Available online at www.sciencedirect.com
Journal of Policy Modeling 31 (2009) 788–802
Terrorism and the effectiveness of security spending in Greece: Policy implications of some empirical findings Christos Kollias a,∗ , Petros Messis b , Nikolaos Mylonidis c , Suzanna-Maria Paleologou c a
Department of Economics, University of Thessaly, 43 Korai str, 38333 Volos, Greece b Department of Business Administration, TEI of Larissa, Greece c Department of Economics, University of Ioannina, Greece
Received 16 June 2008; received in revised form 21 July 2008; accepted 15 September 2008 Available online 20 February 2009
Abstract Greece has over the years faced serious security challenges from domestic as well as transnational terrorist activity. This paper examines empirically the effectiveness of counter-terrorism policy and particularly it focuses on current and investment expenditure on domestic security and public order. Using annual budget data for the 1974–2004 period, it investigates whether current and investment spending by the Ministry of Public Order has been an effective policy measure to counter terrorism. The results seem to suggest that such investment has at best a weak negative impact on internal terrorist actions. The main policy implication of this finding is that investing in counter-terrorist infrastructure and equipment can potentially prove to be an effective policy measure in the fight against terrorism. This, however, may be conditional upon a number of other factors including other anti-terrorist measures such as legislation or how efficiently such expenditure is used. © 2009 Society for Policy Modeling. Published by Elsevier Inc. All rights reserved. Keywords: Terrorism; Greece; Count data analysis; Public order spending
1. Introduction Terrorism presents an important security challenge for most countries, the governments of which are faced with various policy options to counter terrorist threats. Indeed, as pointed out in the relevant literature, complex and multidimensional issues emerge when examining means
∗
Corresponding author. Tel.: +30 24210 74925; fax: +30 24210 74772. E-mail address:
[email protected] (C. Kollias).
0161-8938/$ – see front matter © 2009 Society for Policy Modeling. Published by Elsevier Inc. All rights reserved. doi:10.1016/j.jpolmod.2008.09.008
C. Kollias et al. / Journal of Policy Modeling 31 (2009) 788–802
789
and policies to counter and indeed prevent terrorist attacks (inter alia: Barros, 2003; Barros & Gil-Alana, 2006; Barros, Passos, & Gil-Alana, 2006; Bruck, 2005, 2007; Enders & Sandler, 2006, 2007; Franck, Hillman, & Krausz, 2005; Frey & Luechinger, 2004, 2007). One possible set of policy responses to terrorism could include tightening legislation, perhaps re-evaluating constitutional provisions, legislation changes on how to award citizenship or treat legal and illegal migration flows and even the reassessment of civil liberties in the light of the serious threat posed by terrorist organisations. Clearly, such a set of policy measures entails important dilemmas for western liberal societies that wish to protect citizens and property from transnational or domestic terrorism since state policy makers need to preserve a very delicate balance between, on the one hand, civil liberties and rights and, on the other, tighter legislation and increased powers to the security forces and their various branches (Bruck, 2005; Frey & Luechinger, 2004; Mueller, 2007). An obvious and less controversial policy option is to increase security spending on infrastructure, equipment and training that render the security apparatus more effective in countering terrorism and terrorist organisations (Enders & Sandler, 1993a; Enders, Sandler, & Cauley, 1990). Of course, without the corresponding tightening of the relevant legislation the effectiveness of such expenditure may not be maximised in reducing the scale or the effects of terrorism. This paper attempts to assess empirically the effectiveness of such government outlays and, on the basis of the findings, draw some policy implications in the fight against terrorism. In particular, we study the relationship between security spending in Greece and internal terrorist activity. An important aspect of the data analyzed herein is that the dependent variable in the relevant tests is a count1 of the total number of terrorist incidents in Greece from 1974 to 2004, i.e. a dependent variable that varies from zero to several or even many incidents per annum is used in the empirics. Thus, the paper employs statistical models of counts (non-negative integers with non-negligible probabilities of zero) in the context of time series analysis and uses them to analyze the relationship between terrorist attacks and security spending in Greece. Indeed, count data analysis is more commonly used in other fields of science2 rather than in this type of investigation. Recently, however, it has also been introduced into the investigation of issues related to terrorism (Brandt & Sandler, this issue; Lee, Enders, & Sandler, 2009). The statistical models used are applications of the Poisson distribution and the negative binomial log likelihood. The Poisson distribution is often a reasonable description for events which occur both randomly and independently over time. However, unobserved heterogeneity in economic data often violates the mean equals variance feature of the Poisson. In this case efficient estimates are obtained by assuming that the unobserved heterogeneity is gamma distributed and the count data are negative binomial. 2. Terrorism and security spending in Greece Ever since the early, path-breaking works on terrorism such as Sandler, Tschirhart, and Cauley (1983), the relevant literature on terrorism and its effects has attracted considerable attention and has steadily grown (inter alia: Drakos & Kutan, 2003; Enders & Sandler, 1993b; Enders, Sandler, & Parise, 1992; Llorca-Vivero, 2008). As already stated, the aim of this paper is to investigate the relationship between security expenditure and terrorist activity in Greece. In particular, it sets out to examine how effective such spending is as a policy response to terrorism and to test whether 1 In the context of the present study, the nature of the data dictates the use of count data models regardless of the economic theory. See also Cameron and Trivedi (2006). 2 The modelling of random counts is widespread and long established in the biometric literature. See Patil (1970) for an extensive introduction.
