Measuring Terrorism

Our main dependent variable is the number of terrorist attacks per country-year observation. We apply the inverse hyperbolic sine transformation to this variable to accommodate the influence of outliers; importantly, and in contrast to the log transformation, the inverse hyperbolic sine transformation is also defined for country-year observations with no terrorist activity (e.g., Burbidge, Magee, and Robb Reference Burbidge, Magee and Robb1988).Footnote ⁵

The terrorism data are drawn from the Global Terrorism Database (GTD) described in LaFree and Dugan (Reference LaFree and Dugan2007).Footnote ⁶ The GTD collects information on terrorist activity from reputable media outlets. For a terrorist event to be recorded, it must be documented by at least one high-quality media source and meet the following three criteria: it must (1) be intentional, (2) entail some level of violence or threat of violence, and (3) be committed by non-state actors, meaning that violence by state actors is excluded. Furthermore, the incident must meet at least two of the following three criteria: (1) it must be carried out to achieve a political, economic, religious, or social goal, (2) there must be evidence of an intention to coerce, intimidate, or convey some other message to a larger audience than the immediate victims, and/or (3) the incident must be outside the context of conventional warfare (LaFree and Dugan Reference LaFree and Dugan2007).

Measuring Corruption

Our measure of corruption is the political corruption index from the Varieties of Democracy Dataset (VDEM; Coppedge et al. Reference Coppedge, Gerring, Knutsen, Lindberg, Teorell, Altman and Bernhard2019). Higher values of this index correspond to higher levels of political corruption. This political corruption index is the arithmetic mean of four variables measuring corruption in the (1) executive, (2) legislature, (3) judiciary, and (4) public sector. It covers corruption in the various branches of government and at various levels of government.Footnote ⁷ The corruption index accounts for corruption aimed at influencing policy- and law-making as well as the implementation of these policies and laws. Finally, it covers different forms of corruption, accounting for both “passive” corruption (such as taking bribes) and “active” corruption, for example, in the form of the embezzlement of public resources by public officials and politicians.

VDEM relies on country- and subject-based expert opinion. For instance, to evaluate the extent of legislative corruption, experts are asked to assess to what extent members of the legislature abuse their position for financial gain. To arrive at representative values of political corruption per country-year observation that can also be compared between countries, VDEM then applies item response theory and subjects the individual expert opinion data to other forms of statistical scrutiny to minimize uncertainty and bias (Coppedge et al. Reference Coppedge, Gerring, Knutsen, Lindberg, Teorell, Altman and Bernhard2019).

Empirical Model

To examine the effect of corruption on terrorist activity, we estimate the following model:

(1) $$ \begin{array}{c}terroris{m}_{i,t}=\beta \times corruptio{n}_{i,t-1}\\ {}\hskip8.4em +\hskip2pt \delta {X}_{i,t-1}+{\alpha}_i+{\tau}_t+{\varepsilon}_{i,t},& \end{array} $$

where the (inverse hyperbolic sine transformed) number of terrorist incidents ( $ terrorism $ ) in country i and year t is a function of the country’s political corruption ( $ corruption $ ) in the previous year ( $ t-1 $ ), a set of controls $ X $ as well as country- and year-fixed effects ( $ \alpha $ and $ \tau $ , respectively) to account for time-invariant factors (e.g., culture and norms that affect corruption and/or terrorism) and global time trends.

Below, we consider both a parsimonious and different covariate-adjusted models. Here, the choice of the baseline controls follows the literature on the determinants of terrorism (e.g., Piazza Reference Piazza2008; Krieger and Meierrieks Reference Krieger and Meierrieks2011; Campos and Gassebner Reference Campos and Gassebner2013; Enders, Hoover, and Sandler Reference Enders, Hoover and Sandler2016; Gaibulloev, Piazza, and Sandler Reference Gaibulloev, Piazza and Sandler2017; Gaibulloev and Sandler Reference Gaibulloev and Sandler2019). We include controls for population size and (inflation-adjusted) per capita income. Data on these variables come from the World Development Indicators (WDI; World Bank 2019); both variables are also inverse hyberbolic sine transformed to account for skewness. We also control for democracy, using data from Gründler and Krieger (Reference Klaus and Krieger2016).Footnote ⁸ Finally, we employ an index of state failure from the Political Instability Task Force (PITF 2019) that indicates the extent of large-scale civil warfare and other forms of political instability (e.g., coup d’états).

