You are on page 1of 27

Quantifying fear: the social impact of terrorism1

Juan Prieto-Rodrguez
Universidad de Oviedo, Avda. del Cristo s/n, 33071 Oviedo Tel: +34 985 103768. E-mail: jprietor@uniovi.es

Juan Gabriel Rodrguez*


Universidad Rey Juan Carlos, Campus de Viclvaro, 28032 Madrid Tel: +34 91 4887948. E-mail: juangabriel.rodriguez@urjc.es

Rafael Salas
Universidad Complutense de Madrid, Campus de Somosaguas, 28223 Madrid Tel: +34 91 3942512. E-mail: r.salas@ccee.ucm.es

Javier Suarez-Pandiello
Universidad de Oviedo, Avda. del Cristo s/n, 33071 Oviedo Tel: +34 985 103768. E-mail: jspandi@uniovi.es

Abstract This paper proposes a methodology to measure social impact of terrorism. We define a multidimensional terrorism index based not only on deaths but also on other variables such as injuries, bombs and kidnappings. The weight of each terrorist activity is given by its social impact, which is estimated through its relevance in the media. For this task we build up a new data set from the four most important newspapers in Spain, namely, El Pas, El Mundo, ABC and La Vanguardia. Finally, we evaluate the social impact of ETA terrorism in Spain from 1993 through 2004.

Keywords: terrorism, multidimensional index and social impact weight. JEL Codes: C20, D00 and Z13.

*Corresponding author

We are grateful for assistance with the data base from Jorge Reones. We also acknowledge useful comments and suggestions by the participants in the Lisbon Conference on Defence and Security 2008. This research has benefited from the Spanish Ministry of Science and Technology Projects SEJ200764700/ECON and SEJ2006-15172/ECON. The usual disclaimer applies.

The calculated use of violence or the threat of violence to inculcate fear; intended to coerce or to intimidate governments or societies in the pursuit of goals that are generally political, religious, or ideological. (Definition of terrorism by the U.S. Department of Defense)

1. INTRODUCTION Since September 11th terrorism has became a global problem and a worldwide concern. In fact, terrorist activities are one of the most important worries for those societies that suffer from this problem (for example, Ireland and Spain). Though terrorism is easy to observe, its measurement is a difficult task (see Frey and Luechinger, 2005). The traditional measurement of terrorism based on the number of terrorist events and/or casualties (see, for example, the time series study in Enders and Sandler, 2002) does not consider the consequences for people in terms of economic or utility losses. However, sophisticated techniques also have problems. Impact studies that measure individual losses in terms of monetary revenue (see, among others, Enders and Sandler, 1991; Enders et al. 1992; Enders and Sandler, 1996; Caplan, 2002; Drakos and Kutan, 2003 and Abadie and Gardeazbal, 2003) exclude non-market values so, in this manner, they may underestimate the phenomenon of terrorism. The hedonic market approach relies on the assumption that labor and housing markets are in equilibrium meanwhile the averting behavior approach assumes perfect substitutability between individual and public expenditures for the mitigation of terrorism effects. Moreover, both approaches cannot capture the negative external effects of terrorism over non-use values.2 A

The non-use values are the following: existence value; option value; education value; and, prestige

value.

different method, the contingent valuation survey includes non-use values (see, for instance, Viscusi and Zeckhauser, 2003); however, strategic response biases are possible. Moreover, this method relies on the individual valuation of a specific public good. This evaluation is a demanding cognitive task so superficial responses without adequate consideration of substitutes and the budget constraint may result. The political reaction of voters may capture individuals evaluation of anti-terrorist policies (see, for example, Nacos 1994). The problem with this method is that voters may solely evaluate a government by its outcomes. In this case, citizens could support a particular government that has not undertaken any policy to fight against terrorism if terrorism declines due to other factors. Another method to measure terrorism in the literature is the induced change in happiness.3 This approach captures non-use values though utility losses may reflect not only the effects of terrorism but also government reactions. This paper proposes a class of indices to measure terrorism. In particular, we propose a multidimensional index of terrorism based on a set of dimensions or factors that generate relevant social impact. The basic variables that we consider are the following: number of people murdered or injured; type of attack; scale of terrorism action; and, number of kidnapped people. In the construction of our index we assume a certain degree of substitution among different dimensions or variables. As a consequence, we discard lexicographic orderings. Moreover, we weight dimensions according to social valuation. The valuation that society makes of terrorist activities is not directly observable so we must consider a proxy. We take, in particular, the relevance of terrorist activities in newspapers as a proxy for social valuation. In this manner, we capture utility losses due to terrorism. Furthermore, non-market values and non-use
3

The relationship between economics and happiness is analyzed, among others, by Frey and Stutzer,

2002 and Bruni and Porta, 2005.