790
C. Kollias et al. / Journal of Policy Modeling 31 (2009) 788–802
Table 1 A summary of terrorist incidents in Greece 1974–2004.
1974–1979 1980–1989 1990–1999 2000–2004
Terrorist incidents, A
Incidents with casualties, B
B/A
Number of casualtiesa
110 276 280 61
13 67 43 4
11.8% 24.3% 15.4% 6.6%
107 607 131 8
Source: Global Terrorist Database, http://www.start.umd.edu/data/gtd/ and own calculations. a Fatalities and injuries.
the effectiveness of such government outlays can be established empirically in the case of Greece during 1974–2004. For almost three decades, following the collapse of the military dictatorship in 1974, Greece has suffered from the operation of terrorist groups. During this period a total of 727 terrorist incidents have been recorded, of which 127 are with casualties (fatalities and injuries), totalling 853 people3 (Table 1). Terrorist activity during this period has mainly but not exclusively been domestic. Pre-1974 incidents, such as bombings, are exclusively associated with the resistance against the military junta (1967–1974). Effectively, terrorism and terrorist activity enters the Greek political scene after the 1974 collapse of the dictatorship and in particular in December 1975 with the assassination of the CIA Athens station chief Richard Welch by a local terrorist organisation which, in the years that followed this first attack, became the most important terrorist group that operated in Greece. The main terrorist groups that have operated during the last thirty years are: 17th November (henceforth 17N), the Revolutionary People’s Struggle (henceforth ELA) and the Revolutionary Nuclei (henceforth RN). Of them, 17N was the main terrorist group that has operated with impunity for almost three decades. It was dismantled in 2002 after a bomb accidentally exploded prematurely in the hands of one of its members. As a result of this incident, the group was dismantled, its members arrested, brought to trial and convicted. 17N’s activity included assassinations, bombings and bank robberies to finance its operation. In terms of fatalities and victims, during the 27 years of its operation, 17N was responsible for 23 murders, of both Greek and foreign citizens, and many more injured. ELA was the second most important terrorist group that has operated during this period and considered to be an off-spring of 17N. The group was dismantled in 2003 and its members arrested and convicted. Other groups, such as RN, that have operated during this period were not so active and on the whole less “efficient” and deadly. Terrorist activity during the period under investigation is graphically presented in Fig. 1 the source of which is the Global Terrorism Database4 . Although no particular clear pattern seems to emerge from Fig. 1, it would appear that the 1980s up to the early 1990s was the period when terrorist activity peaked, especially in terms of victims, fatalities and injuries. Perhaps the characteristic that stands out when examining terrorism in Greece is that groups, such as 17N, ELA and RN, have operated for almost three decades with impunity. Indeed there is not a single successful operation by the security apparatus that one can point at, although reportedly there have been cases when the security forces have come close to apprehending members of these groups and perhaps many more that tighter security measures have prevented terrorist attacks. 3 Figures are drawn from “Global Terrorism Database, START/CETIS, accessed via Global Terrorism Database CD, June 2007 edition.” http://www.start.umd.edu/data/gtd/. 4 “Global Terrorism Database, START/CETIS, , accessed via Global Terrorism Database CD, June 2007 edition.” http://www.start.umd.edu/data/gtd/.
C. Kollias et al. / Journal of Policy Modeling 31 (2009) 788–802
791
Fig. 1. Terrorist attacks in Greece 1974–2004.
If one tries to summarise the anti-terrorist policy responses by successive Greek governments to the challenge that terrorist activity has posed during the period in question, then this has essentially taken the form of tighter legislation on the one hand and spending on security on the other. For instance, Law 2928/01 amending the Penal Code and the Code of Criminal Procedure is a recent example of tightening legislation in areas such as criminal organisations (such as terrorist groups), recruitment in such organisations, money laundering, possession of fire arms and explosives. Increased spending on security and tightening up security measures was the other policy option available to Greek governments. This was particularly the case from the late 1990s onwards due to the 2004 Olympics in Athens. Such mega-events offer a target-rich environment for terrorists. Indeed, given the fact that the 2004 Olympics were the first Olympic Games to be organised after 9/11, unprecedented security measures were taken. Reportedly, Greece invested over $1.5 billion on security for the Games. This included spending on training, infrastructure (including infrastructure against nuclear, chemical and biological attacks), surveillance systems (an estimated D 255 million was spend on an advanced C4I system for the host city of Athens), intelligence networks etc while during the Games around 40,000 police and 10,000 troops were deployed. The government body in charge of domestic security in Greece and therefore of the security forces, such as the police, intelligence and antiterrorist units, that are responsible for counterterrorism is the Ministry of Public Order (henceforth MPO). Its outlays, as a share of total government expenditure during the period in question, are graphically depicted in Fig. 2. From the visual inspection of the data it appears that it follows a downward path up to the end of the 1980s with the exception of two years in the middle of this decade. This spike in 1984 and 1985 is probably due to the major reorganisation of the security forces that took place. Specifically, up to then, internal security was the responsibility of the gendarmerie in the regions and rural areas and the police in the major cities. With Law 1481, the gendarmerie was dismantled and incorporated into the police forming a unified national police force. The spike probably reflects the costs caused by this integration and merging of the two forces. Thus, excluding this spike, the overall trend in the 1980s is downward. Then, the series in question, changes into an upward trend roughly from the start of the 1990s. On average, during 1974–2004, MPO expenditure accounted for about 3.4% of total budgetary spending. Although terrorism is by no means the single determinant of this expenditure, the series used here can be regarded as a satisfactory proxy for antiterrorist spending. Having said this, it is also safe to assume that antiterrorist policy and the concomitant costs have been an important factor shaping the ministry’s yearly budget. This is probably more
792
C. Kollias et al. / Journal of Policy Modeling 31 (2009) 788–802
Fig. 2. MPO spending as share of total government expenditure.
