The Roles of Foreign Aid and Education in the War on Terror

Transcription

1 October 12, 2007 The Roles of Foregn Ad and Educaton n the War on Terror by Jean-Paul Azam a and Véronque Thelen b Forthcomng n Publc Choce a: Toulouse School of Economcs (ARQADE & IDEI) and Insttut Unverstare de France, b : Toulouse School of Economcs (ARQADE). Abstract: Ths paper presents a theoretcal framework and some emprcal results showng that the level of foregn ad receved reduces the supply of terrorst attacks by recpent countres, as does the recpent country s level of educaton. Due account s taken of endogenety problems n producng these results. They suggest that Western democraces, whch are the man targets of terrorst attacks, should nvest more funds n foregn ad wth a specal emphass on supportng educaton. Keywords: Terrorsm. Foregn ad. Educaton Correspondng Author: Professor Jean-Paul Azam, ARQADE, Unversty of Toulouse 1, 21 Allée de Brenne, Toulouse, France. E-mal : azam@unv-tlse1.fr

2 1 1. Introducton In the last few decades, terrorsm has manly nvolved attacks on economc or poltcal nterests of rch and democratc countres perpetrated by natonals from developng countres (Krueger and Latn, 2003). 1 Terrorsm thus has an mportant nternatonal dmenson, as most ncdents relate a target country to a dfferent source country, and that relatonshp s typcally asymmetrcal. The two countres nvolved have vastly dfferent levels of economc and poltcal development. Often, the terrorst attack takes place n yet another host country, whch Drakos and Gofas (2006) call the venue. Therefore, the target country s not necessarly n a poston to defend ts nterests drectly, and must delegate part of ts antterrorst actvty to one or several governments, unless t delberately nfrnges on the latter s soveregnty. Ths nter-governmental relatonshp has generally been neglected n the lterature devoted to the fght aganst terrorsm. The latter has manly addressed the ssues of self-protecton aganst terrorst attacks (see Enders and Sandler, 2006) and of crackdown. One excepton s Azam and Delacrox (2006), who suggest that foregn ad s mportant for nducng local governments to fght terrorsm wthn ther sphere of nfluence. They frst brng out a surprsng emprcal relatonshp where the ad flow receved by a country s postvely correlated wth the number of terrorst attacks flowng from t, after controllng only for the recpent country s level of ncome per capta and populaton sze. Furthermore, they provde a smple theoretcal analyss and an econometrc test showng that the latter relatonshp s not a structural equaton but an equlbrum locus relatng two endogenous varables. Ths result and ts theoretcal underpnnngs suggest that donor countres are actvely usng foregn ad as a tool n the war on terror. Wthn that framework, the donor communty s allocatng ad 1 Kurrld-Kltgaard et al. (2006) confrm that democraces tend to be the target of terrorst attacks, wth a quadratc effect reachng a maxmum at very hgh levels of democracy. Beyond that pont, more democracy reduces the number of attacks. They also show that democracy reduces the probablty that a terrorst attack orgnates n any gven country. By contrast, Basuchoudhary and Shughart (2007) show that poltcal rghts are not consstently sgnfcant n ther sample, whle the lack of economc freedom s a key characterstc of source countres.

3 2 across countres to provde (among other thngs) stronger ncentves for fghtng terrorsm to governments facng more mltant groups. They then show that a few dummy varables are suffcent for controllng somewhat for mltancy and producng a structural equaton ncludng a negatve mpact of ad on the supply of attacks by source countres. The present paper presents a more satsfactory structural-form equaton, wth a rcher theoretcal framework, explanng better the role of ad n the fght aganst terrorsm. It presents some stronger estmates for the role of ad and brngs out the mpact of secondary school enrollment n reducng the number of terrorst attacks by country of orgn. Ths debate about the role of foregn ad n the war on terror emerged recently at the hghest poltcal level. George W. Bush advocated ts use as a tool aganst terrorsm n a much cted speech n Monterrey on March 22, He then sad: We fght aganst poverty because hope s an answer to terror (cted n Krueger and Maleckova, 2003, p.119). Ths dea was echoed n varous quarters of the U.S. admnstraton, as well as n academa, and some of ts mplcatons were drawn for strengthenng ad polcy. The Mllennum Challenge Account was created as a new tool for channelng ad to poor countres n the wake of that speech. Alan Krueger crtczed ths proposton n an nfluental op-ed publshed n the New York Tmes (Krueger, 2003). The bottom lne of hs crtcsm s that poverty does not seem to be the man determnant behnd terrorst attacks. The survey data presented n Krueger and Maleckova (2003), as well as n Krueger and Latn (2003), show n fact that terrorsts from dfferent movements, ncludng the Hezbollah, are predomnantly recruted from a relatvely wealthy and educated famly background. Smlarly, the bographes of Al-Qaeda s actvsts analyzed by Sageman (2004) show that they generally have a hgh level of educaton, mostly n scentfc or techncal dscplnes. Hassan (2001) descrbes a sample of Hamas volunteers wth a smlar background. Moreover, she shows that they are gven some ntensve relgous tranng after havng enrolled for a sucde msson. Some other nsghts on the profles of terrorsts can be gleaned from Bloom (2005), Reuter (2004) and Stern (2003). The emergng