Instrumental Variables Approach

The estimates from Equation 1 might be affected by endogeneity bias due to measurement error in the corruption variable, the omission of relevant variables in our empirical model or feedback/reverse causation. To address these endogeneity concerns, we leverage a two-stage least squares (2SLS) IV model of the following form:

(2) $$ \begin{array}{l}corruptio{n}_{i,t}={\beta}_1\times regcorruptio{n}_{i,t}\\ {}\hskip6.4em +\hskip2pt {\delta}_1\times {X}_{i,t}+{\alpha}_{1,i}+{\tau}_{1,t}+{\varepsilon}_{1,i,t},& \end{array} $$

(3) $$ \begin{array}{c}terro{r}_{i,t}={\beta}_2\times {\hat{corruption}}_{i,t-1}+{\delta}_2\times {X}_{i,t-1}\\ {}\hskip-1em +\hskip2pt {\alpha}_{2,i}+{\tau}_{2,t}+{\varepsilon}_{2,i,t},& \end{array} $$

where the first-stage regression (Equation 2) predicts potentially endogenous country-year corruption levels using our IV, regcorruption. The predicted country-specific corruption levels are then used in the second stage to explain terrorism (Equation 3). Importantly, the panel structure of our data and the inclusion of country- and year-specific effects can already ameliorate some concerns about a lack of causal identification. Country-fixed effects can account for certain unobserved factors that correlate with our instrument and increase terrorism through means other than increasing national corruption. Year-fixed effects can factor in the possibility that there are changes over time that are spuriously correlated with both the instrument and terrorism.

Instrument Relevance and Exclusion Restriction

For our IV (the mean level of political corruption in geographically and economically proximate countries) to be valid, it should be sufficiently strong. In our case, regional exposure to corruption should predict political corruption. Indeed, considerable empirical evidence suggests that corruption in proximate countries has predictive power in explaining local corruption (e.g., Becker, Egger, and Seidel Reference Becker, Egger and Seidel2009; Borsky and Kalkschmied Reference Borsky and Kalkschmied2019; Correa, Jetter, and Agudelo Reference Correa, Jetter and Agudelo2016; Dimant and Tosato Reference Dimant and Tosato2018; see also the related literature on the spatial contagion of economic reforms such as Gassebner, Gaston, and Lamla Reference Gassebner, Gaston and Lamla2011; Simmons and Elkins Reference Simmons and Elkins2004).

We argue that corruption levels ought to correlate across space to a common demand for corruption control (or a common tolerance for corruption) that is specific to geographically and economically proximate countries (but differs between geographically and economically diverse countries). Why do we expect this demand for corruption control to be is similar across proximate countries? First, geographical proximity is expected to coincide with common political histories and cultures. For instance, countries that are geographically close tend to have similar religious histories. Religion may, in turn, affect corruption, for example, by shaping how strongly religious dogma affects government policy with respect to measures that punish immoral (corrupt) behavior (e.g., Dimant and Tosato Reference Dimant and Tosato2018; La Porta et al. Reference La Porta, Lopez-de-Silanes, Shleifer and Vishny1999; North, Orman, and Gwin Reference North, Orman and Gwin2013). Second, the economic component of the instrument ought to reflect similarities in production, economic needs, and preferences. For instance, economically proximate countries share a similar demand for internationally mobile factors of production (i.e., physical and human capital). Given that capital and talent are attracted to low levels of corruption (e.g., Dimant, Krieger, and Meierrieks Reference Dimant, Krieger and Meierrieks2013; Poprawe Reference Poprawe2015; Wei Reference Wei2000), this may explain why industries in economically proximate countries demand similar levels of corruption control. Finally, both geographical and economic proximity make it more likely that tolerance for corruption aligns, for example, due to close informational ties and low information and transaction costs.

Figure 3 shows that the levels of corruption across countries are not independent of each other (Map A). For instance, corruption tends to be much higher in Sub-Saharan Africa as compared to Western Europe. As expected, this interdependence is also reflected in the IV we construct (Map B). What is more, comparing both parts of Figure 3 strongly suggests that exposure to regional corruption ought to be predictive of local corruption levels. Indeed, the simple pairwise correlation between both corruption variables is $ r=0.70 $ ( $ p<0.01 $ ) for the largest possible sample (see also Figure B.1 in the Supplementary Material).

