impact of terrorism on tourism demand. However, there is a number of reasons why the subject is still worth being studied. First, there might be a couple of reasons why the terrorism-tourism nexus is not as clear as one might expect. Most of the studies described above emphasize that terrorist attacks alone cannot sufficiently explain the general tourism demand since there are additional factors shaping the behavior of tourists.

The decreasing tourist numbers in Turkey following the terrorist attacks by ISIS, for instance, might have been also caused by a general political instability as the country had suffered from an attempted military coup, subsequent human rights abuses and a reflaming domestic conflict with the Kurds. Second, quantitative studies on the terrorism-tourism nexus are rare within the scholarly literature, which is problematic from a scientific point of view, in so far as the qualitative analysis of single cases cannot be generalized. And third, many studies with a quantitative focus were conducted long time ago, which means that new data could lead to new results.

### 2 Theoretical background

In this study I will analyze the causal impact of terrorist attacks and tourism demand in 156 countries by using a difference in differences regression model. Thus, the primary research question in this study can be formulated as: do terrorist attacks have a negative impact on the tourism demand and, if yes, to what extent? More specifically, I will test a couple of DID regression models with a variety of dependent and independent variables in order to estimate the causal impact of terrorism on tourism consumption. Given this research interest, the main hypothesis (H) will be:

H: Terrorist attacks have a causal impact on the behavior of tourism consumers and decrease the tourism demand in the affected country.

In order to test the main hypothesis, I will create a number of sub-hypothesis, which can be regarded as indicators of the overall tourism demand and which will allow me to draw conclusions about the main hypothesis and the primary research question as described above. The tourism demand as the dependent variable will be tested by three hypothesis with three different dependent variables . First, I will analyze the impact of terrorist attacks on the absolute number of tourist arrivals per year (H_1), which can be formulated as followed:

H_1: Terrorist attacks have a negative causal impact on the absolute numbers of tourists per year in the affected country.

The second dependent variable measuring tourism demand will be the the annual growth rate of the number of tourist arrivals per year (H_2), which can written as followed:

H_2: Terrorist attacks turn out to have a negative causal impact on the growth rate of the number of tourists compared to the year before.

These two sub-hypothesis will be in the center of the regression analysis. Each of the following hypothesis listed below will be, therefore, “subsub-hypothesis” of these two. As described in the literature review, some researchers argue that terrorist attacks have a lagged negative impact on a country’s tourism demand which can last up to one year after the attack. This lagged effect will be tested by an additional hypothesis for each of the sub-hypothesis described above and can be formulated as:

H_3: Terrorist attacks have a lagged negative effect on tourism demand and materialize one year after an attack.

In addition to these hypothesis, which test the impact of terrorist attacks on different dependent variables, I create a set of additional hypothesis to investigate a variety of independent variables which might co-affect the overall tourism demand. This is why my DID regression model will also include variables other than terrorist attacks such as GDP per capita, conflict intensity, the political institutional structure or the population size. Here, I will focus on the following hypothesis:

H_4: Tourism demand is also affected by the GDP per capita and the income level. The higher the GDP per capita in a country, the more tourists are expected to visit the country.

H_5: Tourism demand is influenced by the political institutional structure of a country. The more democratic the institutional structure of a state, the higher the tourism demand.

The main hypothesis (H) and its six sub-hypothesis (H_1 – H_2) will be analyzed by creating and testing a couple of difference in differences (DID) regression models, which will be specified in the next section. There will be three main parts in the analysis. The first part will test the impact of terrorist attacks on the absolute number of tourist arrivals per year (H_1), including the lagged effect and a whole bunch of other independent variables such as GDP per capita, the political institutional structure of a country, and the conflict intensity. The second part will test the other dependent variables in particular the annual growth rate of the tourist number (H_2) plus the lagged effect and the other independent variables. The third part of the analysis discusses the weaknesses of the model and provides alternative explanations and models.

### 3 Empirical background

The following study tries to contribute to the academic literature as it provides a standardized and generalized approach measuring the impact of terrorist attacks on tourism demand. The main part of my empirical design will be a mutiple DID regression analysis which is supposed to estimate the causal impact of terrorism on tourism and which includes several dependent and independent variables. Before a detailed description of the DID regression model will be pursued, however, the section introduces the dataset and its specific variables.