the case from the mid 1990s onwards due to the undertaking of the organisation of the 2004 Olympics. Indeed, the pick towards the end of this period may almost exclusively be attributed to the security costs of organising this international mega-event as well as the emergence, following 9/11, of transnational terrorism as a major security challenge and threat. 3. Econometric methodology 3.1. Basic count data regression The basic nonlinear5 estimation used with count data is the Poisson regression6 . The Poisson distribution seems a natural assumption when dealing with counting problems in econometrics7 . The Poisson distribution has a density: Pr[Y = y] =
e−μ μy , y!
y = 0, 1, 2, . . . ,
(1)
where μ is the rate parameter. The distribution is referred as P[μ] and the first two moments are E[Y ] = μ, Var[Y ] = μ
(2)
which show the equidispersion (equality of mean and variance) property of the Poisson distribution. By introducing the observation subscript t, attached to both y and μ, the iid framework is extended to the regression case. The Poisson regression model is derived from the Poisson distribution by parameterizing the relation between the mean parameter μ and the regressors X. If we incorporate a set of exogenous variables Xtj (j = 1,. . .,K) including a constant, the mean parameter μt is specified as μt = exp(Xtj β),
t = 1, 2, . . . , T
(3)
5 Many count data models are non-linear, though they can be viewed as “generalised linear models” (Nelder & Wedderburn, 1972). 6 For a detailed exposition of the model see Cameron and Trivedi (2006, 1986). 7 See Hausman, Hall, & Griliches (1984).
C. Kollias et al. / Journal of Policy Modeling 31 (2009) 788–802
793
According to Hausman et al. (1984) and Cameron and Trivedi (2006) the advantages and disadvantages of the Poisson specification can be classified as follows. First, the Poisson specification in many ways is analogous to the familiar econometric regression specification (E(yt |Xtj ) = μt ). Second, the zero problem (i.e. yt = 0) is a natural outcome of the Poisson specification. In contrast to the usual logarithmic regression specification, we need not truncate an arbitrary continuous distribution. Likewise the integer property of the outcomes yt is handled directly. In large yt a continuous approximation is often sufficient. But for small yt , as in our case, a specification which models the counting properties of the data seems in order. The shortcomings of the Poisson model are basically that of overdispersion and insufficiency to explain all zeros in the sample. The latter shortcoming should not be a problem in our case since the number of zeros represents only 3% of the sample. As far as concerned the problem of overdispersion, one must not be too hasty and reject the Poisson model since in many cases overdispersion is caused by the distribution of the independent variables.8 3.2. Parametric count regression One way to account for the problem of overdispersion is to use Negative Binomial Models, henceforth referred as Negbin. The Negbin can be obtained by using a mixture distribution. Lets assume that the Poisson parameter μt is distributed randomly so that f (yt |λt ) = exp(−λt )λyt t /yt !. Suppose that the parameter λt is random, rather than being a completely deterministic function of regressor Xtj . In particular, let λt = μt νt , where μt is a deterministic function of Xtj and vt > 0 is iid with density g(vt |α).9 E[λt |μt ] = μt , if E[vt ] = 1, so that the interpretation of the slope parameters stays as in the Poisson model. The marginal density of yt , unconditional on the random parameter vt but conditional on the deterministic parameters μt and α, is obtained by integrating out vt : (4) h(yt |μt , α) = f (yt |μt , νt )g(νt |α)dvt where g(νt |α) is called the mixing distribution and α denotes the unknown parameter of the mixing distribution. The integration defines an average distribution. If f(yt |λt ) is the Poisson density and g(νt ) = νt δ−1 e−vδ δt δ /Γ (δ), vt , δ > 0 is the gamma density with E[vt ] = 1 and Var[vt ] = 1/δ, then we obtain Negbin as a mixture density:10 ∞ −μt νt (μt νt )yt vt δ−1 e−νt δ δδ e h[yt |μt , δ] = (5) dνt yt ! Γ (δ) 0 α−1 yt μt Γ (α−1 + yt ) α−1 = μt + α−1 Γ (α−1 )Γ (yt + 1) α−1 + μt where α = 1/δ, Γ (·) denotes the gamma integral. In this case the first two moments are E[yt |μt , α] = μt 8
See Crepon and Duguet (1997). This is a case of unobserved heterogeneity, since different observations may have different λt (heterogeneity) but part of this difference is due to a random (unobserved) component νt . 10 According to Cameron and Trivedi (2006), Negbin can arise in many different ways and one should not always consider it as a mixture distribution. 9
794
C. Kollias et al. / Journal of Policy Modeling 31 (2009) 788–802
Var[yt |μt , α] = μt (1 + αμt )
(6)
It is clear from Eq. (6) that the variance exceeds the mean (α > 0 and μt > 0). Overdispersion always arises if yt |λt is Poisson and unobserved heterogeneity is of the multiplicative form λt = μt νt , where E[νt ] = 1. Of course, we should be careful when rejecting the Poisson assumption in favour of Negbin because poor performance of the Poisson model can be due to poor specification of the conditional mean function. In this case, the conditional mean function remains the same even under Negbin. Furthermore, Negbin is less robust to distributional misspecification compared to the Poisson model. 