Figure 3.

Corruption across Countries and Exposure to Corruption as Instrument

Note: Map A shows the average level of corruption per country, categorized into quartiles. Map B shows the respective countries’ regional exposure in quartiles, that is, the average level of corruption in economically and geographically proximate countries.

In line with Figure 3, we expect our IV to positively predict local corruption. We assess the strength of our instrument by means of the first-stage F-statistic. The usual rule of thumb to indicate instrument strength ( $ F>10 $ ) has received some criticism for being anti-conservative, meaning that instruments may be weak even if $ F>10 $ (Lee et al. Reference Lee, McCrary, Moreira and Porter2021). Thus, we also report results for the Anderson–Rubin test that is robust to arbitrarily weak instruments (Anderson and Rubin Reference Anderson and Rubin1949; Lee et al. Reference Lee, McCrary, Moreira and Porter2021). A rejection of the Anderson–Rubin test null hypothesis indicates that the coefficient of the endogenous regressor in the structural equation equals zero, which would support the IV estimates. We also report the Anderson–Rubin confidence set (which inverts the Anderson–Rubin test) to further illustrate the trustworthiness of our IV approach in terms of statistical and economic significance (see Stock, Wright, and Yogo Reference Stock, Wright and Yogo2002 for a further discussion).

Figure 3 also speaks to our idea that there is a variable, the demand for corruption control, that is similar across countries that are geographically and economically proximate. However, the corruption control demand variable itself is not observed. Rather, we employ our IV (the mean level of political corruption in geographically and economically proximate countries) to use differences in regional corruption to draw conclusions about differences in this underlying “hidden” variable.Footnote ¹⁰ In arguing that there is an unobserved variable measuring demand for corruption control, we can address the criticism of “spatial instruments” by Betz, Cook, and Hollenbach (Reference Betz, Cook and Hollenbach2018). They argue that instruments that use realizations of endogenous variables in other spatial units are not valid because of simultaneity in the first-stage equation; in our case, local corruption would affect regional corruption and vice versa. However, as the regionally clustered demand for corruption control is causally prior to regional and local levels of political corruption (in that political demand for corruption control induces policy changes related to corruption control), this simultaneity issue does not emerge.

For our IV approach to be valid and causally estimate the effect of corruption on terrorism, the instrument should only affect terrorism via its effect on local corruption. However, as pointed out by Betz, Cook, and Hollenbach (Reference Betz, Cook and Hollenbach2018), there may be various economic, political, and demographic spillovers that could constitute alternative pathways from the instrument to terrorism. For instance, economic downturns in countries that are geographically and economically proximate to the country of interest are expected to correlate with regional levels of corruption (our IV). At the same time, such downturns could spill-over to the country of interest, affecting both local corruption and terrorism by influencing the opportunity costs of non-corrupt and nonviolent economic activities. Such an alternative pathway from regional corruption to terrorism would violate the exclusion restriction.

To address this concern, we implement two additional robustness checks. For one, we use the plausibly exogenous framework of Conley, Hansen, and Rossi (Reference Conley, Hansen and Rossi2012) and developed further by van Kippersluis and Rietveld (Reference van Kippersluis and Rietveld2018). This method allows us to directly examine how plausible violations of the exclusion restriction matter to causal inference. Allowing for violations of the exclusion restriction and still finding that corruption matters to terrorism would raise confidence in our IV approach. For another, we control for a series of observable time-varying shocks that are correlated across countries that are both geographically close and economically similar. For instance, this includes regional levels of economic growth, political instability, and institutional quality. Finding that corruption (instrumented by regional exposure to corruption) affects terrorism even after accounting for factors that might correlate with our IV (and thus potentially account for further transmission channels from our IV to terrorism) would provide evidence in favor of the exclusion restriction.