#### 3.1 Data

In order to conduct multiple DID regression models I had to discover different data sources which were merged into one large dataset consisting of panel data between 1994 and 2014 and cross-sectional variables of 156 countries. The final dataset, as can be seen in Table 1, includes in total twelve variables and more than 3,276 observations. The variables *tourist arrivals per year*, *GDP* and *population size* were collected from World Development Indicators, a dataset constructed by the World Bank. Based on these data I generated the additional dependent variable *annual tourist number growth rate* and the independent variable *GDP per capita*.

The original source of the panel data on terrorist attacks is provided by the National Consortium for the Study of Terrorism and Responses to Terrorism. Here, terrorist attacks are defined as armed attacks on civilians motivated by political, economic, social or religious goals. This dataset provides detailed information on the number of fatalities caused by attacks but also incorporates minor terrorist incidents and kidnappings. The dataset constructed for this study, however, will only include events which caused at least one individual dead and excludes minor events, kidnappings or criminal activities.

Moreover, I collected annual time series data from two datasets provided by the Centre of Systemic Peace, one on institutional structures in different countries, by which I created the interval scaled variable *polity* ranging from -10 (autocracy) to 10 (democracy), and the other one an interval scaled variable measuring the *conflict intensity* with 10 meaning a high intensity and 0 low intensity. The two variables are complex variables based on several qualitative indicators which were operationalized and quantified by the Centre of Systemic Peace and which cannot be explained in detail here. The polity scale, for instance, is calculated by such factors as the competitiveness and openness of the executive recruitment process, political and legal constraints on the government, and the competitiveness of political participation. The conflict intensity scale is a more technical one and is calculated by the number of fatalities, the duration of the conflict, the extent of the dislocation of the population, infrastructural damage, or resource depletion.

#### 3.2 The DID regression model

In the following section I will provide a detailed description of my statistical model, of which the first part will discuss a general discussion of the idea behind difference in differences (DID) models and of which the second part will specify how I am going to apply the model in the following analysis.

Although the origins of the concept of DID analysis are rooted in clinical studies, it has become a popular tool in econometric analysis because the model allows for assessing the causal impact of a certain policy implemented by a government or a business. Let me first explain the idea behind DID analysis by using clinical studies. Within the context of clinical studies a number of sick people gets randomly splitted into two groups: a treatment group, which receives real medical treatment, and a control group which gets a placebo. At the same time, the course of the desease and its progress is observed in the pre- and the post period of the treatment before the results (in most cases an average of a specific value which serves as an indicator for the desease such as the blood level) get compared with each other.

Now the control group serves as a reference group for the treatment group. If the treatment group is recovering but the control group is not, the doctors can conclude that the treatment has a causal impact on those who received the treatment. Let us assume, the desease does not improve in the treatment group and remains at the same level as in the control group. In this case, doctors can conclude that the medication had no causal impact on the treatment group. Or, the control group improves to the same extent as the treatment group does.

In this case, the doctors have to assume that there are other reasons or external factors why improvement occured. The comparison between the two groups and the two time periods, or in other words the difference in differences, can be simply measured by the so-called treatment effect. The treatment effect can be computed by first calculating the difference between the values before and after the treatment of each of the groups and then by calculating the difference between the two differences of each group.

The same logic can be applied to economics. Let us consider a situation, in which we would like to assess the impact of a specific labour market reform, which was unisono implemented by a couple of states in the U.S., on unemployment. We then could split all U.S. states into two groups: one group with those states which had implemented the labour market reform and one group that had not. After that, we would have to observe the unemployment rate before and after the implementation of the policy and could calculate the treatment effect. If the unemployment rates decrease in the treatment group after implementing the reform but do not decrease in the control group, we could conclude that the reform of the labour market had a causal impact on unemployment rates.