4. Empirical analysis 4.1. The data In the empirics that follow, the dependent variable (terr) is the total number of internal terrorist attacks in Greece, over the sample period 1974–2004. As already noted earlier, the series on terrorism incidents are drawn from the Global Terrorism Database11 . In the estimates presented further down the total number of terrorist incidents was used as the dependent variable12 . Two different data sources are employed to measure the total amount of security expenditures in Greece. The first source is expenditure that originates from the MPO’s Ordinary Budget (po) and the second source of expenditure originates from the MPO’s Investment Budget (poi)13 . Also, a set of control variables is used in the tests, such as GDP per capita and total budget expenditures (excluding public order spending) taken from the ordinary and the investment budget, respectively. Finally, a number of country-specific dummy variables were included in the estimations. These consist of a scale dummy to account for the effects of the 2004 Olympics in Athens; a dummy variable that attempts to capture possible effects that the political colour of the incumbent party – centre right or centre left – has on terrorism; and finally a dummy for the dismantling of 17N in 2002. A description of the data used and the variables introduced in the estimations is given in Table 2. 4.2. Empirical specification and results The simple model developed in this section attempts to assess the effectiveness (if any) of security spending on the fight against terrorism. Our empirical approach can be summarised as follows: terrt = α + βPublicOrdert−1 +
k i=1
11
γi Yi,t−1 +
k
δ i Zi + ε t
(7)
i=1
http://www.start.umd.edu/data/gtd/. As suggested by Enders and Sandler (2000, 2002), the regressions, reported further down, were also estimated using the casualties series (fatalities and injuries) instead of the total number of terrorist incidents as the dependent variable but they did not yield any significant results (not reported here). 13 Since 1947 there are two types of budget in Greece: (i) the Ordinary Budget, which mainly contains current expenditures per Ministry, and (ii) the Investment Budget, which in principle contains investment (infrastructure, technology and capital equipment) expenditures per Ministry. This was done because the Greek government officials at that time wanted to make sure that the American Aid, given after Second World War, would be used exclusively for infrastructure development and not spent anywhere else. 12
C. Kollias et al. / Journal of Policy Modeling 31 (2009) 788–802
795
Table 2 Data definitions, sources and basic descriptive statistics. Variable
Definition
Source
Mean
Std. Dev.
terr po
Total number of internal terrorist incidents in Greece Ordinary budget expenditure by the Ministry of Public Order in current prices, logged Investment budget expenditure by the Ministry of Public Order in current prices, logged Nominal GDP per capita, logged Total ordinary budget expenditure (excluding public order expenditures) in current prices, logged Total investment budget expenditure (excluding public order investments) in current prices, logged Scale dummy—2004 Olympics in Athens Political colour dummy (Conservative government = 0; Social-democratic government = 1) Dummy variable—dismantling of 17N (1 in the post-2002 period; 0 elsewhere)
GTB HGAO
23.45 5.51
14.72 1.41
HGAO
1.38
1.61
IFS HGAO
0.95 8.71
1.35 1.63
HGAO
7.00
1.51
poi gdpc tg tgi DO DPOL D17N
Notes: GTB stands for the Global Terrorism Database. HGAO stands for Hellenic General Accounting Office. IFS stands for International Financial Statistics.
where we regress the annual number of internal terrorist incidents with casualties in Greece on a constant, our measure of public order spending (po or poi), a set of control variables Y that might affect the terrorist activities and a set of country-specific dummies Z, with ε being a well-behaved residual. Should public order spending be effective in the fight against terrorism, then po (or poi) is expected to enter the equation with a negative sign. The control variables include economic development (GDP per capita) and total government expenditures (excluding public order). Given that poor economic conditions are often regarded as terrorist breeding grounds (Li & Schaub, 2004); we expect the level of economic development to be negatively related to the number of domestic terrorist incidents. The inclusion of total government expenditures (tg or tgi) in the regression accounts for all these components of security spending not included in po (or poi) (e.g. expenditures undertaken by the Ministries of Defence, Justice, Maritime Affairs, etc.). The hypothesized effect of tg (tgi) on terrorist attacks involves two confounding issues. On one hand, enhanced overall security reduces the ability of terrorist groups to operate, prevents the successful implementation of their plans and increases the probability of punishment. On the other hand, higher overall security may result in backlash attacks against what would be regarded by terrorist groups as authoritarian behaviour by the state and the state apparatus, i.e. a security crack down or even simply an overly harsh government. To the extent that this is the case, the net effect that total government spending has on terrorism is, therefore, ambiguous. Both po (poi) and the set of control variables are in current prices since there is not a separate price deflator for public order and total government expenditures in Greece. Most likely, terrorist attacks casualty series reacts to these explanatory variables only with a lag because the process of choosing and implementing security investments is time consuming and most importantly the effects of such spending become evident with a time lag. Finally, terrorist incidents may affect most of the right-hand variables (GDP per capita, public order and total government spending), thereby introducing simultaneity bias if Eq. (7) is to be estimated in a contemporaneous form. In order to circumvent the aforementioned problems, we estimate Eq. (7) with one-year lag for the independent variables.