values are, in principle, included and the contamination of government reactions is avoided. Nevertheless, the reliability of our measure will depend on the gravity of the gap between public opinion and coverage given to terrorist acts by newspapers. To measure the degree of terrorism impact in newspapers, we have built up a new data set with the incidence of ETA activities in the four most important newspapers in Spain, namely, El Pas, El Mundo, ABC and La Vanguardia. This data set covers the period 1993-2004. Therefore, it provides information for a complete political cycle including three general elections and their corresponding four legislative periods: two governments of the Partido Socialista Obrero Espaol (PSOE, henceforth) and two governments of the Partido Popular (PP, henceforth).4 Once we have estimated the weight for each dimension, the evolution of ETA terrorism in Spain is analyzed. In general terms, we find that ETA terrorism has nonmonotonically decreased since 1994. The paper is organized as follows: we proposed a multidimensional index of social impact of terrorism in Section 2; the data sets used to compute this index and the regression models to empirically determine the weight assigned to each dimension are presented in Section 3; the social impact in Spain of ETA activities during the period 1993-2004 is displayed in Section 4; we comment the usefulness of the proposed index in Section 5; and finally, in Section 6, we discuss the main results.

See Barros and Gil-Arana (2006) for a study on ETA activity during the last 30 years.

2. SOCIAL IMPACT OF TERRORISM: A THEORETICAL PROPOSAL FOR AN INDEX We develop in this section a theoretical proposal for measuring social impact of terrorism. This proposal should be eligible for any country or region and any time. A first decision to be made is to choose between an ordinal and cardinal approach. We focus our attention on cardinal approaches since we are interested in comparisons between periods and regions based on total levels of terrorism activity and not on relative terms. Once this option has been taken, we must decide between a uni-dimensional and multidimensional approach. The problem with a uni-dimensional index of terrorism, for example one exclusively based on casualties, is that other dimensions of terrorism are missed. In fact, people may suffer from great terrorist threats even if there are no casualties. As said above, measurement of terrorism based on the number of terrorist casualties does not consider the consequences for people in terms of economic or utility losses. Consequently, estimations of terrorism based in just one dimension may be biased. We believe, accordingly, that any possible dimension of terrorism must be considered so we advocate for a multidimensional approach. Moreover, we show below that the uni-dimensional approach is a particular case of our proposal. That is, both methodologies converge when the coefficients for all dimensions other than one are not statistically different from zero. In Section 4 we estimate the empirical relevance of including variables other than killed people for ETA terrorism in Spain. An important issue in multidimensional measurement is the choice of dimensions. One possible criterion to choose relevant dimensions is orthogonally. To compute this orthogonally we could apply multivariate analysis using information on different

terrorism dimensions.5 However, it is difficult to find orthogonal dimensions in the terrorism framework. Variables like the number of people murdered or injured, bombs, scale of terrorism action and the number of kidnapped people do not seem to be orthogonal at all. Therefore, we work directly with observable variables which can be, in principle, substitutes or complements. We assume that dimensions are substitutes, think for example about the relationship between total number of killed people and injured people. Nevertheless, some complementarities are possible when dealing with directly observable terrorism variables.6 The degree of substitutability between dimensions is calculated through a regression approach in Section 3. Once a multidimensional cardinal approach has been adopted, there are two main possible ways to define a terrorism index: the axiomatic approach and the information theory approach. In the first case, a set of axioms is imposed over the measure. In the second case, aggregation of dimensions for terrorism assessments adopts the Information Theory approach (see Theil, 1967). In this paper we explore the last approach which is formally presented next.7 Let N = {1, 2, ..., n } be the set of terrorist attacks in a period of time and D = {1, 2, ..., d }
the set of indicators or dimensions, whether they are of a quantitative or qualitative nature. An outcome matrix X is an n x d matrix whose element xij is the outcome of

5 6

Some of these techniques are Principal Components Analysis, Factor Analysis and Cluster Analysis. For instance, the magnitude of a bomb seems to be positively correlated with casualties. However, we

do not have information about the magnitude of a bomb, only about the type of bomb.
7

Measurement of multidimensional poverty applies a somehow related methodology (see Maasoumi

1986). In this field, all relevant attributes of well-being are assumed to be perfectly substitutable, though some scholars have recently suggested the existence of a partial trade-off between attributes (see Bourguignon and Chakravarty 2003, Tsui 2002).

the terrorism attack i in terms of the dimension j. The domain of outcome matrices denoted by is restricted to matrices of nonnegative integer numbers. Then, a multidimensional terrorism index T () for a given period of time is a mapping from matrix X to a number in the set of real-valued numbers: T ( X ) = f ( X ) : .