Although the treatment effect can be calculated by hand, it is more convenient

to compute the difference in differences in a regression model which allows us to take control of covariates and to calculate the standard errors and the significance. The classic DID regression model can be formalized as followed:

Y = β_i + β_t + γ_IT + ε

Here, β_i and β_t are dummy variables indicating the treatment and time period status, which constitute the vertical intercepts of group and time. β_i stands for the treatment/control group dummy and β_t reflects the time dummy. γ_it is the interaction term and is calculated by

γ_it = β_2i ✕ β_3t

The coefficients of the interaction term in a regression model represents the treatment effect. In the following analysis I apply the logic behind difference in differences within the context of the terrorism-tourism nexus and consider tourism demand, to use the methaphor of the clinical study, as the disease and a terrorist attack as the treatment. In doing so, I will split the countries into a treatment group, which was affected by terrorist attacks, and a control group, countries that did not suffer from terrorist attacks. In this study, however, it is not possible to apply the classic DID regression model because I do not have terrorist attacks all occurring at the same time. Rather, the data used in this analysis include many different terrorist attacks between 1994 and 2014 which occur at multiple points in time and, therefore, produce multiple time periods.

In order to deal with these multiple time periods, I created an *attack year dummy * (βt) that switches the multiple time periods for each country on and off with 1 being a country’s year with a terror attack and 0 for no attack. The second dummy variable comprises the *treatment and control group dummy* (βi) with countries who suffered from terrorist attacks being 1 and countries without any attack 0. The interaction term drops out because multiplying βt with βi produces the same dummy as βt. Instead of the interaction term, the attack year dummy will be the coefficient measuring the effect of terrorist attacks on tourism demand. The basic model used in this study can be formalized as:

Y = β_0 + β_iD + β_tD + ε

where Y is defined as the tourism demand, β_tD stands for the attack year dummy and β_iD as the treatment/control group dummy. Moreover, the basic model will vary as there are additional independent variables which will be added in the regression models. The model with all independent variables looks like this:

Y = β_0 + β_iD + β_tD + β_3 + β_4 + β_5 + ε

Here, β_3, β_4 and β_5 are the additional independent variables (GDP per capita, polity, conflict intensity), which are supposed to take control of the covariates. The term ε describes the error of the model.

These formulars will be applied to a series of regression models with different independent and dependent variables, which can be subdivided into two parts. The first part will include first regression (Reg. 1) estimating the causal effect of terrorist attacks on the absolute number of tourist arrivals per year. In doing so, Reg. 1 will be split into several different regressions which are supposed to analyze the effect in the year of an attack, the effect one year after an attack had occured (lagReg 1) and the different independent variables (Reg 1.1 – Reg 1.4).

In the second part I will analyze regression 2 (Reg 2) which focuses on the causal impact of terrorist attacks on another indicator of tourism demand, in particular the annual growth rate of tourist arrivals. Reg 2 will be also subdivided into regressions measuring the lagged effect and the additional independent variables. In order to estimate the lagged effect I have to create an additional dummy that uses the original attack year dummy as a reference and places the lagged effect dummy the year after the attack had occured.

### 4 Results

Before I will start with the regression models I will explore a number of descriptive statistics in order to give an idea of the distributions and the relationships between the variables. In the second part of the analysis I will discuss the results of the regression models and the impact of terrorist attacks on the tourism demand. Following the regression models I will conduct further statistical tests in order to assess the errors and how the model fits.

#### 4.1 Data exploration

The descriptive shows that the data include 21 observations of each of the 156 countries, which gives us in total 3,276 observations. The statistics in Table 2 show that there is a high variation of the number of tourist arrivals. While the lowest number is 3,000 tourists per year, the highest number of tourist arrivals is 83,767,000, which is represented by France in 2014. Another statistical outlier is the United States with 75,011,000 tourist arrivals in 2014. What the descriptive statistics also reveal is that there is an average of 4.3 terrorist attacks for each observation.

However, this high number is mainly due to statistical outliers, in particular 1,250 terrorist attacks in Iraq in 2014 or 336 terrorist attacks caused by Boko Haram in Nigeria. In most observations there was no terrorist attack at all and if there was, the number of attacks was rather low. The data also show a high inequality within the context of the economic development of the observations. In order to deal with these high variations, I will implement a logarithmic transformation of the annual tourist arrivals and the GDP per capita because it allows me to provide a better interpretation of huge data variations between the smallest and the largest countries and to interpret the regression coefficients as percentage points.