796
C. Kollias et al. / Journal of Policy Modeling 31 (2009) 788–802
Table 3 Poisson regression results—ordinary budget data. Variable
1
2
3
4
5
constant
1.780 (1.099) 0.317 (0.242) −0.184 (0.240)
0.234 (1.135) 0.703*** (0.253) −0.490** (0.247)
−0.233*** (0.033)
−0.242*** (0.033) −0.424*** (0.091)
−23.912*** (3.829) 0.635** (0.248) −4.510*** (0.667) 3.192*** (0.446) −0.053 (0.042)
−23.076*** (3.869) 0.874*** (0.257) −4.407*** (0.674) 2.959*** (0.459) −0.069 (0.043) −0.293*** (0.089)
−24.666*** (3.882) 0.600** (0.252) −4.643*** (0.677) 3.312*** (0.452) 0.055 (0.058)
po(−1) gpdc(−1) tg(−1) DO DPOL
−0.860*** (0.325)
D17N
Log-likelihood Pseudo R2 t-test η2
−185.562 0.144 4.597*** 0.234
−174.971 0.193 4.039*** 0.188
−159.181 0.266 2.922*** 0.139
−153.828 0.290 2.634** 0.117
−155.664 0.282 3.094*** 0.132
Notes: Standard errors in parentheses. DO is a dummy variable accounting for the organisation of the 2004 Olympic Games in Athens (see text for further details). DPOL is a political colour dummy variable (conservative = 0; social democratic = 1) referring to which party is in power. D17N is a dummy variable which takes the value of 1 in the post-2002 period and accounts for the dismantling of 17N. *** Significant at 1%. ** Significant at 5%. * Significant at 10%.
Vector Z includes three country-specific dummy variables. The dummy labelled DO assesses the possible impact of the 2004 Olympics on the number of terrorist attacks in Greece. Countries undertaking the organisation of such mega-events that represent a target rich environment for terrorists are forced to heavily invest in tightening security against domestic or foreign terrorist threats. This variable is a scale dummy that escalates from 1 to 7. It takes the value of zero for the period 1974–1997 and the value of 1 g (where g = 0,1,2,. . .,7) for the period 1998–2004 (1998 was the year that Greece officially undertook the organisation of the 2004 Olympic Games)14 . Finally, DPOL is a political colour dummy that takes the value of zero if the political party in power was conservative and the value of one if it was social-democratic. D17N takes into consideration the dismantling of the most prominent and active terrorist group in Greece – i.e. 17N – in 2002. Because terr is event count, OLS estimates are inefficient, inconsistent and biased (Long, 1997). Therefore, we need to apply alternative techniques (Brandt & Sandler, this issue; Lee et al., 2009). The first choice for modelling count data is the Poisson regression model. We begin our investigation by estimating Eq. (7), using different subsets of Y and Z. The Poisson regression results are reported in Tables 3 and 4. Specifically, Table 3 presents the results when ordinary budget data are used, whereas Table 4 reports the corresponding results when investment budget data are used. In a first exercise, we enter a parsimonious set of control and dummy variables as
14
Kollias and Paleologou (2003) use a similar scale dummy.
C. Kollias et al. / Journal of Policy Modeling 31 (2009) 788–802
797
Table 4 Poisson regression results—investment budget data. Variable
1
2
3
4
5
constant
3.250*** (0.049) −0.093* (0.053) 0.192*** (0.049)
3.380*** (0.058) 0.041 (0.063) 0.170*** (0.048)
−0.190*** (0.034)
−0.218*** (0.034) −0.392*** (0.109)
−6.052*** (1.791) −0.125** (0.053) −1.382*** (0.306) 1.565*** (0.301) −0.336*** (0.047)
−7.914*** (1.909) 0.035 (0.061) −1.745*** (0.326) 1.907*** (0.322) −0.404*** (0.050) −0.515*** (0.110)
−6.071*** (1.956) −0.125** (0.053) −1.385*** (0.331) 1.568*** (0.328) −0.337*** (0.068)
poi(−1) gpdc(−1) tgi(−1) DO DPOL D17N Log-likelihood Pseudo R2 t-test η2
−184.819 0.148 4.972*** 0.227
−178.315 0.178 4.310*** 0.200
−171.064 0.211 3.484*** 0.189
−160.075 0.262 3.409*** 0.153
0.008 (0.349) −171.064 0.211 3.483*** 0.189
Notes: See Table 3. *** Significant at 1%. ** Significant at 5%. * Significant at 10%.