Main Results

The main empirical results presented in Table 1 can be summarized as follows. First, the OLS models (specifications 1–3) show a positive and statistically significant association between political corruption and terrorism. Second, the effect of corruption on terrorism is more pronounced in our preferred IV models (specifications 4–6). Here, the impact of regional exposure on corruption in the first stage has the expected effect on local corruption and is sufficiently strong, as indicated by the first-stage F-statistic. The additional IV diagnostics are also sound. Third, introducing the baseline controls to the model does not affect our main empirical conclusion that political corruption encourages terrorism. Concerning these controls, terrorism positively correlates with population size, state failure, democracy, and economic development. These associations are also reported in other studies (e.g., Piazza Reference Piazza2008; Krieger and Meierrieks Reference Krieger and Meierrieks2011; Campos and Gassebner Reference Campos and Gassebner2013; Enders, Hoover, and Sandler Reference Enders, Hoover and Sandler2016; Gaibulloev and Sandler Reference Gaibulloev and Sandler2019). However, due to the lack of an identification strategy associated with estimating these associations, they cannot be given a causal interpretation (Keele, Stevenson, and Elwert Reference Keele, Stevenson and Elwert2020).

Table 1.

Effect of Corruption on Terror Attacks

Note: The table presents main specification results (OLS and 2SLS) of the effect of (exposure to regional) corruption on terrorism in the subsequent year. Robust SE clustered at the country level in parentheses (Driscoll–Kraay SE in model 7). Model 8 uses 2-year instead of 1-year lags. Because several countries in the sample gained independence only after 1970, the number of country-year observations varies from 129 in 1970 to 175 in the 2000s. On average, we observe 160 countries per year. Models 1 and 4 show the parsimonious estimates for the largest possible sample. The remaining models restrict the sample to country-year observations without missing control variables to ensure comparability across specifications. CD-test reports the p-values associated with Pesaran’s test for cross-sectional dependence, where the null hypothesis is cross-sectional independence. PURT-test reports the p-values associated with Choi’s Fisher-type panel unit-root test, where the null hypothesis is that all panels contain a unit root (i.e., are nonstationary) against the alternative that at least one panel is stationary. For the PURT-test, we subtract the cross-sectional averages from the series to mitigate concerns about cross-sectional dependence. * $ p<0.10 $ , ** $ p<0.05 $ , *** $ p<0.01. $

Table 1 also reports some diagnostics and initial robustness checks. For one, there may be concerns about the presence of cross-sectional dependence in the regression residuals, which may affect the validity of statistical inference (e.g., Sarafidis and Wansbeek Reference Sarafidis and Wansbeek2012).Footnote ¹¹ The results of a test for cross-sectional independence of the residuals (Pesaran Reference Pesaran2015) show that for some specifications, cross-sectional dependence is indeed present in the residuals, pointing to a potential violation of the assumption of spatial independence of observations. Therefore, we also run a variant of our baseline model using standard errors proposed by Driscoll and Kraay (Reference Driscoll and Kraay1998), which are robust to heteroskedasticity and autocorrelation, but also to general forms of cross-sectional dependence. As shown in specification 7, accounting for residual cross-sectional dependence in this manner produces even smaller standard error estimates. This suggests that our choice of standard errors (i.e., cluster-robust standard errors) produces rather conservative standard error estimates, so that type I errors are less likely to occur. As another way to address the issue of cross-sectional dependence, we also run a common correlated effects regression within a GMM framework following Pesaran (Reference Pesaran2006). As shown in Table C.1 in the Supplementary Material, our main effect of corruption on terrorism is robust to this alternative specification.

For another, there may be concerns about model choice and lag length selection. As a first way to address this concern, we use 2-year instead of 1-year lags for all independent variables. Specification 8 in Table 1 produces results that are very similar to our baseline estimates. In Appendix C.2 of the Supplementary Material, we consider issues of lag length selection more in-depth by discussing and running a series of error-correction models (e.g., Engle and Granger Reference Engle and Granger1987; De Boef and Keele Reference De Boef and Keele2008). These models allow us to examine how the inclusion of lags of the dependent variable, and additional lags of the independent variables, affect our estimates. At the same time, they allow us to address concerns about non-stationarity and cointegration, while also enabling us to study the relationship between corruption and terrorism in the long run. As shown in Table C.2 in the Supplementary Material, our long-run estimates suggest that higher levels of corruption lead to more terrorism. Indeed, these (dynamic) estimates are rather close to our (static but more parsimonious) estimates reported in Table 1, suggesting that the main results are not affected by different lag structures (concerning both lags of the dependent and independent variables) and non-stationarity. This latter finding speaks to the results from panel unit-root tests following Choi (Reference Choi2001), which we report in Table 1 and which suggest that the residuals from our baseline estimates are not affected by non-stationarity, further ameliorating spurious regressions concerns.