The mean and median of the variable \textit{polity} shows that there are more democratic countries than autocratic governments in the dataset. 50 \% of all countries pocess a score higher or equal than 6, whereas 50 \% have a score between -10 and 6. The variable \textit{conflict intensity} indicates that conflict is a rare exception among the countries in the dataset. Additionally, the descriptive statistics provide information about the number of missing variables (NAs). For instance, there are 447 missing observations of the annual number tourist arrivals. Looking closer at the data, the numbers on tourist arrivals are often missing in cases of high conflict intensities.

Before the analysis will go on with the regression models we have a closer look at the scatterplot matrix in figure 1 which illustrates the distributions and the correlations between the five main variables of Reg. 1. The labelled boxes show the distributions of each of the variables and confirm some of the conclusions of table 2. First, conflict seems to be the rare exception between 1994 and 2014 since the vast number of countries have a conflict intensity of zero. Second, as the distribution of the polity variable shows, most countries have been or have become democracies since 1994.

The relationships between each of the pairs also reveal some interesting insights. First of all, there seems to be no linear relationship between the number of terrorist attacks and tourist arrivals. Instead, the smoothed non-parametric regression line indicates that the number of tourists is going down with higher numbers of terrorist attacks but then increases again due to statistical outliers. By contrast, there is a linear relationship between the tourist arrivals and GDP per capita.

In addition to this, the scatterplot matrix supports the hypothesis that countries with a higher value in the polity scale also have higher number of tourist arrivals, although the relationship is not entirely linear. The last pair, which describes the correlation between conflict intensity and the number of tourist arrivals shows a linear relationship, indicating that a higher conflict intensity — not very surprisingly — imply lower tourist numbers, although the curve does not fall as sharply as one might expect.

Let us take a closer look at the variable *number of terrorist attacks per year*, although this will be not the dependent variable in the regression analysis. To begin with, there is no linear correlation between GDP per capita and the number of terrorist attacks, which might be explained by statistical outliers though. In fact, the graph of this pair shows that the most countries affected by terrorist attacks and the highest numbers of terrorist attacks occur within the lower half of the GDP per capita distribution.

Those countries in the treatment group with a high number of observations of terrorist attacks often represent developing countries such as Burundi, Djibouti, Cote d’Ivoire, Eritrea, Ethiopia, Guinea, Gambia, Guinea-Bissau, Haiti, Kenia, Sri Lanka, Mali, Myanmar, Mozambique, Namibia, Nepal, Pakistan, Sudan, Senegal, Sierra Leone, Tajikistan, Tanzania, Uganda or Yemen. This observation goes hand in hand with the so-called security-development nexus which basically argues that the best possible prediction for insecurity is an underdeveloped economy.

A quite interesting pair is the relationship between the polity variable and the number of terrorist attacks, which seems to support Hegre’s thesis (et al 2001: 33-48) of the inverted u-curve, arguing that the risk of civil war is a function of the level of democracy and that fully democratic states as well as fully autocratic states are likely to be more peaceful as those who are somewhere in between. A large number of the observations with terrorist attacks turn out to be, at least according to the polity scale, illiberal democracies or moderate autocracies such as Angola, Burundi, Bangladesh, the Central African Republic, Cote d’Ivoire, the Democratic Republic of Congo, Colombia, Algeria, Egypt, Ethiopia, Indonesia, Iraq (after 2003), Sri Lanka, Nepal, Pakistan, Russia, Sudan, Uganda, or Yemen.

Last not least, an interesting pair is the relationship between the two variables *GDP per capita* and *polity*, showing a more or less linear correlation between a higher GDP per capita and higher democracy level. As democracies usually imply market economies and allow more economic freedom of their subjects, this is by no means a surprising relationship.

#### 4.2 DID regressions

The following section will present and discuss the results of my DID regression models. While in the classic DID regression model the interaction term estimates the effect of a treatment on the dependent variable, in this case the attack year dummy estimates the impact of a terrorist attack on the tourism demand. A negative coefficient of the attack year dummy will imply a negative causal impact on tourism demand. As can be seen in figure 3 in the appendix, the residuals of the regression models turn out to be heteroscedastic, which means that the standard errors do not follow a normally distributed function.