explanatory variables for the internal terrorist incidents in Greece. In the subsequent columns of Tables 3 and 4 (columns 2 to 5), we explore the impact of augmented alternative set of regressors on terrorist activity. We first focus on our measure of public order spending. Although both po and poi are statistically significant in most instances, they take opposite in sign values. Specifically, poi is negative, thus suggesting effectiveness of counter-terrorism public order spending, whereas po is positive, thereby demonstrating ineffectiveness. Nevertheless, this result is not all together counterintuitive given that the ordinary budget (po) contains spending on such things as salaries, transfer payments to personnel, operating costs etc. In contrast, the investment budget (poi) in principle and by definition contains investment spending. Such outlays probably include spending on such things as infrastructure and capital equipment (for example metal detectors for airports and buildings, surveillance cameras, C4I systems etc). It is reasonable to assume that investing in such areas increases counterterrorist effectiveness (Enders & Sandler, 1993a; Enders et al., 1990). This offers a satisfactory explanation both for the sign and the statistical importance of this particular variable. It follows, that a policy implication of this finding is that investing in antiterrorist infrastructure and capital can prove an effective terrorist thwarting measure. We now turn to the impact of the control variables and the country-specific dummies on the number of internal terrorist activities in Greece. First, it is evident that in most instances the level of economic development, as measured by nominal GDP per capita, reduces the number of terrorist attacks. This is consistent with the findings of Li and Schaub (2004) and Li (2005). In contrast, the effect of total budget spending (tg and tgi) is uniformly positive and statistically significant. This finding could be interpreted as indicating that government activities which enhance overall
798
C. Kollias et al. / Journal of Policy Modeling 31 (2009) 788–802
security, such as the provision of national defence, may induce terrorists to (re)act against what they probably regard as a hardening authoritarian state and overly harsh state behaviour. The 2004 Olympics dummy variable is statistically significant and negative especially when investment budget data are used (Table 4), thereby substantiating the argument that the undertaking of international mega-events has a mitigating effect on terrorism due to the heavy security associated with such events as well as due to the increased international coordination (such as for example exchange of intelligence, know-how etc between countries) to prevent terrorist incidents. In the case of the 2004 Athens Olympics international security coordination was particularly strong (Migdalovitz, 2004). In 2000 the Greek government established a seven nation Olympic Advisory Group made by the USA, the UK, Germany, Israel, Australia, France and Spain. Its members actively participated in training Greek security forces to deal with potential threats from transnational terrorism while intelligence exchange also took place. Other governments, notably that of Russia also provided assistance by sending mobile laboratories for chemical, biological and nuclear attacks while Greece also requested assistance from NATO in the form of AWACS planes for air policing and naval vessels to police extraterritorial waters around Greece. Assistance was also provided by the UN International Atomic Energy Agency against radiological dispersion devices15 . Finally, it is worth mentioning the uniformly negative and statistically significant coefficient of DPOL which seems to be indicating that terrorist activity is possibly influenced by the political colour of the incumbent party. It is possible that a conservative party and its security policy is perceived by terrorist groups to be harder and more authoritarian whereas that of a social-democratic government more liberal. If this is the case, it implies adjustments in the behaviour of terrorists responding to the policy of the party in power. Alternatively, this finding may very well be indicating a more successful antiterrorist policy by the latter, such as tighter security and protection of potential targets, which effectively deters terrorist activity. However, despite its prevalence as a starting point in the analysis of count data, the Poisson specification is often unsuitable. A primary reason for Poisson’s inadequacy as a regression model stems from the equidispersion feature of the Poisson distribution, i.e. the conditional mean of y equals its conditional variance. Unobserved heterogeneity in economic data often causes the variance to exceed the mean resulting in a phenomenon called overdispersion. Wooldridge (2002) proposes a regression based test of the Poisson equidispersion restriction. This is a t-test from a univariate auxiliary regression of the standardised residuals on the predicted values from the Poisson estimation. The results of this test (reported in the penultimate line of Tables 3 and 4) provide evidence of overdispersion. The extent to which the conditional variance exceeds the conditional mean (η2 ) varies from 0.117 to 0.234. Using the values of η2 from the Wooldridge tests, we proceed to estimating regressions based on the negative binomial distribution (Negbin). The Negbin regression results for Eq. (7) are presented in Tables 5 and 6 and show a strong improvement in fit over the Poisson in terms of maximising both the log-likelihood and pseudo R2 . As commonly observed in the presence of overdispersion, the estimated standard errors of the Poisson model are below their Negbin counterparts. Inferences regarding the control variables and the 2004 Olympics dummy variable remain the same as in standard Poisson. The political colour dummy variable (DPOL) retains its sign but becomes statistically insignificant in most instances. We now turn to the main question of this investigation: Has public order spending been effective in fighting terrorism in Greece? Based on Negbin coefficients the answer is ambiguous. Both po and poi are no longer significantly
15
See CRS Report for Congress, RS21833, 30 April, 2004 by Migdalovitz (2004).