4 3 pcture s that terrorsts are men and women n ther twentes wth some post-secondary tranng, mostly n techncal or engneerng educaton. Hence, the mcroeconomc evdence refutes a smple vew that poverty breeds terrorsm, because terrorsts are not recruted among the poorest segments of ther socety of orgn. A smple readng of these results would suggest nstead that wealth and educaton exert a postve nfluence on the decson to engage n terrorst attacks. Smlarly, Paz (2000) documents the fact that hgher educaton expanded throughout the West Bank and Gaza at the tme when the Hamas and Islamc Jhad were ganng popular support. However, the supply of sklled jobs was not expandng n proporton, rasng some doubt about the sgnfcance of ths co-movement. Nevertheless, ths debate s bearng on a fundamental aspect of the relatonshps between the North and the South. It rases the ssue of the contnuaton of foregn ad, nsofar as t s supposed to support educaton and contrbute to growth, whle the global fght aganst terrorsm s domnatng nternatonal relatons. The present paper ams at contrbutng some addtonal arguments n ths debate, usng both theoretcal and emprcal analyss. Presdent Bush advanced a dfferent lne of defense for hs vews about poverty and terrorsm n September In a New York Tmes op-ed publshed on September 11 he wrote: Poverty does not transform poor people nto terrorsts and murderers. Yet, poverty, corrupton and represson are a toxc combnaton n many socetes, leadng to weak governments that are unable to enforce order or patrol ther borders and are vulnerable to terrorst networks and drug cartels (cted n Krueger and Maleckova, 2003, p.140). Ths statement dentfes the government as the crucal actor whose behavor creates the lnk between economc condtons and terrorsm. However, hs use of the expresson weak government mght be msleadng. Krueger and Maleckova (2003) and Krueger and Latn (2003) have found that repressve states are typcal of the countres of orgn of the perpetrators n a cross-country regresson estmated over the perod Ths rases a semantc ssue, as one can make a case that strong states don t need represson, and are better

5 4 equpped for securng cvl lbertes for ther ctzens. However, one potental concern wth these fndngs that cvl lbertes have a negatve mpact on the supply of terrorst events s the endogenety ssue. Arguably, there probably are hghly mltant groups n the countres from whch many terrorsts orgnate, whch mght n turn lead ther governments to adopt a repressve behavor. Frey (2004) argues that the government s facng a tradeoff between usng repressve counter-terrorsm measures ( the stck ) and relyng on more socal spendng for reducng the socal support to the terrorsts ( the carrot ). More mltant groups mght care less for socal support, especally f they have external sponsors (Squera and Sandler, 2006), thus pushng the government to choose more repressve methods. Hence, the presence of a repressve government and the supply of terrorsts mght n fact be jontly determned by the same cause, namely a hgh level of mltancy among some groups n the populaton. The present paper draws two lessons from ths debate, namely that a careful analyss of the role of the government s needed for understandng the lnks between ad and terrorsm, and that the endogenety ssue s crucal for a proper emprcal dagnoss. The emprcal analyss of the number of terrorst events has become a very actve feld of research. Todd Sandler has pursued ths lne of research over about two decades wth varous co-authors, and has publshed recently a synthess of hs fndngs (Enders and Sandler, 2006). These results have brought out the tme-seres propertes of terrorst events, and shown the effectveness of varous countermeasures used by the West. They rase the ssue of deflecton, whereby self-protecton by some countres dverts the flow of terrorst attacks to other countres. Krueger and Latn (2003) and Krueger and Maleckova (2003) have focused nstead on the flow of terrorst attacks orgnatng from dfferent countres and found that the key determnant s the presence of a repressve state n the source country. They underplay the role of economc varables, supportng the vew that poltcal determnants are more mportant. Smlarly, Testas (2004) focuses on poltcal represson rather than on ncome per capta, whch he fnds margnally sgnfcant or nsgnfcant n a sample of Muslm

6 5 countres. He fnds that the mpact of poltcal represson on the supply of terrorst events s non-monotonc, as both low and hgh levels of represson have a postve mpact. Abade (2006) also fnds a non monotonc mpact of the lack of poltcal rghts, usng the World Market Research Center s Global Terrorsm Index as the dependent varable. Ths s a rsk ratng ndex produced by an nternatonal rsk agency. However, Azam and Delacrox (2006) suggest that poltcal represson and the supply of terrorst attacks are jontly determned, both respondng postvely to the unobserved level of mltancy prevalng n some groups n the source country. However, although terrorsts are found n the upper tal of the ncome dstrbuton, as mentoned above, some emprcal studes suggest that the occurrence of terrorst attacks s negatvely correlated wth shfts of that dstrbuton. Several emprcal papers have shown that economc downturns are sgnfcant for explanng upsurges of terrorst attacks (Blomberg et al. 2004, L, 2005). Hence, the macroeconomc evdence seems to contradct the fndngs emergng from ndvdual data. Moreover, Basuchoudhary and Shughart (2007) fnd that t s economc freedom and secure property rghts that reduce the number of terrorst attacks by source countres, rather than poltcal rghts. Lastly, Azam and Delacrox (2006) suggest that these analyses suffer from an omtted-varable bas, as Offcal Development Assstance (ODA) s sgnfcant n ther own structural equaton. Hence, even f foregn ad was neffectve n rasng the recpent s natonal ncome or ts level of educaton, t could have an mpact by some other channels. The present paper offers a possble soluton to ths problem. Frey et al. (2006) have crtczed the standard practce of countng the number of terrorst events for addng up some very heterogeneous events. They argue that evaluatng ther costs would provde a more relevant measure of the severty of terrorst actvty that would offer a more sensble way for computng an aggregate ndex. Cran and Cran (2006) provde an estmate of the GDP losses entaled by terrorsm, suggestng that counter-terrorsm has a hgh socal return. Moreover, Frey et al. (2006) show that non-economc losses

