How much terror? Dissidents,governments,institutions,and the cross-national study of terror attacks

Government coercion

It is widely agreed that states’ coercive behavior influences the behavior of dissidents (e.g. Lichbach, 1987; Enders & Sandler, 1993). Many scholars contend that violent repression raises the costs of dissent (e.g. Gurr, 1970; Tilly, 1978), and some have argued that when the cost of one tactic rises relative to others, dissidents will switch to the less costly tactic (e.g. Lichbach, 1987; Moore, 1998; Enders & Sandler, 2006). We noted above that dissident groups are influenced by their beliefs about the relative costs of alternative tactics. We contend that group’s beliefs about the costs of using violence will be conditioned on the government’s use of violent coercion. We thus hypothesize that state coercion will reduce the level of terror we observe. ³

As often happens in conflict literature, counter-arguments exist to support the conjecture that coercion stimulates, rather than reduces, dissent. These are grievance-based arguments, and the authors who advance them do not dispute that coercion is costly to dissidents. Instead, they argue that coercion alienates people who would have been indifferent or supported the government (e.g. Gurr, 1970; Rasler, 1996; Walsh & Piazza, 2010). Several scholars maintain that arbitrary targeting exacerbates this tendency (e.g. Mason & Krane, 1989; Mason, 2004; Kalyvas, 2006; Findley & Young, 2007), and Wilkinson (2001) in particular advocates that counterinsurgency is a political, rather than military, struggle and that coercion is less effective than investigative police work and prosecution. Each of these arguments suggest that coercion will increase the aggregate level of terror we observe. ⁴

Dissident competition

A prominent argument in the literature on the use of terror by dissidents suggests that groups that compete for the allegiance of the same supporters can get into an outbidding process where each pursues more violent tactics in an effort to show that they are the strongest group (Tillion, 1960: 50–52; Hutchinson, 1972: 391; Rabushka & Shepsle, 1972; Brubaker & Laitin, 1998: 434; Bloom, 2004; Kydd & Walter, 2006). The implication is that terror attacks stimulate more terror attacks. Because we are aggregating the data at a coarse temporal level the hypothesis is not well captured by inclusion of a lagged term in a regression. Instead, we focus attention on the implication that there will be evidence of overdispersion in an event count model of terror attacks. One reason overdispersion occurs is contagion, meaning the occurrence of one event increases the probability of another event occurring in the future (see King, 1989: 768). If outbidding occurs, then one group’s decision to plan and execute terror attacks is a positive function of other groups’ attacks. As such, outbidding implies overdispersion when data are aggregated across large units of time.

Turning to the relationship between violent (non-terror) dissent and terror attacks, outbidding suggests that there will be a positive relationship between the two. That is, as a dissident group observes violent attacks against the state being committed by other dissident groups, competition for supporters should press it toward devoting more resources to violence, including terror attacks.

What impact will nonviolent protest activity have upon dissidents’ use of terror? Nonviolent protest relies on mass mobilization, needing many people to participate (e.g. DeNardo, 1985; Chong, 1991; Schock, 2005; Chenoweth & Stephan, 2011; Cunningham, 2013). It is conceivable that the more nonviolent protest occurs in society, the more successful that tactic will appear to be, which will lead groups to re-evaluate the merits of nonviolent means relative to violent means. Thus we conjecture that nonviolent protest may reduce the number of terror attacks.

Once again, counter-arguments are available. Mullins & Young (2012) argue that the more common is anti-state protest and violence, the more legitimated all forms of anti-state protest and violence become. They show that crime and even capital punishment are positively associated with terror events. In addition, expected utility arguments suggest that dissidents’ beliefs about the probability of success will influence the likelihood that they challenge the state. The more anti-state activity there is, the weaker the state appears to be, and that will increase groups’ beliefs about the likelihood of successfully challenging the state. If observing greater levels of nonviolent dissent leads groups to conclude that the state is weak, then that group might infer that terror attacks will produce concessions. Thus, as with government coercion, we cannot rule out the possibility that nonviolent protest may increase the aggregate number of terror attacks.