C. Kollias et al. / Journal of Policy Modeling 31 (2009) 788–802
799
Table 5 Negbin regression results—ordinary budget data. Variable
1
2
3
4
5
constant
1.834 (3.078) 0.305 (0.682) (0.137 (0.688)
0.327 (3.246) 0.677 (0.728) (0.446 (0.716)
(0.276*** (0.080)
(0.281*** (0.081) (0.375 (0.258)
−23.456** (9.982) 0.643 (0.655) (4.420** (1.733) 3.129*** (1.193) (0.087 (0.102)
(22.406** (10.013) 0.886 (0.699) (4.279** (1.737) 2.862** (1.211) (0.107 (0.104) (0.275 (0.258)
(24.252** (9.769) 0.608** (0.637) (4.565*** (1.696) 3.253*** (1.171) 0.023 (0.143)
po(−1) gpdc(−1) tg(−1) DO DPOL D17N Log-likelihood Pseudo R2
(119.641 0.372
(119.548 0.499
(118.531 0.652
(118.811 0.720
(0.736 (0.683) (118.003 0.676
Notes: GLM robust standard errors corrected for overdispersion. *** Significant at 1%. ** Significant at 5%. * Significant at 10%.
different from zero at conventional significance levels. Clearly, this finding casts doubts on the argument that security spending, at least in the case of Greece, proved to be an effective antiterrorist policy instrument. Overall, our findings provide mixed evidence on the impact of security spending on terrorism. The outcomes depend on the data and methodology used. For instance, using investment budget Table 6 Negbin regression results—investment budget data. Variable
1
2
3
4
5
constant
3.231***
3.347***
(0.133) (0.048 (0.133) 0.197 (0.130)
(0.167) 0.048 (0.157) 0.188 (0.131)
(0.246*** (0.084)
(0.264*** (0.085) (0.329 (0.276)
−6.464 (4.689) (0.096 (0.136) (1.458* (0.810) 1.632** (0.789) (0.366*** (0.103)
(7.541*
(6.961 (5.391) (0.096 (0.140) (1.539* (0.921) 1.716* (0.908) (0.396** (0.184)
poi(−1) gpdc(−1) tgi(−1) DO DPOL
(4.423) 0.041 (0.155) (1.688** (0.771) 1.840** (0.747) (0.408*** (0.102) (0.462* (0.274)
D17N Log-likelihood Pseudo R2 Notes: See Table 5.
(119.826 0.389
(119.698 0.463
(118.126 0.502
(117.429 0.612
0.168 (0.834) (118.098 0.502
800
C. Kollias et al. / Journal of Policy Modeling 31 (2009) 788–802
data and a standard Poisson regression, the results indicate a significant negative effect of lagged public order investments on terrorist attacks. The concomitant policy implication of this finding being that such investment can prove an effective weapon against terrorists. In the corresponding Negbin regression, this negative effect becomes insignificant. Based on statistical criteria, and given the evidence of overdispersion, one should probably put more emphasis on the Negbin regressions and their results. However, as Crepon and Duguet (1997) suggest, overdispersion does not always invalidate the Poisson model especially when this is caused by the distribution of the independent variables. Therefore, on the basis of this observation, we may tentatively conclude that our findings, albeit mixed, indicate a relative effectiveness of counter terrorism public order investments. 5. Concluding remarks and some policy implications For almost three decades Greece has faced many internal terrorist attacks and has suffered from the operation of domestic terrorist groups such as 17N and ELA. Their operation has claimed the lives of many people – Greek as well as foreign citizens – and caused injuries to hundreds more. Anti-terrorist policy responses by successive Greek governments have essentially taken the form of tighter legislation on the one hand and spending on security on the other. The paper set out to examine empirically the effectiveness of security spending as a policy option against terrorism. It focused on investment expenditure by the Ministry of Public Order, the government body in change of domestic security and antiterrorism. Such spending improves the capacity and the effectiveness of the security forces in the fight against terrorism. The findings reported herein, albeit mixed and dependent upon the econometric methodology employed, seem to indicate that spending on infrastructure, technology and capital equipment can potentially prove to be an effective counterterrorist policy option. Of course, other low or noncost policy measures such as legislation can also be effective weapons in the fight against terrorism and indeed maximise the effectiveness of such investment expenditure as the one examined here. Nevertheless, in line with the findings of previous studies, the main policy implication of the results reported herein is that investing in antiterrorist infrastructure and capital can prove an effective terrorist thwarting measure. Spending on security may be seen as an investment. It may have short run returns but it may also very well yield fruits in the long run. A lot depends not only in the amounts invested but also on how efficiently such resources are allocated and indeed used. In any case, the fight against terrorism requires a consistent, coordinated and enduring effort that also focuses on alleviating the generating causes of this problem. Acknowledgments The paper has greatly benefited from the insightful comments and constructive suggestions of an anonymous referee. We are grateful to him/her for spotting flaws and shortcomings. The authors also wish to thank participants of the 2008 Lisbon Conference on Defence and Security for their valuable input on a previous version of the paper. The usual disclaimer applies. References Barros, C. P. (2003). An intervention analysis of terrorism: The Spanish ETA case. Defence and Peace Economics, 14(6), 401–412.