7 6 represent a sgnfcant component of the socal cost of terrorst events. A more lethal event would reduce lfe satsfacton by a larger amount than a less lethal one, over and above any economc loss experenced by the survey respondents. Hence, a correct measurement of the damage nflcted by the terrorst events would gve more weght to the more lethal and volent attacks, focusng attenton on the most damagng ones. Nevertheless, an emprcal analyss of the number of events orgnatng n each country remans a useful task to perform n order to dentfy the possble polcy handles avalable for reducng ther flow. The measurement of the socal cost of terrorsm s probably more relevant for the target or the host country than for the source country, whch s the am of the analyss presented below. We focus here on the crosscountry determnants of the number of terrorst attacks by country of orgn. The ndvdual-level fndngs descrbed above about the mpact of wealth on terrorsm also rase a major challenge to ratonal-choce theorsts. Hgher wealth and educaton ncrease the opportunty cost of takng rsk n perpetratng a terrorst attack, and stll do not deter those who cross the lne. Three man arguments have been offered to reconcle ths fndng wth ratonalty. The frst one s based on the assumpton of ratonng on the volunteers market; there s an excess supply of volunteers for terrorst mssons, and the organzatons pck the most educated ones, lkely to be more effcent than the others. Bueno de Mesquta (2005) has developed ths lne of analyss, suggestng that polces mprovng the economc stuaton, ncludng, perhaps, foregn ad, could nevertheless play a favorable part n reducng moblzaton and volence. Hs model predcts that we could observe both the mcro-result descrbed above that terrorsts have an above-average educatonal level, and the opposte macro-level effect, namely that an ncrease n the general level of educaton n the populaton could reduce the supply of terrorst attacks. Ths would work ndrectly, through a reduced level of moblzaton mpactng on the terrorst organzaton s ncentves. A dfferent argument s used by Azam (2005), assumng that terrorsts are motvated by altrusm towards the next generaton. People wth a lower rate of tme preference wll nvest more resources n

8 7 educaton, and wll also be more wllng to sacrfce ther own lfes for the sake of the future generaton. Hence, the effect of educaton on the opportunty cost of puttng one s lfe at rsk mght be offset by ts postve mpact on nter-generatonal altrusm. That model suggests that some types of ad polces mght be effectve aganst terrorsm, dependng on the effect that they have on the tradeoff facng the potental terrorsts. Here as well, the mpact of a hgher level of educaton n the populaton on the supply of attacks cannot be sgned unambguously, whle the most educated ones would stll be more nvolved than the others. Berman and Latn (2005), Ferrero (2006) and Wntrobe (2006) provde a thrd lne of argument, and analyze nstead the socal pressure dmenson of the decson to opt for sucde bombng. More educated people would be more senstve to some knd of socal pressure, especally when they are students lvng far away from ther famles. They then tend to create a closed crcle wth other students. Sageman (2004) llustrates ths phenomenon wth examples from Al- Qaeda actvsts. The Hamas volunteers descrbed by Hassan (2001) provde a slghtly dfferent pcture, where socal pressure s more dffuse and pervasve. However, she shows that ntensve relgous tranng only comes after the decson to go for a sucde msson has been taken. These models do not have strong mplcatons for ad polcy. The four papers cted above (Azam, Berman-Latn, Ferrero and Wntrobe) vew sucde attacks as the ultmate test of the ratonal-choce approach to terrorsm. The present paper does not address emprcally ths ssue, but the theoretcal framework takes due account of these fndngs that educaton mght have a postve mpact on the value that some actvsts attach to performng a terrorst attacks. 2 The next secton brefly revews the lterature on the allocaton of foregn ad across countres, whch suggests that the latter s predomnantly determned by poltcal consderatons, rather than by the am of fghtng poverty per se, despte declaratons to the 2 Moreover, all these papers analyze the mcro-foundatons of terrorsm n general, whle the present one s restrcted to transnatonal terrorsm. It seems safe to assume that ther predctons extend to ths settng as well.

9 8 contrary. Secton 3 presents the smple model that captures the delegaton problem nvolved n the fght aganst transnatonal terrorsm. It s amed at brngng out the roles of foregn ad and educaton n the war on terror. The emprcal results are presented n two steps. Secton 4 presents the data used and a smple estmaton of the mpact of ad and educaton on the number of attacks orgnatng from each country. Secton 5 provdes a slghtly more sophstcated analyss, drected at testng whether an endogenety bas s present n that equaton. Secton 6 offers some concludng comments. 2. The Poltcal Determnants of the Allocaton of Foregn Ad across Countres As mentoned above, a careful analyss of the role of the government s needed for understandng the lnks between ad and terrorsm. Ths s consstent wth the theoretcal lterature on ad, whch s largely couched n the framework of the prncpal-agent model. There s now a szable lterature dscussng varous aspects of ad, whch has percolated somewhat n the polcy debate (e.g., World Bank, 1998). The basc structure of the theoretcal model of ad vews the recpent government as the agent of a foregn power, the donor. Both players have some common nterest, whch s wdely assumed to be poverty allevaton, albet wth dfferent weghts (Adam and O Connell, 1999, Azam and Laffont, 2003, Svensson, 2000 and 2003). Any other common nterest could be ncluded n the model, wthout changng the basc structure. Then, the am of the analyss s to brng out the mplementaton problems nvolved and to dscuss solutons that can make ad effectve, by the donor s standards. Azam and Saad-Sedk (2004) go one step further by lookng at the foregn power s choce between gvng ad and mposng sanctons. They provde a case study of the Iraq Kurds after the Provde Comfort operaton protected them aganst Saddam Hussen s persecuton, and conclude that they benefted from some economc growth under the alled forces mltary sheld. Secton 3 uses a smple nstance of such a prncpal-agent model where the donor s usng the recpent government as a delegate for performng some tasks on ts behalf.