Plausibly Exogenous Framework

While the exclusion restriction cannot be tested directly, there are ways to probe it. We first rely on the plausibly exogenous method of Conley, Hansen, and Rossi (Reference Conley, Hansen and Rossi2012). This method relaxes the assumption of perfect instrument exogeneity, instead allowing for violations of the exclusion restriction. The plausibly exogenous methods rests on the following 2SLS model:

(4) $$ \begin{array}{l}corruptio{n}_{i,t}={\beta}_1\times regcorruptio{n}_{i,t}+{\delta}_1\times {\boldsymbol{X}}_{i,t}\\ {}\hskip6.5em +\hskip2pt {\alpha}_{1,i}+{\tau}_{1,t}+{\varepsilon}_{1,i,t},& \end{array} $$

(5) $$ \begin{array}{l}terro{r}_{i,t}={\beta}_2\times {\hat{corruption}}_{i,t-1}+\gamma \times regcorruptio{n}_{i,t-1}\\ {}\hskip4.4em +\hskip2pt {\delta}_2\times {\boldsymbol{X}}_{i,t-1}+{\alpha}_{2,i}+{\tau}_{2,t}+{\varepsilon}_{2,i,t}.& \end{array} $$

Regional exposure to corruption—the instrument—can now directly affect terrorism in the country of interest i in the second stage via $ \gamma $ , meaning that the exclusion restriction is violated; in case of $ \gamma $ being equal to zero, the exclusion restriction would be valid. By considering different values of $ \gamma $ , we can investigate how violations of the exclusion restriction matter to our IV estimates. Below, we use the union of confidence interval approach proposed by Conley, Hansen, and Rossi (Reference Conley, Hansen and Rossi2012). For our analysis, we consider various values of $ \gamma $ , where the upper bound is chosen according to the reduced-form equation. These reduced-form estimates are reported in Panel A of Table 2 and indicate—consistent with our IV approach—that higher levels of regional corruption correlate with higher levels of terrorism in the country of interest. While we assume per the exclusion restriction that this effect is only due to the role of shared tolerance of corruption (approximated by regional corruption) in local corruption, by means of the plausibly exogenous approach, we can assess how robust our findings are to different degrees of instrumental invalidity. As a lower bound, we consider a value of $ \gamma =-1 $ , which would imply that more regional corruption is associated with less terrorism in the country of interest.

Table 2.

Zero-First-Stage Tests

Note: Full sample refers to the baseline sample of 167 countries without missing control variable values; we report the baseline results for comparison. Robust SE clustered at the country level in parentheses. * $ p<0.10 $ , ** $ p<0.05 $ , *** $ p<0.01 $ .

Figure 4 presents the upper and lower bounds for the corruption coefficient, applying the plausibly exogenous approach for our baseline empirical model. There are two conclusions. First, if $ \gamma <0 $ , the upper bound effect of local corruption on terrorism becomes larger. Such a relationship might emerge when regional corruption positively correlates with certain predictors of terrorism. Second, as long as $ \gamma <1.5 $ , the 95% confidence interval does not include zero (for 90% CI, this $ \gamma $ threshold is 1.9). Taken together, the results suggest that $ \gamma $ has to reach a value that is in excess of 42% of the size of the reduced-form estimate for the coverage area to include zero (53% of the reduced form effect size for the 90% CI). This, in turn, suggests that our IV estimates appear to be robust to high degrees of instrumental invalidity.

Figure 4.

Testing Instrument Invalidity via Union of Confidence Interval Approach

Note: The figure shows the upper and lower bounds of the 95% confidence intervals based on the union of confidence interval approach with varying degrees of instrument invalidity.