This index must fulfill the following normalization property: T(X) = 0 whenever xij = 0 for all i and j. However, this index cannot be normalized to 1 for the case terrorism is maximum because, in principle, there is no upper bound for terrorist activity. Now, we discuss the theoretical properties of this index trying to explain its relevance. A first possibility for T(X) could be the lexicographic ordering of social preferences. Given two scenarios X and Y (for example, two countries or two periods of time for the same country) the lexicographic ordering is as follows:

x 1 y1 x1 = y1 , x2 y2 T(X) T(Y) = ... x j = yj , x d y d j d


where x j = xij denotes aggregate value for any j dimension. We assume j is ordered
i =1 n

according to certain social hierarchical relevance, such as j = 1 is the number of killed people in a terrorist attack, j = 2 is the number of injured people and so on. In words, as long as the first dimension of X is larger than that of Y, X shows a higher degree of terrorism, and this is so even if the rest of dimensions are larger for Y. But as soon as the first dimension become equal only the second dimension is relevant.

Unfortunately, the lexicographic ordering is just an extreme possibility. There are two main problems with this ordering. First, it assumes that there is no substitution between dimensions at all. For example, one murder is always worst than any amount of injured people. Second, it is not marginally affected by small variations in xij, that is, it is not continuous because it may exhibit jumps. To avoid these problems we adopt a different approach. The idea is to replace the n pieces of information on the values of terrorist attacks for the various dimensions by a composite indicator xc = ( xc1 ,..., xcd ) which is a vector of d scalars, one for each dimension. For this, the vector ( x1 j ,..., xnj ) corresponding to the dimension j is replaced

by the scalar xcj . This scalar may be considered as representing the level that dimension j derives from the terrorist attacks that have occurred in a society during a given period of time. In this paper, we consider that xcj = x j though other alternatives are possible because terrorist attacks are not necessarily alike. For instance, we could weight more those terrorist attacks that take place before an election (see Barros et al., 2006 and Berrebi and Klor, 2006), or those which occur at the beginning of the period when people have not got used to them yet. To make things simple we assume that terrorist attacks are equally important so they will receive the same weight. Nevertheless, we introduce a dummy for each year in the proposed regression model to control for the year when terrorist attacks take place (see Section 3). Next, we have to select appropriate weights for the dimensions in the composite vector xc . We propose to weight each dimension by its presence in the media. That is, we consider that the main purpose of terrorist activity is to make terrorists goals notorious to society. However, the valuation that society makes of terrorist activities is not directly observable so we proxy the social relevance of each dimension by its presence

in television, newspapers, reviews and so on. The advantage of this weighting procedure is that utility losses of terrorism are considered. Moreover, non-market values and nonuse values are captured without the contamination of government reactions. The estimation procedure of these weights is presented in Section 3. Finally, what is the proposed model or aggregation function for the index T()? In principle, we could think about the translog function or the constant elasticity of substitution function (CES) because of their flexibility to support different alternatives about substitution between variables. Unfortunately, the estimation of these functions requires the specification of each component xcj in logarithms so zero values are not allowed. All terrorist activities take this value from time to time (fortunately!), consequently, we need to rely on a different model. A first candidate is the CobbDouglas function which defines imperfect substitution between terrorism dimensions but it is not possible for the same reason above. We eventually propose a linear and quadratic aggregation functions. In the linear case, the terrorism index is the following:
d

T ( X ) = j xcj
j =1

where j is the weight of dimension j. This functional form assumes perfect substitution between dimensions though they have different weights. Because of this disadvantage we propose a quadratic model. In this case, the terrorism index is the following:
d

T ( X ) = j xcj + 11 xc21 + 22 xc22 + 12 xc1 xc 2


j =1

where dimensions 1 and 2 are the number of killed people and injured people, respectively. Note that the rest of explicative variables (j = 3, ..., d) will be dummies so it does not make sense to extend the quadratic form to these variables. The advantage of this functional form is that the marginal rate of substitution is not constant (see next section).

3. ESTIMATION OF WEIGHTS
To make our index T() applicable we need to estimate the weight for each dimension. As said in the previous section, we propose to weight each dimension by its presence in the media as a proxy for social valuation. Accordingly, we build up a data set with information about terrorist activities in newspapers. For this empirical exercise, we focus on a long term terrorist problem in Spain since the last years of Francos era, the terrorist activities of ETA. Since the 11th March 2004 train bombing was not an ETA attack, it is excluded from our study.8

3.1. DATA BASE


The data set Terrorism in Western Europe: Events Data (Engene, 2006) is a valuable source of information for analysing patterns of terrorism in Western Europe. From this data set, we obtain information for the following dimensions of terrorism: total number of killed people; total number of injured people; type of attack (letter bomb, car bomb, other bomb, rocket or grenade attack, armed attack and other attack); kidnappings; and,

International terrorism is analysed, among others, in Engene (2004), Bellany (2007) and Barros et al.

(2007).