10 9 The emprcal lterature on the allocaton of ad across developng countres takes on board the poltcal economy dmenson. Ad to developng countres has been ncreasngly delvered as program ad, condtonal upon the recpent government undertakng varous polcy reforms. The nfluental paper by Burnsde and Dollar (2000) has brought out that the effect of ad must be analyzed whle takng due account of some heterogenety among recpent countres. They favor an ndex of the qualty of macroeconomc polces as ther heterogenety parameter. However, ther results suggest that ad-effectveness, as measured by ts mpact on growth, s not the key determnant of the allocaton of ad across countres. Usng also cross-country regresson analyss, Svensson (1999) shows that ad s more effectve n affectng growth n more democratc countres, but s not allocated to the latter more favorably. Ths also suggests that ad allocaton s governed by other consderatons. There s thus probably a hdden agenda besde the generous drve to allevate poverty. Ths ssue s analyzed by Alesna and Dollar (2000), who show that colonal hstory and strategc allances are the man determnants of the amount of ad receved. They also show that, n the tme-seres dmenson, democratzaton s often followed by ncreased ad, although there s no sgnfcant statc effect of democracy. By contrast, Berthélemy and Tcht (2004) fnd a sgnfcant postve mpact of the Freedom House ndex of cvl lberty and poltcal rghts, n a panel data analyss over the perod , for 137 ad recpent countres and 22 blateral donors. Ths s confrmed n Berthélemy (2006), usng a dfferent estmaton method. The latter two studes brng out that most blateral donors are also guded by ther self-nterest n allocatng ther ad, and n partcular by ther commercal relatonshps. Fleck and Klby (2006a) show that the latter play also an mportant part n determnng the allocaton of U.S. blateral ad across countres, whle more conservatve presdents are more nfluenced by commercal nterests than more lberal ones. Fleck and Klby (2006b) suggest that such a dagnoss can be extended to the case of the World Bank, whose ad-allocaton behavor s sgnfcantly nfluenced by U.S. tradng and poltcal

11 10 nterests. However, one may wonder whether tradng flows are exogenous. Although most blateral donors have formally ruled out ted ad, towards the end of that sample perod, some mplct and subtle ways of tyng ad lkely reman n operaton. Moreover, ad helps fnancng the trade defct of developng countres, and ths s boostng ther mports from ndustralzed donors. Hence, some reverse causaton between ad and trade mght also be at work. Chauvet (2002) looks at the relatonshp between ad allocaton across countres and soco-poltcal nstabltes. She shows that the ad flow s somewhat drected at governments that are under poltcal threat, whle t shes away from the threats that are drected more specfcally at the economy. Ths s related to the fndng by Coller (2007) that a nonneglgble fracton of ad leaks out fnancng the mltary, wthout much effort by donors to avod t, whle ad s pretty neffectve at lftng people out of poverty. These results agan support the vew that donors are poltcally motvated when gvng ad, whle economc ssues play a secondary role. Smlarly, Azam and Delacrox (2006) show that ad allocaton across countres s nfluenced by the donors goal of fghtng terrorsm. The present paper also analyzes a poltcal dmenson of the allocaton of ad across countres, by lookng at ts relatonshp wth terrorsm. Instead of lookng at the ndrect lnk between ad and poverty, and then n turn to the addtonal lnk between poverty and terrorsm, as done n the polcy debate mentoned above, we are lookng here at the drect lnk between ad and terrorsm. 3. The Model We model a donor that allocates ad between a number of countres, whch are lable to produce some terrorst attacks aganst the donor. In each of these countres, the government s able to exert some effort to fght terrorsm, at a cost. Then, ad s a way to defray the recpent government for ths cost of effort. The government s acton exerts ts nfluence on the value of a terrorst ht for the actvsts, whch also depends on some dosyncratc

12 11 mltancy parameter. Hence, for each country, three players are nvolved: () the terrorst group determnes the number of attacks perpetrated aganst the donor, () the local government exerts some effort to deter these actons, whle () the donor provdes some ad for compensatng the government. Ths model s an extenson of the one presented n Azam and Delacrox (2006) 3 amed at capturng the choce between the carrot and the stck as a way of controllng terrorsm n the ad-recpent country. The Three Agents We capture ths framework usng the followng specfcaton. Denote Y the gven ncome of the donor, A the total amount of ad delvered, and H the total number of terrorst attacks httng the donor and comng from the dfferent countres. Assume that the donor ncurs a cost ψ ( H ) because of these attacks, assumed ncreasng and convex. Now, denote a the ad gven to country { 1,..., n}, and h the number of attacks orgnatng n. Then, by defnton, A a = and H = h. The donor s also assumed to splt ts ad flow between general budget support to the government denoted b and an amount s whch s earmarked for supportng educaton. The latter captures the socal concern expressed by the donor n ts relatonshp wth the recpent government. Hence, a = b + s,. Country s government values the ad flow b as well as the level of human captal acheved n the country denoted k. The latter produces a level of utlty ( ) u k (assumed ncreasng and concave) for the government, whch s meant to capture all the postve fallout of human captal, bult up through educaton and health, rangng from the mproved socal welfare to the ncreased fscal resources that a more sklled populaton s lable to pay. Ths utlty functon thus captures both the pecunary and non-pecunary benefts that the ruler gets from the level of human captal present n hs country. The unt cost of producng ths level of human captal s denoted (, ) δ ε, whch s decreasng n the donor s earmarked contrbuton s 3 The workng paper verson of that paper s longer and more comprehensve than the publshed one, and s avalable as IDEI Workng Paper No. 324 at