This can be seen in the Q-Q plot in figure 3 located in the appendix. Because of this reason, I did not use the standard errors of the ANOVA analysis or other standard error formulas, but rather needed to compute white-robust standard errors. In addition to the graphs in the appendix, I compared the usual standard errors and the white robust standard errors, of which the latter showed much better results. The following regression tables already include white-robust standard errors which are represented by the terms within the slashes below the regression coeffecients.

In the first part I will analyze the regression models with the annual number of tourists as the dependent variable. Here, I will focus on the coefficient of the attack year dummy and its lagged effect (lagged attack year dummy). Additionally, the first part will test the additional independent variables *GDP per capita*, *polity* and *conflict intensity*. In the second part I will concentrate on the dependent variable *annual tourist growth rate* and estimate the lagged effect and the additional independent variables as described above. In addition to this, I will discuss a couple of weaknesses of the regression models in the third part.

#### 4.2.1 The impact of terrorist attacks on the number of tourists per year

As can be seen in table 3, the first set of regressions (Reg 1.1 — 1.4) measuring the effect of terrorist attacks on the number of tourist arrivals in the affected year cannot confirm the conventional view on terrorism-tourism nexus. Instead of a negative effect, as the terrorism-tourism nexus would predict, the coefficients of Reg 1.1 — 1.4 show a positive coefficient of terrorist attacks on the number of tourists.

Due to the fact that the regressions estimate the effect on the *logarithm* of the absolute number of tourism, the results can be interpreted as percentage points. As Reg 1.1 estimates, a terrorist attack was followed by an average increase in tourist numbers by 2 percentage points. Taking control of the covariates logBIP per capita, polity and conflict intensity the coefficients increase even more, predicting an average increase in tourist numbers of more than 64, 66 and 73 percentage points following a terrorist attack. Additionally, the coefficients have a high level of significance.

At the same time, the effects of the polity variable is very low with an average increase of the tourist numbers per year of 3 percentage points if the polity scale increases by 1. The level of the conflict intensity, however, shows a slight negative impact on the tourist numbers per year, decreasing the number of tourists by almost 4 percentage points following one level higher in the intensity scale. A very strong impact, by contrast, can be seen in the GDP per capita level. One percentage growth of the GDP per capita leads on average to a rise of the tourist numbers by 91 percentage points. The strong impact of the GDP per capita can be also seen in the R^{2} explaining 45 % of the given variance.

The picture slightly changes when we look at the regressions estimating the lagged effect (table 4), though not very much. The first regression (lagReg 1.1), which computes the lagged effect as the independent variables on the tourist numbers without any other additional independent variables, shows a negative coefficient, decreasing the number of tourists per year by 4 percentage points.

Given these results of the regression models, it is hard to confirm the terrorism-tourism nexus. The regression model lagReg 1 might show a negative coefficient, but the evidence for the terrorism-tourism nexus is still too weak. First, the coefficient is very low with a coefficient of -4.4 percentage points one year after an attack. Second, the negative effect of lagReg 1.1 is not significant, which means that the risk of a coincidental relationship is very high. And third, as soon as we take control over the other independent variables, the coefficients of the attack year dummies massively increase. In short, there is no hard evidence supporting the terrorism-tourism nexus.

#### 4.2.2 The impact of attacks on the annual growth rate of tourists

To check these results I will test another set of regressions (Reg 2.1 – 2.4) which are supposed to estimate the impact on the annual growth rate of tourist arrivals, an additional variable indicating tourism demand. For this, I calculate the growth rate of the tourist arrivals per year by computing the annual change (in %) in the absolute numbers of the tourist arrivals. From a technical point of view, this means that the coeffecients of the attack year dummy will be negative, if the annual growth rate of the tourist arrivals will be lower than the year before a terrorist attack had occurred.