Another informal test of the exclusion restriction is the so-called zero-first-stage test, which considers a sub-sample for which the first stage is zero (van Kippersluis and Rietveld Reference van Kippersluis and Rietveld2018). For this sub-sample, the reduced form should then also be zero if the exclusion restriction is satisfied. Because there are no countries that saw zero changes in corruption over the 1970–2018 period, we construct a “quasi zero-first-stage group,” for which the responsiveness of local to regional corruption (indicated by the respective standard deviation of local corruption per country) over the observation period is as close to zero as possible.

We consider three quasi zero-first-stage country groups that include 32 countries (approximately one-fifth of the total sample) and 19 countries, respectively. The first group includes countries for which local political corruption was the least responsive.Footnote ¹² Eighteen out of 32 countries were OECD members before 1990, which could raise concerns that this zero-first-stage may not be informative about the corruption–terrorism nexus in non-OECD countries. Thus, we create a second country group that includes those 32 non-OECD countries (nonmembers before 1990) that saw the lowest levels of responsiveness of their local corruption levels.Footnote ¹³ Because this group contains several small nations, we define a third zero-first-stage group comprising of those non-OECD countries with the lowest corruption responsiveness and at least two-million inhabitants. Table 2 shows that the effect of regional corruption on local terrorism (reduced form) is not significant in the quasi zero-first-stage samples (Panel A) and that the effect of regional on local corruption (first stage) is similarly imprecisely estimated (Panel B). These findings indicate that the direct effect of regional exposure to corruption on terrorism is negligible. While the zero-first-stage test can never verify the exclusion restriction, our findings support the notion that this restriction is satisfied.

The results from the zero-first-stage test also provide informative values for $ \gamma $ within the plausibly exogenous framework (van Kippersluis and Rietveld Reference van Kippersluis and Rietveld2018). For “true” zero-first-stage cases, the reduced-form estimated effects of regional corruption on terrorism would be equal to $ \gamma $ .Footnote ¹⁴ For our analysis, we can use the reduced-form estimates for the quasi-zero-first-stage country groups as an input for the plausibly exogenous approach to investigate how such plausible violations of the exclusion restriction affect our 2SLS estimates. Using $ \gamma =0.040 $ (quasi-zero-first-stage countries in column 1 of Table 2) yields a β associated with political corruption of 6.851 [95% CI: 2.281, 11.421]. Similarly, using $ \gamma =-0.218 $ yields β=7.152 [95% CI: 2.334, 11.969] and $ \gamma =-0.446 $ yields β=7.443 [95% CI: 2.335, 12.552]. That is, the implied violations of the exclusion restrictions do not result in the inclusion of zero in the 95% confidence interval associated with $ {\beta}_2 $ , meaning that political corruption exerts an unfavorable effect on terrorism when instrumented by regional corruption under plausible violations of the exclusion restriction. Indeed, the same conclusion can also be drawn from Figure 4, where $ \gamma $ -values of $ 0.040 $ , $ -0.218 $ , and $ -0.446 $ , respectively, clearly do not include zero.

Finally, van Kippersluis and Rietveld (Reference van Kippersluis and Rietveld2018) argue that incorporating uncertainty around the direct effect of the regional corruption on terrorism is meaningful, given potential differences in covariates between the zero-first-stage and full sample. They suggest—as a rule of thumb—fixing the variance in such a way that the normalized difference in direct effects between the (quasi) zero-first-stage and the full sample does not exceed one-quarter in 95% of cases, meaning that the variance–covariance matrix $ \omega $ is given by $ \omega ={(0.125\times \sqrt{s_0^2+{s}_1^2})}^2 $ , where $ {s}_0 $ is the standard error of $ \gamma $ in the quasi-zero-first-stage and $ {s}_1 $ is the standard error of $ \gamma $ for the remaining sample. We can incorporate this uncertainty by applying the local-to-zero approach of Conley, Hansen, and Rossi (Reference Conley, Hansen and Rossi2012). Here, we find that for the first quasi-zero-first-stage country group (with $ \gamma =0.040 $ ), $ \omega =0.655 $ , which yields a $ \beta $ associated with political corruption of 6.226 [95% CI: 0.974, 11.447]. For the first alternative quasi-zero-first-stage country group (with $ \gamma =-0.242 $ ), $ \omega =0.043 $ , resulting in $ \beta $ =6.726 [95% CI: 2.414, 11.038], while for the second alternative group (with $ \gamma =-0.446 $ ), $ \omega =0.140 $ , yielding $ \beta $ =7.168 [95% CI: 2.697, 11.640]. Hence, allowing for uncertainty does not overrule our finding that more corruption leads to more terrorism, where the difference between covariates for the alternative quasi-zero-first-stage country group (from which OECD members are excluded) appears to be especially benign.

