Expanded Discussion of Drone Strike Databases from Chapter 3

In Chapter 3, I briefly compared the three major databases on the US drone campaign in Pakistan to help establish its relatively discriminate nature. Here I provide some more background on these databases, including their methodologies and results.

One active database on US drone strikes in Pakistan is maintained by the New America Foundation (NAF), an American think tank that has developed a significant national security focus. NAF’s drone database draws on major Pakistani newspapers (e.g., Dawn, The Express Tribune, The News) as well as credible international news outlets (e.g., AFP, Associated Press, Reuters) for its information. The database relies on at least two reputable sources to verify each strike, recording the number of “militant,” “civilian,” and “unknown” casualties in these sources and providing a range whenever there is variation. As of this writing, the database records 414 US drone strikes in Pakistan, resulting in 2,366–3,702 total casualties, with an estimated 245–303 of these listed as civilians and another 211–328 listed as unknown. If we treat all unknown casualties as civilians, this would yield a ratio of 4.6 militants for every civilian killed using the averages of the ranges provided (this goes up to roughly 9:1 if we omit all unknowns as suggested in Plaw and Fricker Reference Plaw and Fricker2012).

Another major active strike-tracking database is maintained by the Bureau of Investigative Journalism (BIJ), a non-profit news organization based out of the UK. The BIJ database also relies on reputable Pakistani and international media reporting, supplementing these sources with leaked government documents, academic research articles, and independent fieldwork visits to Pakistan. The database uses a minimum of four different sources to verify each strike and presents narrative descriptions of each event as well as active links to all media sources used to investigate it. Overall, it is increasingly seen as the most transparent, complete, and reliable publicly available drone strike database by academics (Bauer, Reese, and Ruby Reference Bauer, Reese and Ruby2021). To date, the BIJ database records 430 US drone strikes in Pakistan, resulting in 2,515–4,026 total casualties, of which 424–969 are civilian. This database has higher civilian casualty counts than the NAF database largely because it counts anyone reported as “tribesmen” and anyone under eighteen as civilians (meanwhile, the NAF classifies tribesmen as unknown and uses fourteen as the age of adulthood). Still, using the means of the BIJ ranges produces a similar militant-to-civilian casualty ratio of 3.7:1.

A third major active, publicly available source of drone strike data comes from the UMass DRONE Database Project, an academic dataset from the University of Massachusetts-Dartmouth. Like the other sources discussed earlier, this database uses mostly Pakistani and international media outlets, but instead of treating all credible sources as equally valid, it prioritizes the highest-quality reporting in terms of (1) level of detail, (2) range of sources, and (3) recency of publication (Plaw and Fricker Reference Plaw and Fricker2012). Like the NAF database, the UMass Drone Database Project classifies casualties as “militant,” “civilian,” and “unknown,” and provides ranges for each type of casualties based on the sources used. At the time of this writing, the database was currently being rebranded and moved to a new online home, but I was able to obtain its most recent figures from the individual managing that transition who graciously shared the data with me. The database contains 452 drone strikes in Pakistan resulting in 3,250 overall deaths, with 179 of these civilian and 486 unknown. Once again, if we (conservatively) treat all unknown individuals as civilian, this yields a militant-to-civilian casualty ratio of around 3.9:1 (or 14.4:1 if we exclude unknowns).

Finally, a fourth major source of drone strike information that was popular for many years but is no longer actively maintained is from the Long War Journal (LWJ), an online foreign policy outlet affiliated with the think tank The Foundation for the Defense of Democracies. The LWJ also relies on Pakistani and international press reports as well as its own reporting, which draws heavily on US intelligence sources. This database has limited summary statistics available online, but as mentioned earlier it is unfortunately no longer actively maintained. As of May 1, 2017, when I was last able to access the database, the LWJ database recorded 392 American drone strikes in Pakistan with 2,799 militants and 158 civilians killed. The LWJ database thus provides a militant-to-civilian casualty ratio of 16.7:1, significantly greater than the NAF and BIJ estimates. Because the database is not very transparent in its methodology, is no longer active, and relies heavily on government sources, however, the estimate should be treated with appropriate caution and as an upper bound in terms of its militant-to-civilian casualty ratio.

Alternative Modeling for Analysis of Drone Support in Chapter 3

Table A3.1 replicates Chapter 3’s analysis of Pakistani drone attitudes from 2009 to 12, but with ordered logistic regression models rather than ordinary least squares or linear regressions. As is clear, the sign and significance of all of the key predictors remains unchanged, with perceptions of excess collateral damage exhibiting a strong negative association with Pakistani drone support. Additionally, as in the linear regression results, the coefficients on this variable are the largest of the three drone related items and among the largest in the models overall. To see this more clearly, I calculated different variables’ average marginal effects on the probability of holding each value of the dependent variable – that is, of viewing American drone strikes as a very bad, bad, good, or very good thing. This reveals that “turning on” the collateral damage variable – that is, going from a zero to a one on the variable – increases the probability of harboring a strongly negative view of drones by 23 percentage points. In contrast, the largest shift associated with turning on the military necessity variable for any level of the DV is eight percentage points, and it is just four percentage points for the sovereignty violation variable. This is consistent with the relative sizes of the linear regression coefficients on these variables in the main text.

Supporting Information and Analyses for Chapter 4

Figure A4.1 Distribution of survey responses and relevant violent events across Iraq

Additional Analyses for Chapter 5

Figure A5.1 Predicted probability of discernment confidence with alternate measure of temporal experience