13 12 s and n the country s own past nvestment n human captal, whch we call educatonal captal from now on, denoted ε. The latter captures all the socal and physcal nfrastructure that the country has n the educaton (and health) sector, ncludng ts cultural tradtons, ts schools and unverstes, as well as ts stock of traned teachers, etc. The donor s earmarked contrbuton to the educaton sector s s modeled here as reducng the cost of expandng human captal borne by the government. Ths specfcaton entals that such an earmarked contrbuton to the educaton sector s not perfectly fungble wth ether local funds or other ad flows. Ths may capture, for example, the use of dfferentated nputs lke hghly qualfed teachers, wthout local substtutes, that would not be avalable wthout the donor s nterventon, or scholarshps for jonng select nsttutons abroad. We can thus defne the government s proft from human captal nvestment as: (, ) max ( ) (, ) π s ε = u k δ s ε k. (1) k It s easly checked that ths proft functon s ncreasng n ts two arguments, reflectng the cost-savng effects of both the local educatonal captal and the donor s earmarked contrbuton. Applyng Hotellng s lemma, (1) mples that country s human captal level may be wrtten as an ncreasng functon of s and ε : k (, ) = k s ε. (2) The country s government also ncurs a cost ξ ( r ) (assumed ncreasng and convex) when performng an amount r of represson aganst terrorsts, wth ( ) ξ 0 = 0. Then, the ad contract wll specfy how much represson the local government s expected to perform aganst the terrorsts wthn ts sphere of nfluence n return for the ad receved, ncludng the earmarked contrbuton to the educaton sector. Ths vew of the ad contract captures the dea that the foregn power has to delegate part of the protecton of ts nterests aganst terrorsm to local governments, usng ad as a means of defrayng the costs of dong so ncurred by the

14 13 local government. In order for ths contract to be acceptable by the latter, the trplet { b, s, r } must fulfll the followng government s partcpaton constrant, where we normalze ( ) π 0, ε = 0 : ( ) ( ) b + π s, ε ξ r 0. (3) Let h be the number of attacks perpetrated by country s terrorst organzaton aganst the foregn power s nterests. The terrorst organzaton attaches a unt value θ v( k ) to these attacks, where θ s the mltancy parameter, assumed known to both the donor and the government, and v( k ) s an ncreasng functon capturng the postve mpact of human captal on the value attached by the terrorst organzatons to the attacks perpetrated aganst the foregn power s nterests. Ths postve effect of human captal on the value of terrorst attacks for the perpetrators s meant to capture the fact mentoned n the ntroducton that terrorsts generally have an above-average educaton level, as frst ponted out by Krueger and Maleckova (2003) and Krueger and Latn (2003). The terrorst organzaton s also ncurrng a cost ( h, k, r) ω for perpetratng ts attacks. Ths cost functon s naturally assumed to be ncreasng and convex wth respect to h, and ncreasng n r. The mpact of k s less clearcut, and probably combnes two opposng effects. There s frst a postve mpact, as more educated people have a hgher opportunty cost, whch the terrorst organzaton wll most probably take nto account. There s then a second mpact gong n the other drecton, as more educated people are probably more effcent at performng the attacks. The latter effect has been analyzed by Bueno de Mesquta (2005), as dscussed n the ntroducton. However, we do not need to make a frm assumpton regardng the margnal effects of represson and educaton on the terrorst organzaton s costs, as they do not affect the model s man predctons. The latter only depend on the cross-second partal dervatves,.e., the mpact of these varables on the margnal cost of perpetratng an attack for the terrorst organzaton. It

15 14 s natural to assume that represson ncreases the margnal cost of perpetratng an attack. Denotng cross-second dervatves by subscrpts, ths means that ω hr > 0. The tme lne of the game s as follows: () the donor offers the ad contract descrbed above; () the government exerts the agreed level of represson and chooses ts preferred educaton level, both assumed to be perfectly observable and contractble by the two partes; () the terrorsts launch ther chosen number of attacks h ; and lastly (v) the ad s delvered and consumed. Hence, ths game can be solved by backward nducton. We frst derve the terrorst organzaton s best-response functon, as a functon of the government s polcy varables. Then, the attacks supply curve s derved at the country level, by brngng n the government s preferred mx of represson and educaton expendtures, takng due account of the donor s nfluence on that choce. The terrorst organzaton chooses ts level of attacks h wth a vew to maxmzng: ( ) ω(,, ) θ v k h h k r. (4) Then, ts best-response functon h h( θ, k, r) condton θ v( k ) = ω ( h, k, r), where ω ( ) h = may be derved from the frst-order h denotes the dervatve of the cost wth respect to the level of attacks. Denotng ω > 0, ω > 0 and ω the relevant second dervatves of the cost functon, the latter mples: ( ) ( ) hh hr hk ( ) ( ) ( ) ( ) ( ) h v k hr ' h ω v k h θ ωhk = > 0, = < 0 and = θ ω r ω k ω hh hh hh. (5) The sgns of the frst two partal effects on the left-hand sde are farly ntutve, and do not call for much comment: more mltant groups produce more attacks, whle a greater represson effort by the government reduces the number of attacks. The thrd effect s ambguous, as more human captal ncreases the value of terrorst attacks whle ts mpact on the margnal cost of these attacks s tself ambguous, as dscussed above. As mentoned n the ntroducton, ths ambguty s compatble wth the predctons of ether Bueno de Mesquta