Institutions as context

The socio-economic context and other important factors

Macroeconomic performance is associated with grievances: the lower the level of performance, the greater the average level of grievances among the population. While grievances are not sufficient to produce dissident activity, the intensity and scope of grievances should covary positively with the amount of dissident activity, including acts of terror (e.g. Gurr, 1970; Crenshaw, 1981; Blomberg, Hess & Weerapana, 2004a,b). ⁹ That said, Krieger & Meierrieks (2011: 10) report that studies have found little support in favor of a relationship between macroeconomic performance and transnational terror attacks. Given that these results might be a function of analyzing transnational terror attacks, however, we include macroeconomic performance in our model.

Ethno-linguistic composition refers to how heterogeneous the population of a state is. There is a large literature on the effect of ethnic heterogeneity on large-scale conflict (Horowitz, 1985; Posen, 1993; Ellingsen, 2000; Sambanis, 2001; Reynal-Querol, 2002; Fearon & Laitin, 2003; Habyarimana et al., 2007). While the theoretical links between ethnic heterogeneity and conflict remain somewhat obscure (Saideman, 2010) and the empirical findings concerning this relationship are mixed, we conjecture that if ethnic diversity is thought to be a cause of violence between governments and dissidents then we should not rule out the possibility that it will exhibit some relationship with terror attacks (Krieger & Meierrieks, 2011: 12). ¹⁰

The relationship between physical quality of life and terror attacks remains unclear. This variable has been shown to be negatively correlated with state failure (Goldstone et al., 2010), and Piazza (2008b) finds state failure has a positive impact on international terror attacks. As mentioned above, Coggins (2015) also finds that political collapse is positively related to both domestic and international attacks. As such, the higher the level of quality of life, the lower we might expect terror incidence to be. However, Coggins (2015) also finds that measures of human security are not strongly associated with terror attacks, and in one case there is a positive relationship between the two.

The size of the population has been shown to have a positive impact upon terror attacks (Krieger & Meierrieks, 2011; Gassebner & Luechinger, 2011). Finally, a state’s involvement in external conflict may affect terror attacks (Drakos & Gofas, 2006).

Terror events

To measure the number of terror events at the country-year level we use the Global Terrorism Database (GTD 1.1), available through the Interuniversity Consortium for Political and Social Research (ICPSR). ¹¹ Specifically, we used the GTD 1.1 events data to produce an event count of all types of terror events. The GTD 1.1 dataset codes all terror events regardless of the nationality of the perpetrators and targets and it thus allows us to include domestic as well as international terror events. The temporal domain of our study is 1971–2007. The countries and years included in the analysis can be found in the Online appendix.

Noisy and missing data

As is common in CNTS datasets, ours is full of potentially noisy indicators with many missing values. For this reason we employ Bayesian factor analysis to create measures for several of our concepts. These models have two advantages over classic factor analytic techniques – they directly estimate the values of the latent trait and provide measures of uncertainty for these estimates. When using the resulting variable in subsequent predictive models, this uncertainty propagates directly into estimation.

Another advantage of the Bayesian factor model is the ability to avoid listwise deletion. The model permits one to use differing numbers of indicators for each country. Listwise deletion can lead to biased results, and we are able to retain every observation for which we have values for most, but not all, of our variables. This is particularly helpful in a cross-national setting where it is likely that data will be missing. The estimates will have larger standard errors for country-years with more missing data, which is a desirable result since we should be more uncertain when we have less information. ¹² The analysis below would not be possible outside of a Bayesian framework: the classic model would retain, at best, roughly 3.5% of the observations, and that for only a handful of countries.

Observed indicators for measurement models

Here we briefly discuss the data used to create our latent variables. More detailed information can be found in the Online appendix.

Political institutional context

The indicators used to construct measures of participation, contestation, and veto players are listed in Table I. The Suffrage indicator from Paxton et al. (2003) measures the percentage of the adult population that is eligible to vote. Vanhanen’s (2000) participation measure, Part, indicates the proportion of the adult population that actually votes. Finally, a political liberties measure, Political rights, is taken from Freedom House’s 2008 data. Political liberties include electoral processes and participation.