10

type of target (military, police, public service, political institution, business and civil). The descriptive statistics for these variables are shown in the appendix (see Table A.1.) Moreover, using information contained in this data set to locate ETA activities through time, we have collected daily data from the four most important newspapers in Spain, namely, El Pas, El Mundo, ABC and La Vanguardia. The first three newspapers are national meanwhile the last one is local, in particular, Catalan. Moreover, the ordering of national newspapers attending to their conservatism is the following: ABC; El Mundo; and, El Pas. We have considered the following twelve variables: photograph in the cover page; percentage of the news about terrorism on the cover page; percentage of the news about terrorism in the editorial section; total number of pages; photograph on the first interior page; percentage of the news about terrorism on the first interior page; photograph on the second interior page; percentage of the news about terrorism on the second interior page; photograph on the third interior page; percentage of the news about terrorism on the third interior page; percentage of the news about terrorism on the second day editorial; and, second day total number of pages. The descriptive statistics for these variables are shown in the Table A.2. (see appendix). Information about ETA activities has been collected for a complete political cycle including three general elections and their corresponding four legislative periods: two governments of the PSOE from 1993 through 1996 and from 2004 through 2008; and two governments of the PP from 1996 through 2000 and from 2000 through 2004. Note that the data set only provides information for the first year of PSOEs second government though there was a truce from the 22nd of March 2006 through the 30th of December 2006.

11

3.2. REGRESSION ANALYSIS


The first step in the estimation of weights is to define an appropriate dependent variable. This variable has to represent the social valuation of terrorist activities. For this we apply Factor analysis to the data collected from Spanish newspapers. In model 1 we only consider information about the following day, in particular: photograph on the cover page; percentage of the news about terrorism on the cover page; percentage of the news about terrorism on the editorial section; and, total number of pages. The results for this model are shown in Table 1.

Table 1. Factor analysis of model 1. Eigenvalue Difference 2.46068 2.11617 0.34451 0.45422 -0.10970 0.05324 -0.16295 . Weight Factor 1 0.7526 0.8935 0.7066 0.7724 Proportion 0.9716 0.1360 -0.0433 -0.0643 Cumulative 0.9716 1.1077 1.0643 1

Factor1 Factor2 Factor3 Factor4

Variable Cover photograph % Cover % Editorial Total number of pages N LR test [independent vs. saturated chi2(66)] Kaiser-Meyer-Olkin measure AIC (Factor1) BIC (Factor1)

Uniqueness 0.4336 0.2016 0.5007 0.4034 537 1262.83 0.6888 199.4424 216.5863

The eigenvalue of a particular factor captures its variance. Accordingly, the column of proportions gives the part of total variance that is explained by each factor. Thus, the first factor accounts for more than 97% of total variance meanwhile the second factor accounts for 13% of total variance and so on. The last two eigenvalues are negative 12

because the matrix has not full rank, that is, although there are 4 factors the dimensionality of the factor space is smaller. Given these results we have taken the factor 1 as our dependent variable. Note that the weight of each variable of information in the construction of Factor 1 is provided below the eigenvalues. Moreover, it is shown the Kaiser-Meyer-Olkin measure of sampling adequacy. This is an index for comparing the magnitudes of observed correlation coefficients with the magnitudes of partial correlation coefficients. Large values for the Kaiser-Meyer-Olkin measure indicate that a factor analysis for a set of variables is a good idea. We also consider an extensive model (model 2) which includes information about the second day. Results are shown in the appendix (see Table A.3.). It is apparent from the results in Table 1 and Table A.3. that the first factor explains the major part of total variance. In this sense the result of model 1 is robust. Moreover, the Kaiser-MeyerOlkin measure increases from 0.69 to 0.82. However, the required level of information, twelve variables instead of four, significantly reduces the net benefit of extending the model. For these two reasons we compute the dependent variable, social impact, by the Factor 1 in model 1. Once social impact has been estimated by Factor analysis, we regress this dependent variable on the terrorism dimensions specified above. The estimated coefficient of a particular terrorism dimension will represent the part of social impact that such dimension explains. We have also included in the regression a dummy for each year to control for time and a dummy variable for El Pas, El Mundo and La Vanguardia to control for the difference between newspapers. In Table 2 the results for the linear and quadratic specifications are shown. Note that the variables other attacks, civil target and ABC are used as references so they are not included in the regression.

13

Table 2. Regression of model 1.


Social impact (model 1) Total killed Total killed squared Total injured Total injured squared Total killed * Total injured Letter bomb Car bomb Other bomb Rocket or grenade attack Armed attack Kidnapping Military target Police target Public service target Political institutions target Business target El Mundo El Pas La Vanguardia Constant N R2 F Lineal 0.22748*** (0.038) ----0.01313*** (0.003) --------0.42541* (0.224) 0.46301*** (0.135) 0.04976 (0.148) -0.26524 (0.374) 0.71809*** (0.140) 0.78329*** (0.256) 0.25213* (0.133) -0.09364 (0.104) -0.01358 (0.128) 0.41084*** (0.108) -0.27707 (0.197) -0.27326*** (0.084) -0.35276*** (0.085) -0.49669*** (0.083) -0.46497** (0.201) 537 0.4929 19.064 Quadratic 0.62002*** (0.088) -0.06931*** (0.016) 0.04791*** (0.009) -0.00039*** (0.000) -0.00569*** (0.002) 0.49863** (0.218) 0.46053*** (0.133) 0.0472 (0.147) -0.10021 (0.364) 0.64656*** (0.136) 0.86321*** (0.249) 0.17283 (0.130) -0.23505** (0.107) 0.07702 (0.129) 0.34341*** (0.107) -0.23694 (0.191) -0.26716*** (0.081) -0.35538*** (0.082) -0.49014*** (0.081) -0.62766*** (0.198) 537 0.527 19.481

***: Significant at the 1% level. **: Significant at the 5% level. *: Significant at the 1% level. Standard deviations in parentheses.