Robustness of Instrumental Variable Approach

We consider several additional ways to further probe the robustness of our IV results, focusing on instrument construction and potential violations of the exclusion restriction, which we discuss in detail in Appendix B of the Supplementary Material. First, we show in Table B.1 in the Supplementary Material that the results are robust to a more detailed definition of local exposure (18 instead of 6 UN geographical regions), to using yearly or 1995-fixed WDI income levels and to a continuous measure of geographical proximity (log capital distance) instead of world regions. Each of these alternative IV constructions renders effects of political corruption on terrorism comparable to our main specification. Second, we show that random assignment of baseline instrument values to other countries produces zero results. These placebo checks confirm that the geographical and economic ties between regional and local corruption are essential to the relevance and validity of our baseline IV approach. They speak to the notion that there is an unobserved corruption tolerance variable that is non-randomly distributed across geographically and economically proximate countries, which we can capture with our regional corruption instrument. Third, the exclusion restriction may be violated due to shocks that are correlated within regions and simultaneously affect local corruption and terrorism (e.g., Betz, Cook, and Hollenbach Reference Betz, Cook and Hollenbach2018). To address such concerns, we control for a series of time-varying variables that ought to capture the role of regionally correlated economic, political, institutional, and demographic shocks. Controlling for these regional shocks does not affect our main empirical conclusion, which raises further confidence in our assumption that the exclusion restriction is valid (Table B.2 in the Supplementary Material).

Further Robustness Checks

Having provided evidence that our IV approach is sound, we consider further ways to probe the robustness of our empirical findings in Appendix C of the Supplementary Material. In detail, we show that our findings are also robust to changes to our baseline model, for example, in terms of the operationalization of controls (Table C.1). For example, Enders, Hoover, and Sandler (Reference Enders, Hoover and Sandler2016) and Gaibulloev, Piazza, and Sandler (Reference Gaibulloev, Piazza and Sandler2017) suggest that per-capita income and democracy may share a nonlinear relationship with terrorism; our results are robust to the inclusion of such nonlinearities. Similarly, the interpretation of the results does not change with the inclusion of additional control variables (Table C.4). It is also not affected by alternative measurements of the dependent variable, for example, by measuring terrorism in per capita terms (Table C.5); by considering alternative modes of terrorist attacks (e.g., bombings vs. kidnappings; Table C.6); by differentiating between domestic and transnational terrorism as well as between government and civilian terrorist targets (Table C.7); and by dropping potentially influential cases such as countries which exceptionally high levels of terrorist activity or political corruption (Table C.8). Finally, in Appendix D of the Supplementary Material, we investigate the role of specific subtypes of corruption in terrorism. We find that corruption in the executive, legislative, and judicial branches encourages more terrorist activity, which tends to point to a generalized relationship between political corruption and terrorism. However, the effect of public sector corruption on terrorism—while having the expected sign—is not estimated precisely enough to fully support this notion. For this case, IV diagnostics suggest that our IV is too weak to allow for proper identification.

Exploration of Mechanisms

As another important contribution, we study the potential mechanisms through which political corruption may translate into increased terrorist activity. Above, we discussed three transmission channels: (1) an inadequate provision of public goods (which, in turn, exacerbates economic grievances), (2) stronger political dissatisfaction related to the reduced viability of peaceful political contestation, and (3) reduced counter-terrorism capacity. To provide suggestive evidence on their role, we consider the following 2SLS system:

(6) $$ \begin{array}{c}corruptio{n}_{i,t}={\beta}_1\times regcorruptio{n}_{i,t}+{\delta}_1\times {\boldsymbol{X}}_{i,t}\\ {}\hskip1em +\hskip2.5pt {\alpha}_{1,i}+{\tau}_{1,t}+{\varepsilon}_{1,i,t},& \end{array} $$

(7) $$ \begin{array}{c}m{v}_{j,i,t}={\beta}_2\times {\hat{corruption}}_{i,t-1}+{\delta}_2\times {\boldsymbol{X}}_{i,t-1}\\ {}\hskip-2em +\hskip2pt {\alpha}_{2,i}+{\tau}_{2,t}+{\varepsilon}_{2,i,t}.& \end{array} $$

Employing our usual IV approach and including the baseline set of controls and fixed effects, we investigate the effect of political corruption on six mediator variables (mv). In addition to estimating the causal effect of corruption on these mediators, we also study the correlation between these variables and terrorism.Footnote ¹⁵ The link from the various mediator variables to local terrorism should not be interpreted causally.