16 15 (2005) or Azam (2005) concernng the macro-mpact of educaton on the level of terrorst actvty. The Attacks Supply Curve Now, the donor wll choose ts preferred ad contract wth a vew to maxmzng ( ) Y A ψ H over all countres, subject for each one of them to the terrorsts best-response functon h h( θ, k, r) k (, ) =, the government s preferred human captal nvestment polcy = k s ε and the government s partcpaton constrant (3), whch may be wrtten as: (, ) ( ) a s π s ε + ξ r. (6) Ths problem can be decomposed nto two steps: () Ad composton problem: the effcent attacks supply curve s determned for each country by mnmzng h, usng s and r as control varables, gven the level of a, the local government s partcpaton constrant (6), ts preferred educaton polcy (2), and the terrorst organzaton s best response functon. 4 () Ad allocaton problem: the donor s optmal allocaton of ad across countres s determned by mnmzng A ψ ( H) constrants. 5 +, takng all the effcent attacks supply curves as The donor has no reason to leave any postve rent to the local government, and thus (6) wll hold wth equalty. Then, solvng smultaneously these three equatons allows us to establsh proposton 1. Proposton 1: () The number of terrorst attacks orgnatng from country can be wrtten as the followng attacks supply curve: 4 The second-order condton for ths problem s messy, and nvolves second- and thrd dervatves for whch not much ntuton seems compellng. It requres the terrorsts best-response functon to be quas-convex n s and r,, after substtutng for the employment polcy, and the government s partcpaton constrant to be concave n the same space. We smply assume that t holds n the relevant neghbourhood. 5 The second-order condton requred here s not more enlghtenng than at the prevous footnote. It essentally requres the effcent attacks supply curves derved at (7) below to be convex n a.

17 ( θ, ε, ) 16 h = h a, (7) such that, denotng λ > 0 the Lagrange multpler assocated wth the government s partcpaton constrant: ( ) ( ) h v k h λ δ ε h = > 0, = < 0 and = λ < 0. (8) θ ω ε δ s a hh () The donor s optmal allocaton of ad across countres s determned by settng: ( ) { } λ = 1 ψ ' H, 1,..., n. (9) Proof: Proposton 1 () s establshed by mnmzng the terrorsts best-response functon = ( θ,, ) subject to the government s preferred educaton polcy k k( s, ε ) h h k r =, and the government s partcpaton constrant wrtten as an equalty a s π ( s, ε ) ξ( r) frst-order condtons for ths problem yeld: (( π s) 1) h λ h = and = λξ ' k k s r ( r ) = +. The. (10) Then, the frst-order condton for the government s problem (1) can be used to derve the followng partal dervatves: π k k δ = and = δ s s s u'' k s ( ), (11) whch can be substtuted n (10). The resultng expresson can then be used wth the terrorst organzaton s best response functon and the government s partcpaton constrant to yeld the dervatves gven n (8) by takng ther total dfferental and rearrangng the terms. Then part () s establshed by mnmzng A ψ ( H) supply curves as constrants. +, takng all the effcent attacks A notceable property of the attacks supply curve, as expressed n (8), s that the sgns of the mpacts of the educatonal captal and of ad do not depend on the terrorsts

18 17 parameters, once (9) s taken nto account, but only on the cost functons enterng the donor s and the government s objectve functons. In partcular, the fact that the value attached by the terrorsts to the attacks v( k ) was assumed above to be ncreasng n the level of human captal s rrelevant for these predctons. Ths suggests of course that the observed fact that terrorsts have an above-average educaton level mentoned n the ntroducton s rrelevant for ad polcy. Proposton 1 tells us that even n that case, more educatonal captal ends up reducng the number of terrorst attacks at the cross-country level, because ts effect s more than compensated by an adjusted level of represson; for a gven ad level, a hgher level of educatonal captal allows the local government to reach ts human captal objectve at a lower cost, and thus frees some resources for performng more represson. 6 Hence, for some values of the parameters, t s possble that the donor wll elct more represson from the government by earmarkng more fundng to educaton 4. Emprcal Analyss We test the man predctons of the model descrbed above usng a dataset smlar to the one used by Krueger and Maleckova (2003) and Krueger and Latn (2003), based on the same source. We thus focus on the number of terrorst events per country of orgn of the perpetrators 7. These data are avalable on the nternet, n the database provded by the Internatonal Polcy Insttute for Counter-Terrorsm (ICT). We have extended the sample used n the two Krueger et al. papers cted above to cover the perod from January 1990 to March More precsely, the number of terrorst events s computed from a set of 1119 terrorst ncdents, takng place between January 1990 and March (Azam and Delacrox, 2006). All these events are transnatonal n that the target and the source countres are dfferent. These terrorst attacks are aggregated over the perod mentoned above to produce a number 6 The appendx wraps up the theoretcal analyss by explanng how the a s and H are jontly determned. 7 Hence, a terrorst attack perpetrated by two terrorsts from dfferent countres s counted as two events. 8 Alexandra Delacrox has produced ths data set when she was a DEA student at Toulouse Unversty.

19 18 of attacks orgnatng from each country durng that perod. Durng the latter, the terrorst attacks orgnated from 80 source countres n our sample, as we have excluded Afghanstan, Bosna, Cuba, Myanmar, Somala and Yugoslava, for lack of some data on the explanatory varables. Thus, we are left wth the source countres of terrorst attacks accountng for slghtly less than half the 176 countres n our sample. Table A1 n the appendx gves the full lst of the number of attacks by country of orgn used n our estmatons. Only 18 countres are the source of more than 18 attacks, whle 18 countres are the source of only 1 event over that perod. Not surprzngly, the West Bank and Gaza Strp provde the largest number of attacks, wth 400 terrorst events orgnatng there. However, the Mddle East s not the only source of terrorsm as Sub-Saharan Afrca, South Asa, Europe and Latn Amerca are all represented n the top-10 source countres. Many OECD countres are also the source of terrorst attacks, ncludng France, the U.K. and the U.S.A. The man goal of our emprcal analyss s to test the predctons that foregn ad and educatonal captal have a negatve mpact on the supply of terrorst attacks orgnatng n the recpent counntres. We use the standard measure of foregn ad, namely Offcal Development Assstance (ODA). Ths varable aggregates the dsbursements of loans (wth a hgh enough grant component) and grants by offcal agences of the members of the Development Assstance Commttee (DAC) to promote economc development and welfare n the recpent countres. These data are measured n constant 2000 U.S. dollars. In the sample used n ths artcle, 24 countres are ad donors, manly among OECD member countres. The model presented above suggests that ths s an endogenous varable, nsofar as the donor s allocatng ad wth a vew to controllng the supply of terrorst attacks from recpent countres. Moreover, there s no compellng argument for decdng whether ODA should be measured per capta or as a rato to GDP. Whle the former specfcaton seems to capture better the potental benefts that the country wll get from ad, the latter seems more approprate for measurng the need for t. Consequently, we use both specfcatons n what