14

First, the quadratic specification achieves a better fit than the linear specification. Besides the square terms and the cross effect are negative and statistically significant at 1%. A negative value for the square terms implies certain saturation in the social impact of total killed and injured people. In fact, these two variables reach their maximum social impact at 4 killed people and 60 injured people, respectively. Second, kidnapping reaches the largest coefficient. We could think, therefore, that the best strategy for a terrorist group like ETA, in terms of social impact, is kidnapping instead of killing people. In fact, kidnapping also allows terrorist groups to collect money. However, we think that terrorist groups like ETA do not widely adopt this strategy because kidnapping is a great source of risk and cost for the gang. Moreover, the fact that the variable armed attack has also a larger coefficient than total killed people in the quadratic specification reinforces our multidimensional approach, in the sense that total killed people does not include the whole terrorism phenomenon. Next section shows the difference between the proposed multidimensional index and a measure based exclusively on killed people. Third, newspapers have a different sensitivity towards terrorism. As said above, the variable ABC is used as reference so the negative sign of the coefficients for El Pas, El Mundo and La Vanguardia implies that these newspapers are less sensitive towards terrorism than ABC. Thus, the ordering of the Spanish newspapers according to this sensitivity is the following: ABC; El Mundo; El Pas; and, La Vanguardia. The more conservative is a national newspaper, the greater the sensitivity towards terrorism. Note that the inclusion of a dummy variable for the type of newspaper allows us to control for newspapers ideology. Unfortunately, it does not permit to control for the gap between public opinion and coverage given to terrorist acts by newspapers. Media

15

coverage of terrorist events and publics attitudes and perceptions may be different if we face a common-interest game. In this case, terrorists get free publicity for their cause and media make money as reports of terror attacks increase newspaper sales (see Rohner and Frey, 2007). Fourth, we assumed in Section 2 that total number of killed and injured people were substitutive variables. Moreover, the quadratic specification of our model implied a non-constant marginal rate of substitution between both dimensions. Both assumptions are confirmed in Table 3 where the estimations for the marginal rate of substitution at different quantiles are shown. Bear in mind that the marginal incidence measures are the derivative of social impact with respect to the number of killed and injured people. Meanwhile, the marginal rate of substitution is the ratio of both variables. The rate of substitution for the mean is 12.22. Finally, we have also computed a regression for social impact in model 2. The results are shown in the appendix (see Table A.4.). In general terms, estimations in Table 2 are replicated for a more extensive definition of social impact.

Table 3. Marginal Rate of Substitution.


Marginal incidence of killed Quantile People 0.6200 5% 0.6200 10% 0.6200 25% 0.4814 50% 0.4757 75% 0.2916 90% 0.2574 95% Marginal incidence of Injured People 0.0479 0.0479 0.0479 0.0422 0.0414 0.0295 0.0248

Marginal Rate of Substitution 12.9413 12.9413 12.9413 11.4021 11.4794 9.8803 10.3677

16

4. SOCIAL IMPACT OF ETA TERRORISM IN SPAIN


The final step is to calculate the multidimensional index T(). For this, we have used the estimated coefficients in Table 2, after eliminating the variables of control (dummies for time and newspapers). To make comparisons, we have also computed a terrorism index based exclusively on the total number of killed people (lexicographic ordering). The number of attacks, deaths and the value of index T by year are shown in the appendix (see Table A.5.). Moreover, we have applied nonparametric techniques to smooth the diary evolution of deaths and index T. In particular, we have used the Nadaraya-Watson nonparametric smoother (Nadaraya, 1964 and Watson, 1964) with the Epanechnikov kernel (Epanechnikov, 1969). In Figure 1 we show the evolution of both indices of terrorism. Note that the index T has no dimensions so the scale in the vertical axis corresponds to the number of killed people. Moreover, there was a truce from 16th of September 1998 through 3rd of December 1999. The vertical lines in Figure 1 represent this period of time. Two facts become apparent from Figure 1. The first fact is that both measures need not to coincide in their diagnosis. That is, one index can indicate that terrorism is increasing meanwhile the other can indicate the opposite (see, for example, the first half of 1995, the beginning of 1997 and 2004). A more important fact is that the killed people index does not include the whole information about terrorist activities so it smoothes terrorism. This causes a more linear evolution of the killed people index in comparison with the proposed index T() which shows higher crests and deeper valleys. For the whole period we observe that ETA terrorism has non-monotonically decreased since 1994.