The tables 6 and 7 depict the results of the second set of regressions. However, the regressions show only a slight change in the overall picture. The coefficient of the attack year dummy on the annual tourist growth rate might be negative in Reg 2.2 and Reg 2.3 but the coefficients are extremely slight and they are not significant. The same is true for lagReg 2.3 which estimates the lagged effect of terrorist attacks on the annual tourist growth rate. To sum up, only three out of eight regressions show a slightly negative attack year dummy coefficient and none of them is significant. In other words, there is again no hard evidence for the terrorism-tourism nexus.

#### 4.2.3 Alternative explanations and models

The answer why the terrorism-tourism nexus is not as straight forward as many expected might be a complex one. Looking closer at the data, there are a complex set of variables and a variety of trends that could diminish, or at least mitigate, the negative effects of terrorist attacks on the overall tourism demand. Of course, there are specific examples in the dataset showing a decrease in tourist arrivals after a terrorist attack. This includes countries such as the United Kingdom in the late 1990s, the United States after 9/11, Yemen since 2006 or Kenya since 2011.

However, most of the annual tourist numbers of countries in Asia, Africa or in South America are growing, despite the fact that many of them frequently suffer from terrorist attacks. India, for instance, has been constantly affected by a certain degree of terrorism since the 1990s, but the long-term trend shows a constant year to year growth of its tourism industry. A similar pattern can be seen in countries such as Bangladesh, Pakistan, Myanmar, Nepal since 2003, China, Indonesia, Thailand, Cambodia, the Philippines, Uganda, Angola, Sudan, the Central African Republic since 2002, Algeria, the Democratic Republic of Congo since 2002, Ethiopia, Israel (since 2003), Lebanon until 2011, Iran, Colombia, Mexico, Peru, Bosnia and Herzegovina, or Russia. Turkey too, has been affected by terrorist attacks in almost all of the years between 1994 and 2014 but had a constant year to year growth of its tourist numbers from 7,083,000 in 1995 to 39,811,000 tourist arrivals in 2014.

The left graph in Figure 2 displays the panel data of the global trend in the number of tourist arrivals of all countries between 1995 and 2014, whereas the right graph shows that the global GDP massively increased between 1994 and 2014, both of which might be part of the explanation why the terrorism-tourism nexus is not as straight forward as our intuition expect. A wide range of members of the treatment group, most of which are listed above, have a growing population, high GDP growth rates and a rising international and domestic tourism industry. Tourist numbers in these emerging countries continue to rise even after a terrorist attack because modernization and emerging middle classes encourage their domestic and international tourism industries.

Another variable that correlates with the annual tourist numbers is the size of the population of a country. Since most of the emerging markets and societies also have higher population growth rates, this is a factor that can also be part of the explanation. In other words, there are a variety of factors that diminish or mitigate the terrorism-tourism nexus such as growing populations, rising economies and emerging middle classes, all of which lead to growing domestic and international tourism markets.

Given this complex set of variables which probably affect the relationship between terrorist attacks and tourist numbers, we can try to find regression models which take these variables into account. For this reason, I designed a couple of additional regression models. In the first regression model listed below (Reg 3 and lagReg 3) I have repeated Reg 1.2 and lagReg 1.2 with the extension that I included the country variable as fixed effects. Fixed effects allow me to control non-random variables with an unobserved heterogeneity, all of which are constant over time.

Furthermore, this time I exclude the variables *polity* and because the regression models above did not show any considerable linear influence on tourism demand. Neither could this variable explain much of the variation measured in R^{2}. As the coefficients of Reg 3 in Table 8 show, the overall picture changes considerably with terrorist attacks decreasing the number of tourists by 8 percentage points in the year of the attack.

The model lagReg 3, which estimates the lagged effect one year after an attack had occurred, shows a similar pattern with terrorist attacks decreasing the number of tourists by 10 percentage points. At the same time, all of the variables have a high level of significance and much lower standard errors than in the regressions in the sections before.

Moreover, I repeated the regression models Reg 1.2 and lagReg 1.2 but included the variable *logarithmic population* in order to take control over variations in the population change (Reg 4 and lagReg 4). This time the effect of terrorist attacks is even stronger with a decrease in the number of tourist arrivals by 32 percentage points in the year of the attack and 35 percentage points one year after an attack had occured. Moreover, all of the coefficients are highly significant and the standard errors are much lower than in the previous regression models.