20 19 follows, showng that the standard econometrc procedure for choosng between two nonnested hypotheses s not conclusve n ths case. It s less straghtforward to fnd the rght measure of educatonal captal n a comparable way across countres. We have chosen to use the enrollment rate n secondary educaton n our man equaton, but the appendx presents some robustness checks. Ths varable s thus lable to be endogenous on two accounts. Frst, the theoretcal model above predcts that the donor wll earmark some funds for nfluencng the decson of the local government to fght terrorsm wthn ts sphere of nfluence by reducng the cost of nvestng n educaton. Second, ths varable s bound to measure educatonal captal naccurately, and thus to nclude some measurement error entalng a correlaton between the ncluded varable and the resduals. Both ODA and secondary school enrollment are probably correlated wth the level of economc and poltcal development of the recpent countres. It s thus mportant to control for the latter n order to mtgate the rsk of fndng a spurous correlaton wth ad and educaton due n fact to under-development. We use GDP per capta and populaton sze as controls n our regressons. GDP n constant 2000 U.S. dollars s dvded by mdyear populaton. However, because represson s endogenous n our theoretcal framework, and optmzed out for dervng the structural equaton (7), we do not nclude a control varable capturng cvl lberty or democracy. The latter are mplctly determned by the exogenous varables n the model, lke mltancy and educatonal captal, and the level of foregn ad. The source of data for all these explanatory varables s the World Bank s World Development Indcators onlne (WDI). We use a sample of 176 countres, averagng populaton, GDP per capta, ODA per capta, ODA as a percentage of GDP and secondary school enrollment over the perod Table A.2 n the appendx provdes some summary statstcs for these data. ODA per capta s Offcal Development Assstance dvded by the mdyear populaton sze. A majorty of countres have an ODA per capta between 0 and 50 dollars. Countres recevng more than 100 dollars per capta are the poorest among the

21 20 developng countres. The majorty of countres wth a rato of ODA to GDP hgher than 30% are n Sub-Saharan Afrca. Dummy varables are used for ndcatng these countres, as a check on possble non-lnearty or a dfferental treatment for them by the donors. For measurng the level of secondary educaton, we use the gross enrollment rate,.e., the rato of total enrollment, regardless of age, to the populaton of the relevant age group. The resultng rato can therefore be hgher than 100 %. The majorty of countres have a gross percentage of secondary school enrollment between 90 and 100 % but many countres have a much lower rate, especally n developng countres. Fnally, as emphaszed by Azam and Delacrox (2006), t s very mportant to try to control for mltancy n order to dentfy the correct structural equaton for the attacks supply curve. The crucal element of ther dentfcaton strategy s the sgn of the mpact of ad, whch s negatve n the structural equaton, whle there s a postve correlaton across countres f no attempt s made to control for mltancy. We follow the same route here and use some dummy varables as proxes for capturng ths unobserved varable. 9 After a lttle expermentaton, the followng dummy varables turned out to be useful: West Bank and Gaza, Camp Davd (Egypt and Israel), Chna and Inda, Latn Amercan countres, ASEAN countres before 1990 (Indonesa, Malaysa, Phlppnes, Sngapore and Thaland), OECD countres before 1990, Sub-Saharan Afrcan countres and former USSR countres. Of course, these dummy varables also control for other country characterstcs that may affect the supply of terrorst attacks, lke geography and cvlzaton. Moreover, these varables can also capture some effects of the educatonal captal that are not captured by the secondary school enrollment rate. The attack supply curve cannot be analyzed by standard least squares estmaton because the dependent varable takes only non-negatve nteger values correspondng to the 9 A potentally useful varable for capturng some aspects of ths mltancy level s the ethnc tenson ndex that affects sgnfcantly the level of terrorst attacks by country of orgn n Basuchoudhary and Shughart (2007). Testng for ths falls outsde the scope of the present paper.

22 21 number of terrorst events, whle most countres are the source of no attacks at all. The standard methods for analyzng such count data nclude the Posson regresson and the negatve bnomal model (Cameron and Trved, 1986). The latter s less restrctve than the former because t has one more free parameter. The assumed equalty of the condtonal mean and varance functons s the most restrctve assumpton of the Posson regresson model. The negatve bnomal specfcaton ntroduces an ndvdual, unobserved effect nto the condtonal mean. Accordng to ths model ndvduals have a constant but unequal probablty of experencng an event. We use the latter specfcaton, followng McCullagh and Nelder (1983). We tested t aganst the Posson model, and concluded that the negatve bnomal model s always preferred. Therefore, n the followng analyss we only present the negatve bnomal results. Both the theoretcal model presented above and the econometrc results presented n Azam and Delacrox (2006) conclude that the number of terrorst events by country of orgn and the amount of ad are smultaneously determned. However, because t s well known that most methods amed at correctng for the endogenety bas are lable to reduce the effcency of the estmators, we start the analyss by performng the estmatons wthout takng care of t. In the next secton, we then test whether these benchmark estmates are msleadng by testng for the presence of a sgnfcant endogenety bas. Table 1 presents two dfferent estmated equatons of the number of terrorst events orgnatng from each country, estmated on the 176-country sample and where the varable ODA s n percentage of GDP. Equaton 1 ncludes populaton sze as a control varable, whch turns out to be nsgnfcant. Equaton 2 then s estmated whle excludng populaton from the lst of explanatory varables. Ths does not affect much the other coeffcents. These equatons provde some support for the two mantaned hypotheses, as the rato of ODA to GDP and secondary school enrollment have a sgnfcant mpact wth the predcted sgn. Regardng the other control varables, per capta GDP shows up wth a negatve sgn. Ths s