17

Figure 1. Evolution of ETA terrorism in Spain (1993-2004).

5. ECONOMIC APPLICATIONS
In previous sections we have described the contents and main properties of the proposed index of terrorism. Next, we suggest some fields where this index can be applied. Obviously, this is not a closed list of issues, but only a sort of examples exploring where the use of our proposal can be useful. Evaluation of terrorism costs In the economic literature, some studies have recently assessed the costs of terrorism. They adopt a macroeconomic perspective (see Caplan, 2002 and Abadie and Gardeazbal, 2003) or a microeconomic perspective based on the quality of life lost (see Frey et al., 2007) or the actual loss of human life (Riera et al., 2007). We think that our

18

index, insofar as it approximates the social relevance of terrorism, can be used as a complementary tool in this field. Economic impact of terrorism in development Another issue that has received increasing attention is the analysis of the determinants for development (see for example Olson, 1996 and Murdoch and Sandler, 2002). A line of research that has given some fruits in this field explores the impact of terrorism in development. In particular, terrorism has been analyzed through its effects on capital markets (Chen and Siems, 2004), foreign investment (Enders and Sandler, 1996) and economic activities linked to specific sectors, typically tourism (Enders et al., 1992 and Llorca-Vivero, 2008). A terrorism index like the one we are proposing may help researchers to explain these phenomena. For example, economic activities like tourism and foreign investment crucially depend on the perception that foreign people have on countrys security. Accordingly, it might be a good idea to study the incidence of foreign public opinion on economic growth. The difference with respect our study would be the source of information: foreign newspapers. Terrorism, institutions and public choice A third field where our indicator can be applied is Institutional Economics. More precisely, the proposed terrorism index can be used to study the roots of institutions and processes of collective choice. A good example of this is the number of studies devoted to analyze the connections between economic variables like poverty, cultural variables like education and terrorism (see, among others, Berreby, 2007 and Krueger and Malekov, 2003). Another example from a more theoretical view is the number of studies that question the irrationality of terrorism (Caplan, 2001a, 2001b and 2008) or its effects on the duration of Governments (Saez, 2002).

19

Terrorism and public economics Finally, another research guide that can be supported by an instrument such as the one proposed here is the analysis of terrorism effects on the allocation of public spending. Thus, terrorism affects public spending not only through protection of citizens (defence and police spending) (Higgs and Kilduff, 1993 and Prez-Fornis et al., 2004), but also through the support of victims and reconstruction of damages. In all these cases, possibly more, we think that a multidimensional index based on the social valuation of terrorism such as the one suggested here, can help to understand the economic roots and effects of this social scourge.

6. CONCLUSIONS
This paper proposes a multidimensional index of terrorism. We construct this measure by aggregating the different dimensions of terrorism, namely, killed people, injured people, bombs, kidnappings and targets. For this aggregation we estimate the weight of each dimension by regressing the social impact of terrorism on these dimensions. However, social impact is not directly observable so we consider information in the media as a proxy. Finally, we apply our index to the terrorism of ETA in Spain and compare it with the evolution of the total number of killed people. The comparison shows that the effort to compute the proposed index is worthwhile. Let us finish this paper with the following comment. The nature and political objectives of national and international terrorist groups are usually different. Thus, national terrorism is usually linked to territorial defined political objectives, independence in most of the cases. On the contrary, global terrorism is usually focused on factors like

20

religion and the socio-economic organization of a nation. However, both kinds of terrorism manifest themselves through their impact on the citizens state of mind. The methodology proposed here measures the social impact of terrorism so, in principle, it can be applied in the measurement of both kinds of terrorist activities.

REFERENCES
Abadie, A. and J. Gardeazbal (2003): The economic cost of conflict: A case study of the Basque Country, American Economic Review, 93: 113-132. Barros, C. P. and L. A. Gil-Alana (2006): ETA: a persistent phenomenon, Defence and Peace Economics, 17 (2): 95-116. Barros, C. P., J. Passos and L. A. Gil-Alana (2006): The timing of ETA terrorism, Journal of Policy Modelling, 28 (3): 335-346. Barros, C. P., J. R. Faria and L. A. Gil-Alana (2007): Terrorism against American citizens in Africa: Related to poverty?, Journal of Policy Modelling, 30 (1): 5569. Bellany, I. (2007): Terrorism: facts from figures, Defence and Peace Economics, 18 (2): 101-112 Berrebi, C. (2007): Evidence about the link between education, poverty and terrorism among palestinians, Peace Economics, Peace Science and Public Policy, 13(1): Article 2. Berrebi, C. and E. F. Klor (2006): "On Terrorism and Electoral Outcomes: Theory and Evidence from the Israeli-Palestinian Conflict", Journal of Conflict Resolution, 50: 899-925. Bourguignon, F. and S. Chakravarty (2003): "The measurement of multidimensional poverty", Journal of Economic Inequality, 1(1): 25-49. Bruni, L. and P.L. Porta (2005): Economics and Happiness, Oxford University Press.