The first three contestation indicators (Parcomp, Xrreg, and Xrcomp) come from the Polity IV dataset (Marshall & Jaggers, 2009). Parcomp measures ‘the extent to which alternative preferences for policy and leadership can be pursued in the political arena’ (Marshall & Jaggers, 2009: 29). Xrreg indicates ‘the extent to which a polity has institutionalized procedures for transferring executive power’ (Marshall & Jaggers, 2009: 18). Xrcomp measures the competitiveness of executive recruitment, meaning the extent to which current political institutions provide equitable access to leadership positions. Vanhanen’s (2000) contestation measure, Comp, is constructed by subtracting the vote share of the largest party from 100. Military executive is a dichotomous variable drawn from the Database of Political Institutions and is coded as 1 if the executive is a military officer.

We use four data sources to measure the veto concept: the Political Constraints Data (Henisz, 2000), DPI (Keefer, 2005), Clague et al. (1999), and the International Country Risk Guide (Political Risk Group, nd). PolconIII is a continuous variable ranging from 0 to 1 that measures the difficulty with which any single government actor is able to change policy. J is a binary measure that indicates the presence of a judicial veto. From DPI we take the variable Checks, which measures the number of veto

OPEN IN VIEWER

Table I.

Measures of political institutions

Concept	Variable	Source
Participation	Suffrage	Paxton et al. (2003)
Participation	Part	Vanhanen (2000)
Participation	Political rights	Freedom House (2008)
Contestation	Parcomp	Polity IV (Marshall & Jaggers, 2009)
Contestation	Xrreg	Polity IV (Marshall & Jaggers, 2009)
Contestation	Xrcomp	Polity IV (Marshall & Jaggers, 2009)
Contestation	Comp	Polyarchy, Vanhanen (2000)
Contestation	Military Executive	DPI, Keefer (2005)
Veto players	PolConIII	Henisz (2000)
Veto players	J	Henisz (2000)
Veto players	Checks	DPI, Keefer (2005)
Veto players	CIM	Clague et al. (1999)
Veto players	Corruption	Political Risk Group (nd)
Veto players	Expropriation risk	Political Risk Group (nd)
Veto players	Rule of law	Political Risk Group (nd)
Veto players	Repudiation risk	Political Risk Group (nd)

points in government, including legislative chambers and opposition parties. We also use several indicators that capture the risk of arbitrary confiscation of wealth or property. The Clague et al. (1999) CIM measure is calculated as the proportion of the total money supply that is in banks or otherwise invested, higher numbers indicating greater trust in the enforcement of property rights. The four measures drawn from the ICRG data are Corruption, Rule of law, Expropriation risk, and Repudiation risk (risk of repudiation of contracts by government actors).

Measurement models for latent variables

For each measurement model the observed values y of the latent concept X are modeled with a distribution appropriate to the observed indicator. Continuous indicators, and ordinal indicators with six or more categories, were assumed to follow normal distributions, and were all standardized prior to estimation. Formally, for continuous indicators we assume:

y i , t ∼ N ( μ i , t , σ 2 ) μ i , t = β X i , t

For binary indicators we assume:

y i , t ∼ B e r n o u l l i ( p i , t ) l o g i t ( p i , t ) = β X i , t

Ordinal variables with fewer than six categories were assumed to follow categorical distributions with J categories:

Pr ( y i = 1 | X i , t ) = Λ ( τ 1 − β X i , t ) Pr ( y i = j | X i , t ) = Λ ( τ j − β X i , t ) − Λ ( τ j − 1 − β X i , t ) Pr ( y i = J | X i , t ) = 1 − Λ ( τ J − 1 − β X i , t )

where Λ is the c.d.f. of the standard logistic distribution and the τ s are ‘cut-point’ parameters that indicate the value of the latent variable at which the observed variable changes categories. After the models converged we sampled at least 1K values of the posterior distributions for X_i. We use the means and variances of the draws from the posteriors of the X_is to sample values of the latent variables in the predictive model. ¹³