23 22 consstent wth casual observaton, whch suggests that terrorsts manly come from developng countres. As many characterstcs of the level of economc and poltcal development are known to be correlated wth per capta GDP, one should not nfer too much from ths result. It does not mply that per capta ncome strcto sensu s a key factor n explanng the supply of terrorst events. Krueger and Maleckova (2003) have presented a strong argument aganst such a vew. Nevertheless, t s a useful control varable to have, n order to dsentangle the effect of foregn ad from that of under-development. Table 1: Number of Terrorst Events Orgnatng From Each Country (Wth ODA n Percentage of GDP) Equaton 1 Equaton 2 Varables Coeffcent z-statstcs p-value Coeffcent z-statstcs p-value Intercept Populaton 4.12E GDP p.c E E ODA (% GDP) Secondary enrollment (% gross) West B. Gaza "Camp Davd" Sub-Saharan OECD Nb. Obs R² Pseudo-R² LR-Stat ch2(7 df) Note: Equaton 1 and 2 are negatve bnomal regressons estmated by maxmum lkelhood. Three of the dummy varables ncluded are sgnfcant. The West Bank and Gaza dummy varable s an obvous canddate for beng used as a proxy for mltancy. The estmates confrm that t s hghly sgnfcant. The Camp Davd one, whch ndcates Egypt and Israel, also has a sgnfcant postve sgn, whch s less than 2/3 as hgh as that for the

24 23 West Bank & Gaza Strp. Lastly, the dummy varable ndcatng the OECD countres also has a sgnfcantly postve sgn, suggestng the presence of a hgh degree of mltancy over the perod consdered. A glance at table A.1 shows that sx OECD countres have produced at least sx terrorst attacks, ncludng Span (31), Turkey (28), Greece (11), France (7), Ireland (6) and Italy (6). The mportance of these dummy varables s confrmed by runnng these estmatons wthout them. 10 The results are shown n the appendx, where the equatons run wthout the dummy varables have no sgnfcant lkelhood rato tests. Table A.3 presents ths result for ODA as a percentage of GDP and ODA per capta. Hence, these dummy varables contan some relevant nformaton for dentfyng our equaton, although they most probably fall short of measurng mltancy wth any accuracy. Table 2: Number of Terrorst Events Orgnatng From Each Country (Wth ODA per Capta) Equaton 3 Equaton 4 Varables Coeffcent z-statstcs p-value Coeffcent z-statstcs p-value Intercept Populaton 3.50E GDP p.c E E ODA p.c Secondary enrollment (% gross) West B. Gaza "CampDavd" Sub-Saharan OECD Nb. Obs R² - - Pseudo-R² LR-Stat ch2 (7 df) Note: Equaton 3 and 4 are negatve bnomal regressons estmated by maxmum lkelhood. 10 The dummy varables for Sub-Saharan Afrca and OECD are kept n the equatons presented at tables 1 and 2 for the sake of comparson and testng because they are sgnfcant n ether one or the other.

25 24 Usng ODA per capta nstead of ODA as a rato to GDP does not affect much the dagnoss, as shown by table 2. The results are qualtatvely the same as n the prevous case, the same coeffcents beng sgnfcant wth the same sgn. Hence, we can draw the same concluson regardng ODA and secondary educaton, although the latter s only sgnfcant at the 10 % level when the nsgnfcant populaton sze s dropped, whle t s not sgnfcant otherwse. Another dfference s that now, t s the Sub-Saharan Afrca dummy whch s sgnfcant, whle the OECD one becomes nsgnfcant. Over the perod very few Afrcan countres are concerned wth terrorsm, whle soco-poltcal nstabltes of another knd are endemc, ncludng coup d etat, cvl wars and rots. One can observe that the majorty of countres concerned by terrorst actvtes were not undergong a cvl war. Hence, both equatons lead to the same concluson, namely that ODA, under both specfcatons, the secondary school enrollment rate, and GDP per capta negatvely affect the supply of terrorst events by country of orgn. These results shed some lght on the lessons drawn from the mcroeconomc data dscussed n the ntroducton. Even f terrorsts are predomnantly recruted among the relatvely wealthy and educated socal strata, as mentoned above, the secondary school enrollment rate and GDP per capta have a negatve nfluence on the number of nternatonal terrorst events perpetrated by ndvduals from each country. Ths seemngly contradctory result s one of the predctons of the theoretcal model presented above, and these emprcal results thus provde some support for the latter. The overall performance of these equatons seems relatvely weak, when lookng at 2 the pseudo- R s, but the lkelhood rato tests accept the overall sgnfcance of the equatons. 11 Moreover, we tred to choose between the two specfcatons presented usng the J-test proposed by Davdson and MacKnnon (1982) for choosng between two non-nested models. The dea s that f one model s the correct one, then the ftted values from the other 11 We ran some of these regressons usng both Stata and E-Vews, and found much hgher pseudo-r 2 wth the latter. We just present here the smaller of the two, for the sake of remanng on the safe sde.