### 5 Discussion and Conclusion

For as long as studies have analyzed the terrorism-tourism-nexus, the media and pundits have warned about terrorism harming the tourism industry. One of the arguments made in the academic literature is that terrorist attacks show a negative impact because of the flexibility of the supply- and demand-side because consumer preferences can change quickly. This thesis cannot be confirmed by this analysis as most countries in my dataset facing terrorist threats show annual growth rates in tourism numbers. Moreover, risk perception theory argues that terrorist attacks have an impact on the behavior of tourism consumers, which can lead to a decreasing tourism demand. With the exception of a few examples, this argument could not be verified by the main regression models because a variety of other factors diminish or mitigate the negative impact.

Nevertheless, tourism industries might collapse as the terrorism and conflict gets worse. Future research could address this issue by asking at which level of violence rising tourism industries stop to grow and which role the media perception plays in this. Also, a variety of scholars have pointed out that there might be a couple of different variables affecting tourism demand such as human rights violations. Assuming that democracies imply fewer human rights violations than autocratic states, this thesis could not be confirmed with hard evidence. Although the scatterplot matrix and the regression models showed a slight impact, the correlation and impact were negligible, or at least they were not linear and would, therefore, require another set of logical arguments to explain a non-linear correlation.

As the DID regression models show, the reality is more complex than the terrorism-tourism-nexus makes us to believe. In the vast majority of cases the tourist numbers are not lower after an attack than before and even continue to grow after a terrorist attack had occured. As the regression models indicate, not even the annual growth rate of the tourist numbers was lower than before. As the data reveal, most of the countries affected by terrorist attacks are emerging countries with growing GDPs and populations, emerging middle classes and rising domestic and international tourism industries. A modest negative impact on tourism demand can be seen only if we account for these factors and assume that these variables remained constant over time.

As most of the authors dealing with the issue pointed out, there are additional factors which explain tourism demand such as prices, media coverage or the general attractiveness of a tourist destination. Although this analysis suggested that the terrorism-tourism nexus might be more complex than many believe, absolute numbers of tourist arrivals can be best predicted, at least according to my findings, by looking at the population number and the GDP per capita.

### References

Ahlfeldt, Gabriel M; Franke, Bastian; Maennig, Wolfgang (2015) Terrorism and International Tourism: The Case of Germany, in: Jahrb¨ucher für Nationalokonomie und Statistik, Vol. 235, No. 1, pp. 3-21.

Arana, Jorge E.; Leon, Carmelo J. (2008) The Impact of Terrorism on Tourism Demand, in: Annals of Tourism Research, Vol. 35, No. 2, pp. 299-315.

Baker, David Mc. A. (2014) The Effects of Terrorism on the Travel and Tourism Industry, in: International Journal of Religious Tourism and Pilgrimage, Vol. 2, No. 1, pp. 58-67.

Diriye, Hussein Gure (2015) The Effect of Terrorism on Tourism in Africa in the 21st Century: A Case Study of Kenya, University of Nairobi.

Enders, Walter; Sandler, Todd (1991) Causality between transnational terrorism and tourism: The case of Spain, in: Terrorism, Vol. 14, No. 1, pp. 49-58.

Hegre, Havard; Ellingson, Tanja; Gates, Scott; Gleditsch, Nils Petter (2001) Toward a Democratic Civil Peace? Democracy, Political Change, and Civil War, 1816-1992, in: The American Political Science Review, Vol. 95, No. 1, March, pp. 33-48.

Martin, Thomas L.; Gu, Zheng (1992) Terrorism, Seasonality, and International Air Tourist Arrivals in Central Florida, in: Journal of Travel and Tourism Marketing, Vol. 1, No. 1, pp. 3-15.

Neumayer, Eric (2004) The Impact of Political Violence on Tourism, in: Journal of Conflict Resolution, Vol. 48, No. 2, April, pp. 259-281.

Neymayer, Eric; Pl¨umper, Thomas (2016) Spatial spill-overs from terrorism on tourism: Western victims in Islamic destination countries, in: Public Choice, No. 169, pp. 195-206.