Cross-sectional covariation

As discussed above, we expect contagion among attacks: at the yearly level outbidding will produce highly variable numbers of attacks from country to country. Researchers commonly employ the negative binomial model for count data with such characteristics. This model allows the variance of the dependent variable to be larger than the mean by including an additional parameter relative to the Poisson model. In our data we have the additional complication of panel-specific overdispersion. That is, not only are the data overdispersed relative to a Poisson distribution, but the extent to which the variance in the counts exceeds the mean varies by country. To address this issue we estimated a Poisson-lognormal mixture model with random effects for overdispersion (Cameron & Trivedi, 1998; Winkelmann, 2008). The Poisson-lognormal model allows us to estimate the expected number of counts per country-year as a multiplicative function of the covariates described above and a country-specific, normally distributed random effect.

The predictive model is:

y i , t ∼ P o i s s o n ( λ i , t ) log ( λ i , t ) = α i + X i , t − 1 β + Z i , t − 1 γ α i ∼ N ( μ i , σ 2 ) μ i = α 0 + W i δ

where the α _i s are the country-specific overdisperson parameters, X is a matrix of latent variables, and Z is a matrix of observed variables. ¹⁴ In the second-level equation, the α _i s are modeled as a function of time-invariant country-level characteristics, W_i. ¹⁵ During estimation values of X, the set of latent variables, are drawn from their respective posterior distributions from the measurement models. This allows us to account for the uncertainty in the latent variable estimates. ¹⁶

Parameter estimates

Table II displays the means, standard deviations, and 95% credible intervals for the regression coefficients. All four of the Behavior variables, all five of the Institutions variables, and two of the four Socio-economic structure variables have posterior densities for which at least 95% lies above or below zero: Macroeconomic performance and Ethno-linguistic fractionalization appear not to influence terror attacks. All but two of the variables that do have an impact increase the incidence of terror: both Veto players and the size of the Coercive bureaucracy are associated with lower numbers of terror attacks. Government coercion, Nonviolent dissent, Violent dissent, International conflict, Contestation, Participation, liberal democratic institutions (Assoc./press/rel./speech), Physical quality of life, and the size of the Population are all positively associated with the number of terror attacks that occur in a country.

Table II also reports, σ ², the variance of the country-specific random effects, which serve as country-specific overdispersion parameters. The variance is 4.30, which suggests that the amount overdispersion varies quite a bit across countries. ¹⁷ The posterior distributions of the overdispersion parameters (α_i) are displayed in Figures 1 and 2. These figures allow us to see where overdispersion is most prominent, and thus where contagion effects are most likely. The largest estimate is that for Peru, which has a random effect statistically distinguishable from the rest of the countries in the analysis. This suggests strong contagion, and is consistent with accounts of competition and outbidding between the Shining Path and other leftist groups during the 1980s (Ron, 2001; Kydd & Walter, 2006). The second and third largest random effects are for El Salvador and Colombia, which is also consistent with a contagion hypothesis. Like Peru, El Salvador and Colombia both experienced large numbers of attacks during periods of violence between the state and several groups with similar ideologies that were competing for the population’s support, circumstances which Kydd & Walter (2006) identify as conducive to outbidding. ¹⁸

OPEN IN VIEWER

Table II.

Summary of coefficient posterior distributions

Variable	Coef.	Std Dev.	5% CI	95% CI
Behavior
Govt. coercion t − 1	0.18	0.01	0.15	0.20
Nonviolent dissent t − 1	0.05	0.01	0.04	0.05
Violent dissent t − 1	0.06	0.01	0.05	0.06
Int’l violence t − 1	0.05	0.02	0.02	0.08
Institutions
Contestation t − 1	0.27	0.03	0.21	0.32
Participation t − 1	0.12	0.04	0.05	0.19
Veto t − 1	−0.21	0.02	−0.26	−0.16
Assoc./press/rel./speech t − 1	0.12	0.03	0.06	0.18
Coercive bureaucracy t − 1	−0.20	0.02	−0.24	−0.16
Varying socio-economic structures
Macroecon. performance t − 1	−0.03	0.04	−0.13	0.07
Physical qual. of life	0.56	0.04	0.49	0.63
Population	0.10	0.01	0.10	0.11
Sticky socio-economic structures
Ethno-ling. frac.	0.92	0.56	−1.21	2.02
Random effects
σ2	4.30