C. Kollias et al. / Journal of Policy Modeling 31 (2009) 788–802
801
Barros, C. P., & Gil-Alana, L. (2006). ETA: A persistent phenomenon. Defence and Peace Economics, 17(2), 95–116. Barros, C. P., Passos, J., & Gil-Alana, L. (2006). The timing of ETA terrorist attacks. Journal of Policy Modeling, 28(3), 335–346. Brandt, P., & Sandler, T. (this issue). Hostage taking: Understanding terrorism event dynamics. Journal of Policy Modeling. Bruck, T. (2005). An economic analysis of security policies. Defence and Peace Economics, 16(1), 375–389. Bruck, T. (Ed.). (2007). The economic analysis of terrorism. Routledge. Cameron, C., & Trivedi, P. (1986). Econometric models based on count data: Comparisons and applications of some estimators and tests. Journal of Applied Econometrics, 1(1), 29–53. Cameron, C., & Trivedi, P. (2006). Microeconometrics: Methods and applications. Cambridge: Cambridge University Press. Crepon, B., & Duguet, E. (1997). Research and development, competition and innovation pseudo-maximum likelihood and simulated maximum likelihood methods applied to count data models with heterogeneity. Journal of Econometrics, 79, 355–378. Drakos, K., & Kutan, A. L. (2003). Regional effects of terrorism on tourism in three Mediterranean countries. Journal of Conflict Resolution, 47(5), 621–641. Enders, W., & Sandler, T. (1993a). The effectiveness of anti-terrorism policies: Vector-autoregression intervention analysis. American Political Science Review, 87(4), 829–844. Enders, W., & Sandler, T. (1993b). Terrorism and foreign direct investment in Spain and Greece. Kyklos, 49(3), 331–352. Enders, W., & Sandler, T. (2000). Is transnational terrorism becoming more threatening? A time series investigation. Journal of Conflict Resolution, 44(3), 307–332. Enders, W., & Sandler, T. (2002). Patterns of transnational terrorism, 1970–1999: Alternative time series estimates. International Studies Quarterly, 46(2), 145–165. Enders, W., & Sandler, T. (2006). The political economy of terrorism. Cambridge: Cambridge University Press. Enders, W., & Sandler, T. (2007). An economic perspective on transnational terrorism. In T. Brück (Ed.), The economic analysis of terrorism. Routledge Studies in Defence and Peace Economics, pp. 13–28. Enders, W., Sandler, T., & Cauley, J. (1990). Assessing the impact of terrorist-thwarting policies: An intervention time series approach. Defence Economics, 2(1), 1–18. Enders, W., Sandler, T., & Parise, G. (1992). An econometric analysis of the impact of terrorism on tourism. Kyklos, 45(4), 531–554. Franck, R., Hillman, A. L., & Krausz, M. (2005). Public safety and the moral dilemma in the defense against terror. Defence and Peace Economics, 16(5), 347–364. Frey, B. S., & Luechinger, S. (2004). How to fight terrorism: Alternatives to deterrence. Defence and Peace Economics, 14(4), 237–249. Frey, B. S.,& Luechinger, S. (2007). Decentralization as a response to terror. In T. Brück (Ed.), The economic analysis of terrorism. Routledge Studies in Defence and Peace Economics, pp. 224–230. Hausman, J., Hall, B. H., & Griliches, Z. (1984). Econometric models for count data with an application to the patents–R & D relationship. Econometrica, 52, 909–938. Kollias, C., & Paleologou, S. M. (2003). Domestic political and external security determinants of the demand for Greek military expenditure. Defence and Peace Economics, 14(6), 437–445. Lee, B. S., Enders, W., & Sandler, T. (2009). 9/11: What did we know and when did we know it. Defence and Peace Economics (in press). Li, Q. (2005). Does democracy promote or reduce transnational terrorist incidents? Journal of Conflict Resolution, 49(2), 278–297. Li, Q., & Schaub, D. (2004). Economic globalization and democracy: An empirical analysis. Journal of Conflict Resolution, 48(2), 230–258. Llorca-Vivero, R. (2008). Terrorism and international tourism: New evidence. Defence and Peace Economics, 19(2), 169–188. Long, J. S. (1997). Regression models for categorical and limited dependent variables. Thousand Oaks, CA: Sage. Migdalovitz, C. (2004). Greece: Threat of terrorism and security at the Olympics, CRS Report for Congress, RS21833. Mueller, D. (2007). Rights and citizenship in a world of global terrorism. In T. Brück (Ed.), The economic analysis of terrorism. Routledge Studies in Defence and Peace Economics, pp. 209–223. Nelder, J. A., & Wedderburn, R. W. (1972). Generalized linear models. Journal of the Royal Statistical Society, Series B, 135, 370–384.
802
C. Kollias et al. / Journal of Policy Modeling 31 (2009) 788–802
Patil, G. P. (1970). Random counts in models and structures Pennsylvania: The Pennsylvania State University Press. Sandler, T., Tschirhart, J. T., & Cauley, J. (1983). A theoretical analysis of transnational terrorism. American Political Science Review, 77, 36–54. Wooldridge, J. M. (2002). Econometric analysis of cross section and panel data. The MIT Press.