How (not) to study suicide terrorism

Today is the 20 year anniversary for 9/11. That made me look into one of the most salient methodological discussions on how to study suicide terrorism within political science.

Suicide terrorism is a difficult topic to study. Why? Because we cannot learn about the causes (or correlates) of suicide terrorism from only studying cases of terrorism. Pape (2003) studies 188 suicide attacks in the period 1980-2001. He concludes that there is a strategic logic to these attacks, namely that they pay off for the organisations and groups pursuing such attacks.

Ashworth et al. (2008) use simple statistics such as conditional probabilities to show that there are problems with the paper in question, namely that the original paper “samples on the dependent variable.” I especially liked this formulation in the conclusion: “It is important to note that our critique of Pape’s (2003) analysis does not make the well-known point that association does not imply causation. Rather, because Pape collects only instances of suicide terrorism, his data do not even let him calculate the needed associations.”

Pape (2008) provides a reply to the critique raised by Ashworth and colleagues. He first brings a long excerpt from his book not taking the critique of Ashworth et al. into account. Then, he writes: “One might still wonder whether the article is flawed by sample bias because it considered systematically only actual instances of suicide terrorism. The answer is no, for two reasons. First, the article did not sample suicide terrorism, but collected the universe of suicide terrorist attacks worldwide from 1980 through 2001. […] There is no such thing as sample bias in collecting a universe. Second, although it is true that the universe systematically studied did not include suicide terrorist campaigns that did not happen, and that this limits the claims that my article could make, this does not mean that my analysis could not support any claims or that it could not support the claims I actually made.”

Importantly, just because you might have the universe of suicide terrorist attacks, you should still treat it as a sample (especially if you want to make policy recommendations about future cases we have not seen yet). In other words, this is a weird way of defending your flawed analysis. In an unpublished rejoinder, Ashworth (2008) provide some additional arguments to why the response to the criticism is flawed. Also, Horowitz (2010) shows that when you increase the universe of cases, Pape’s findings do not hold.

The debate is more than ten years old but reminiscent of similar contemporary debates on data and causality. Accordingly, I find it to be a good read for people interested in research design, data and inference — and it’s a good case to discuss what can (not) be learned from ‘selecting on the dependent variable’. Last, and most importantly, if you want to understand this amazing tweet, it is good to be familiar with the debate.