Model fit

In order to assess model fit we present the Deviance Information Criterion (DIC) (Spiegelhalter et al., 2002) for different model specifications. Similar to the AIC, models with smaller DICs are more able to accurately reproduce the observed data. We present the DIC for each of five model specifications: the full model (including all variables), a model with only behavioral variables, a model with only institutional variables, a model with only structural variables, and a model with only random intercepts. Table III shows that the full model offers a large improvement in fit over the intercept-only model, and also permits us to evaluate the relative impact that each block of variables has upon model fit. In comparison to the null model, the block of dissident and state Behavior variables make the largest contribution to model fit, the political Institutions block has the second largest influence, and the block of Socio-economic structures has the smallest. This is powerful evidence in favor of the focus upon dissident and state behavior that we advocate.

Substantive effects

To evaluate the impact of individual variables we take 1,000 random draws from each of the coefficient distributions, producing 1,000 predicted values (for each observation) for which all of the Xs are set to their observed value. We do this a second time, setting the X of interest to its observed value plus one standard deviation. ¹⁹ We then take the difference between these counts. ²⁰ Table IV reports the median and 95% credible intervals of these differences for all country-years in the sample. Summarizing these changes hides considerable variation across countries. In the Online appendix we report the expected change, with a measure of uncertainty, for each country in the sample.

Political institutions

All five institutional variables exhibit a substantial impact on the use of terror, though Participation and Freedom of association/press/religion/speech have smaller impacts than the other three.

OPEN IN VIEWER

Figure 1.

Country-specific overdispersion parameters, top half

OPEN IN VIEWER

Figure 2.

Country-specific overdispersion parameters, bottom half

OPEN IN VIEWER

Table III.

DIC for variable blocks

Model	DIC
Random intercepts	109,072
Structural block	107,481
Institutions block	106,225
Behavior block	102,128
Full model	87,674

OPEN IN VIEWER

Table IV.

Summary of distribution of change in predicted count

Variable	2.5 P-tile	Median	97.5 P-tile
Behavior
Govt coercion	0.01	0.23	14.32
Violent dissent	0.01	0.17	10.07
Nonviolent dissent	0.01	0.18	10.73
Int’l violence	0.01	0.09	5.58
Institutions
Contestation	0.01	0.47	29.05
Participation	0.01	0.16	10.53
Veto	−17.75	−0.29	−0.01
Assoc./press/rel./speech	0.01	0.17	10.71
Coercive bureaucracy	−17.25	−0.28	−0.01
Varying socio-economic structures
Population	0.05	4.00	243.41
Macroecon performance	−2.88	−0.01	0.75
Physical quality of life	0.02	1.23	75.46

Table IV indicates that Contestation is associated with an increase in the expected number of terror attacks, offering support for Chenoweth’s (2010) argument that political competition stimulates all types of political activity, including the use of terror. Increasing the value of Contestation by one standard deviation, while holding all other variables at their observed values, produces a median increase in the number of attacks of 0.47. This cuts against the more commonly advanced hypothesis that institutions which lower the public’s costs of entry into the political arena reduce grievances and discourage the use of violent tactics to promote policy change. Similarly, participation has a positive impact.

Turning to Veto, Table IV indicates that increasing the value of the veto points variable by one standard deviation decreases the number of terror attacks by 0.29, contrary to results reported in Li (2005) and Young & Dugan (2011).

Our measure of Freedom of association, press, religion, and speech has a positive impact on the number of terror attacks a country experiences, which is consistent with results reported in previous studies by Eubank & Weinberg (1994), Lai (2007), Chenoweth (2010), and Bell et al. (2014).

Our final institutional variable, capacity of the Coercive bureaucracy, exhibits a negative relationship with terror tactics: the median effect across all country-years is a reduction of 0.28 terror attacks. This provides support for our conjecture, based on Gurr’s (1988) argument about garrison states, that a large coercive bureaucracy indicates that past state repression was successful in quelling dissent, and that the state is likely to respond to dissent with repression in the present, and should thus discourage the use of terror by dissidents.