First, health inequality and education equality (both VDEM) measure the provision of public goods. Furthermore, these measures relate to their eventual relationship with inequality and exclusion. We expect corruption to reduce health and education equality and lower levels of equality to correlate with higher levels of terrorist activity. Second, we consider political accountability (VDEM) and political protest (Banks and Wilson Reference Banks and Wilson2013). The former variable refers to the extent that citizens, civil society organizations, and the media can hold the government accountable, while the latter measures the number of nonviolent or unorganized political protests (strikes, riots, and anti-government demonstrations). We expect corruption to adversely affect government accountability and lead to more political protest, and low levels of accountability and high levels of political protest to breed terrorism. Third, we employ the indicators military capacity and territorial control using data from the National Material Capabilities Dataset updated from Singer (Reference Singer1988) and VDEM, respectively. The former variable is a composite measure of military capacity that accounts—via principal component analysis—for a country’s total military spending, number of military personnel, and per-capita military spending.Footnote ¹⁶ The latter variable indicates the percentage of territory a state has effective control over, where such control may be disputed by, for example, criminals, warlords, or insurgents. Both variables thus relate to a country’s counter-terrorism capacity. We anticipate corruption to lower this capacity and lower capacity to correlate with increased terrorist activity.

The empirical results reported in Table 3 can be summarized as follows.Footnote ¹⁷ First, political corruption reduces both health and education equality (Panel A). This is consistent with the idea that corruption undermines (e.g., by distorting public investment) the provision of public goods that could otherwise ameliorate socioeconomic inequality. Countries with inadequate public goods provision also see more terrorism (Panel B). For instance, this finding is in line with earlier contributions stressing the inequality–terrorism nexus (e.g., Ezcurra and Palacios Reference Ezcurra and Palacios2016; Krieger and Meierrieks Reference Krieger and Meierrieks2019). Second, we find no evidence that corruption affects political accountability and protest, nor that a lack of accountability correlates with more terrorism. This may indicate that the role of corruption in the viability of peaceful political contestation is less pronounced, at least concerning the political input (accountability) and output (protest) variables we consider in this article. In line with Campos and Gassebner (Reference Campos and Gassebner2013), there is, however, evidence that more political protest is associated with more terrorist activity. Third, political corruption reduces both military capacity and the state’s territorial control. What is more, both variables negatively correlate with terrorism, which speaks to earlier empirical analyses of the unfavorable relationship between state weakness and terrorism (e.g., Lai Reference Lai2007).

Table 3.

Potential Mechanisms

Note: The table explores potential mechanisms by replacing the main outcome (terrorist attacks) with key determinants of public goods (models 1 and 2), political grievances (models 3 and 4), and counter-terrorism capacity (models 5 and 6). OLS and robust SE clustered at the country level in parentheses. * $ {}p<0.10 $ , * $ p<0.05 $ , *** $ p<0.01 $ .

Corruption and Other Types of Conflict

In principle, corruption may also matter to other forms of conflict. Estimating a parsimonious specification of our 2SLS model, however, we find no effect of corruption on incidences of ethnic, civil, and international war (Table D.2 in the Supplementary Material; the conflict data are from Marshall Reference Marshall2019). Moreover, in Table 3, we have already shown that corruption does not affect protest (strikes, riots, and demonstrations). These tentative findings suggest differences between the determinants of terrorism and other forms of political violence (see also, e.g., De La Calle and Sanchez-Cuenca Reference De La Calle and Sanchez-Cuenca2012; Sambanis Reference Sambanis, Keefer and Loayza2008). However, we invite future research to undertake more elaborate assessments of whether the violent consequences of corruption go beyond terrorism.