21

Caplan, B. (2001a): Rational ignorance versus rational irrationality, Kyklos, 54(1): 326. Caplan, B. (2001b): Rational irrationality and the microfoundations of political failure, Public Choice, 107(3/4): 311-331. Caplan, B. (2002): How does war shock the economy, Journal of International Money and Finance, 21: 145-162. Caplan, B. (2008): Terrorism: The relevance of the rational choice model, Public Choice, forthcoming. Chen, A. H. and T. F. Siems (2004): The effects of terrorism on global capital markets, European Journal of Political Economy, 20(2): 349-366. Drakos, K. And A.M. Kutan (2003): Regional effects of terrorism on tourism: three Mediterranean countries, Journal of Conflict Resolution, 47 (5): 621-641. Enders, W. and T. Sandler (1991): Causality between transnational terrorism and tourism: the case of Spain, Terrorism, 14: 49-58. Enders, W., T. Sandler and G. F. Parise (1992): An econometric analysis of the impact of terrorism on tourism, Kyklos, 45: 531-554. Enders, W. and T. Sandler (1996): Terrorism and foreign direct investment in Spain and Greece, Kyklos, 49(3): 331-352. Enders, W. and T. Sandler (2002): Patterns of transnational terrorism, 1970-99: alternative time series estimates, International Studies Quarterly, 46: 145-165. Engene, J.O. (2004): Terrorism in Western Europe: Explaining the trends since 1950, Cheltenham UK/Northampton MA, USA: Edward Elgar. Engene, J.O. (2006): Terrorism in Western Europe: Events data, University of Bergen. Epanechnikov, V.A. (1969): Nonparametric estimation of a multivariate probability density, Theory of Probability and its Applications, 14: 153-158. Frey, B. S. and A. Stutzer (2002): What can economists learn from happiness research?, Journal of Economic Literature, 40 (2): 402-435. Frey, B. S. and S. Luechinger (2005): Measuring terrorism in Law and the State, eds. A. Marciano and J-M. Josselin, pp. 142-181. Edward Elgar Publishing Ld., Cheltenham, UK. Frey, B. S., S. Luechinger and A. Stutzer (2007): Calculating tragedy: Assessing the costs of terrorism, Journal of Economic Surveys, 21(1): 1-24. Higgs, R. and A. Kilduff (1993): Public opinion: a powerful predictor of U.S. defence spending, Defence and Peace Economics, 4 (3): 227238. Krueger, A. and J. Malekov (2003): Education, poverty and terrorism: Is there a causal connection?, Journal of Economic Perspectives, 17(4): 119-44. Llorca-Vivero, R. (2008): Terrorism and international tourism: new evidence, Defence and Peace Economics, 19 (2): 169188.

22

Maasoumi, E. (1986): The measurement and decomposition of multi-dimensional inequality, Econometrica, 54, 991-997. Murdoch, J. and T. Sandler (2002): Civil wars and economic growth: a regional comparison, Defence and Peace Economics, 13 (6): 451464. Nacos, B. L. (1994): Terrorism and the Media. New York: Columbia University Press. Nadaraya, E.A. (1964): On estimating regression, Theory Probability Applied, 10: 186190. Olson, M. (1996): Big bills left on the sidewalk: Why some nations are rich and others are poor, Journal of Economic Perspectives, 10: 3-24. Prez - Fornis, C., M. D. Gadea and E. Pardos (2004): Gasto en defensa y renta en los pases de la Alianza Atlntica (1960-2002), Hacienda Pblica Espaola Revista de Economa Pblica, 170(3): 137-153. Riera, A., A. M. Ripoll and J. M. Sbert (2007): Estimacin del valor estadstico de la vida en Espaa: una aplicacin del mtodo de salarios hednicos, Hacienda Pblica Espaola Revista de Economa Pblica, 181(2): 29-48. Rohner, D. and B. Frey (2007): Blood and ink! The common-interest game between terrorists and the media, Public Choice, 127(1): 129-145. Sez, J. L. (2002): Economa y poltica en la duracin de los Gobiernos: El caso de Espaa (1977-2001), Hacienda Pblica Espaola Revista de Economa Pblica, 161(2): 69-98. Theil, H. (1967): Economics and Information Theory. Amsterdam: North Holland. Tsui, K. (2002): Multidimensional poverty Indices, Social Choice and Welfare, 19: 69-93. Viscusi, W. K. and R. J. Zeckhauser (2003): Sacrificing civil liberties to reduce terrorism risks, Journal of Risk and Uncertainty, 26: 99-120. Watson, G.S. (1964): Smooth regression analysis, Sankhya Series A, 26: 10116.

