“Draining the Swamp”: An Empirical Examination of the Production of International Terrorism, 1968–1998

Abstract

One central element of the current war on terrorism is “draining the swamps,” addressing conditions within a state that produce international terrorism. This paper empirically examines what factors lead a state to become a “swamp,” drawing on a theoretical approach that guides current U.S. policy. This theory looks at the ability of a state to impose costs on terrorist groups within its own borders. The lower the operating costs within a state, the greater the amount of terrorism produced within that state. Using data on the number of international terrorist events originating from a state from 1968 to 1998, an empirical model incorporating variables designed to test this theoretical argument as well as relevant control variables is employed. Strong support was found for the state strength approach, suggesting that one way to address the threat of international terrorism is to strengthen a government's ability to control its own territory.

Keywords:

• terrorism
• political violence
• state strength

Acknowledgments

The author would like to thank Glenn Palmer, Emily Lai, Kelly Kadera and several anonymous reviewers for their comments on this paper.

Notes

¹This is from the 2002 U.S. National Security Strategy, Chapter 3, Strengthening Alliances to Defeat Global Terrorism and Work to Prevent Attacks Against US.

²There are numerous other examples including Plan Colombia and U.S. troops in Afghanistan and Iraq, as well as several other states.

³The U.S. has used airstrikes in Yemen, Sudan, Somalia, and Afghanistan (before 2001) to try and reduce the abilities of terrorist groups.

⁴An alternative approach would be to model the production of fatal attacks. This is not used for the following reasons. First, it is not clear that excluding fatal attacks is necessary or useful. Given that acts of terrorism are designed to have an impact beyond the immediate audience, whether an act kills someone or not is not the only intention. For example, when groups like the IRA call in their bombs before detonation, they reduce the likelihood of death but are still attempting to influence politics within their region. Second, looking at fatal attacks ignores the response of target groups. Because this paper looks at the production and not targeting of terrorism, the acts of targets in stopping attacks and preventing deaths could pose a problem. If a group attempts a fatal attack and fails, it is counted as a nonfatal attack. However, conditions within a state allowed for the production of that attack. It did not become fatal because the target was able to stop it. Thus, this paper uses incidents and not just fatal incidents.

⁵While it would be ideal to have data on all types of terrorist events, the ITERATE data is very useful for this analysis, especially given the policy focus of dealing with states that produce international terrorism as opposed to states that produce and are the targets of terrorism. Also, ITERATE picks up some cases of states that are targeted by the terrorism they produce. ITERATE codes issues of separatism (i.e., Northern Ireland, Basques) as cases of multiple nationalities, leading to their inclusion in the data. Finally, no other data source has the temporal or spatial coverage needed to test this hypothesis broadly.

⁶See Mickolus (1982) for a broader discussion of these data.

⁷Occasionally, there are more than ten groups in a state. ITERATE provides a three-digit code similar to originating state code. For example, there are more than ten groups originating from the United States (country code 002). Thus, the White Panthers, a U.S.-based group, has an initial three-digit code of 003. For all of these cases, the groups were researched using a variety of web sources to determine their country of origin. The primary source is the MIPT terrorist knowledge database.

⁸An alternative procedure to code this variable would be to use the home country of the individual terrorists. Because the hypothesis is about conditions that allow a group to operate within a state, this paper uses the location of the group. A model was run using the location of the individuals and the results are essentially the same. The only difference is that the international war variable is no longer statistically significant.

^a 193 states included before exclusion based on missing data; 8 states are removed because of missing data.

¹⁰In addition, the Minorities at Risk (MAR) Economic Discrimination Index, which measures economic policies that discriminate against a minority is tested to see whether there is a difference between these two variables. This variable ranges from 0 (no discrimination) to 4 (restrictive policies). Because the MAR data operate at the group level, for any given state, the highest differential score of all the groups is used to aggregate to the state level. So if a state has three groups, the value used in this analysis is the highest differential score of the three groups.

¹¹One potential concern with using GDP per capita is that this variable might be endogenous with terrorism (Abadie, 2004) as terrorism may influence the economy of a state. While this is a potential concern, there are three reasons that suggest its inclusion. First, it has been used in other research that examines the effect of a state's economy on political violence. Most notably, Fearon & Laitin (2003) present it as a central explanation of whether a state is likely to experience a civil war. Also, it has been used in several studies of terrorism (Li, 2005; Li & Schaub, 2004). Second, lagging the dependent variable by a year should help to address this potential problem by explaining the incidents of terrorism at time t with the GDP per capita at time t−1. Also, because this paper looks at the production and not targeting of terrorism, it is less clear that there is an endogenous relationship between a state's wealth and its level of produced terrorism. States targeted by terrorism may be likely to experience a decline in their GDP, but it is unclear if this relationship is true for producers of terrorism. Finally, while some research (Abadie, 2004) has attempted to address this endogeneity, instrumenting GDP per capita with variables that are also not likely to be influenced by terrorism is difficult. Abadie (2004) is looking only at a cross section and thus uses a state's landlocked status as an instrument. However, this is not as useful for data that also vary across time. Also, Abadie (2004) finds no real difference between the effects of the economy on terrorism between models that use GDP per capita and the instrumented GDP per capita.

¹²One caveat with this variable is that the U.S. started its list in 1979, while the data begins in 1968. Thus, there are 11 years where states that could be state sponsors are not identified. A sample is run that is 1980 onwards to check the validity of the state sponsor results. State sponsors are Iraq (1979–1981, 1990–1998 [end of data]), Libya (1979–1998), Syria (1979–1998), South Yemen (1979–1990), Cuba (1982–1998), Iran (1984–1998), North Korea (1988–1998), and Sudan (1993–1998). One notable country not listed here is Afghanistan. The Taliban did not control Afghanistan until 1996, when bin Laden and Al Qaeda moved there from the Sudan. This variable is examined with and without Afghanistan.

¹³This model is used instead of a Poisson because there is likely to be overdispersion in the data. For example, terrorists are likely to try to plan simultaneous terrorists incidents, either to demonstrate their strength, retaliate against government actions, or wear down the government in order to achieve a more favorable deal (Kydd & Walter, 2003). As such, the occurrence of one terrorist event is likely to be related to the occurrence of other terrorist events in that same period. Also, other studies have used a zero-inflated model (Drakos & Gofas, 2006) to model the potential for under-reported terrorist events. Using a variety of independent variables for the inflation model, Vuong tests indicated that a zero inflated model did not significantly contribute anything more than a negative binomial model, thus this paper presents the results from a negative binomial.

¹⁴Another method to deal with this problem is to use fixed effects for each state. A model was run using this method and of the hypothesized variables, the international war and population variables were no longer significant. However, this model is not presented in the paper because it potentially leads to the removal of useful information because of a lack of variation within a state for the time period under study. For states that produce no terrorism, these are removed from analysis because the fixed effects perfectly predict the outcome (King, 2001). Also, fixed effects have the potential to mask slowly changing or unchanging variables (Beck & Katz, 2001), like the state sponsorship of terrorism variable, which is no longer significant in a fixed-predict effects model.

∗p < 0.05

∗∗ p < 0.01

∗∗∗ p < 0.001 two-tailed tests, Robust standard errors (SE) are reported in parentheses.

¹⁵Also, to examine whether the inclusion of the potentially endogeneous civil war variable is biasing the other estimates, Model 6 in Table 1 is re-analyzed without the civil war variable. The results for the other variables are unchanged.

¹⁶The results for the economic discrimination variable (which starts in 1980) are similar, although the level of statistical significance is less (p < 0.1).

† two standard deviations below mean

†† two standard deviations above the mean

††† mean