23

Appendix

Table A.1. Descriptive statistics of the dependent and explicative variables Variable Social Impact (from Factor analysis) Total killed Total injured Letter bomb Car bomb Other bomb Rocket or grenade attack Armed attack Kidnapping military target police target public service target political institutions target business target El Mundo El Pais La Vanguardia Mean 0.02127 0.68691 3.24099 0.02657 0.29791 0.29981 0.00190 0.27894 0.02657 0.10436 0.15939 0.07780 0.19355 0.03985 0.25237 0.24288 0.26186 Std. Dev. 0.93458 0.97433 11.23905 0.16096 0.45778 0.45861 0.04356 0.44890 0.16096 0.30602 0.36639 0.26811 0.39545 0.19579 0.43479 0.42923 0.44006

Table A.2. Descriptive statistics of the factor variables Variable Cover photograph % Cover % Editorials Total number of pages First interior page photograph % first interior page Second interior page photograph % Second interior page Third interior page photograph % third interior page % second day editorials Second day total number of pages Mean 0.57169 0.42719 0.14022 3.80447 0.94896 0.87600 0.74488 0.71248 0.50466 0.48678 0.15604 1.97816 Std. Dev. 0.49529 0.30491 0.15661 3.20286 0.22029 0.20650 0.43634 0.40977 0.50044 0.46370 0.28968 2.70367

24

Table A.3. Factor analysis of model 2. Eigenvalue 5.51367 1.11297 0.73747 0.45645 0.4149 0.09835 -0.02477 -0.05671 -0.08143 -0.10583 -0.12146 -0.21953 Factor1 0.6543 0.8181 0.6504 0.8001 0.2756 0.5346 0.7634 0.7888 0.8044 0.8234 0.3431 0.586 537 4769.01 0.8150 2099.245 2150.451 Difference Proportion Cumulative 4.40069 0.7138 0.7138 0.37551 0.1441 0.8579 0.28102 0.0955 0.9534 0.04154 0.0591 1.0125 0.31655 0.0537 1.0662 0.12312 0.0127 1.0789 0.03194 -0.0032 1.0757 0.02472 -0.0073 1.0684 0.0244 -0.0105 1.0578 0.01563 -0.0137 1.0441 0.09807 -0.0157 1.0284 . -0.0284 1 Uniqueness 0.5719 0.3307 0.577 0.3598 0.9241 0.7142 0.4172 0.3778 0.3529 0.322 0.8823 0.6566

Factor1 Factor2 Factor3 Factor4 Factor5 Factor6 Factor7 Factor8 Factor9 Factor10 Factor11 Factor12 Variable Cover photograph % Cover % Editorial Total number of pages First interior page photo % first interior page Second interior page photo % Second interior page Third interior page photo % third interior page % second day editorial Second day total number of pages N LR test [independent vs. saturated
chi2(66)]

Kaiser-Meyer-Olkin measure AIC (Factor1) BIC (Factor1)

25

Table A.4. Regression of model 2. Social impact (model 2) Lineal Total killed 0.24340*** (0.038) Total killed squared ----Total injured 0.01264*** (0.003) Total injured squared ----Total killed * Total injured ----Letter bomb 0.28526 (0.228) Car bomb 0.43928*** (0.138) Other bomb 0.10644 (0.152) Rocket or grenade attack -0.32302 (0.718) Armed attack 0.73931*** (0.143) Kidnapping 0.81077*** (0.261) Military target 0.19802 (0.136) Police target -0.12392 (0.106) Public service target -0.14652 (0.134) Political institutions target 0.36595*** (0.112) Business target -0.27711 (0.201) El Mundo -0.37360*** (0.087) El Pas -0.51043*** (0.088) La Vanguardia -0.68079*** (0.086) Constant -0.28539 (0.210) N 527 R2 0.5161 F 20.51

Quadratic 0.64576*** (0.089) -0.06886*** (0.016) 0.05781*** (0.009) -0.00051*** (0.000) -0.00690*** (0.002) 0.37267* (0.220) 0.41836*** (0.135) 0.0738 (0.149) -0.13739 (0.689) 0.67114*** (0.138) 0.90410*** (0.251) 0.10822 (0.131) -0.26248** (0.108) -0.0327 (0.134) 0.30206*** (0.110) -0.21882 (0.193) -0.36738*** (0.083) -0.51101*** (0.084) -0.67303*** (0.082) -0.45265** (0.205) 527 0.5579 21.625

***: Significant at the 1% level. **: Significant at the 5% level. *: Significant at the 1% level. Standard deviations in parentheses.

26

Table A.5. Number of attacks, deaths and value of index T.


Year 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 Attacks 6 4 10 10 16 6 -22 18 8 9 4 Deaths 2 6 10 5 9 5 -26 13 3 3 2 Accumulated value of T 6.4 7.3 13.9 11.2 18.1 8.9 -33.2 23.0 7.0 5.9 2.7

27

You might also like