The New Barbary Wars: Forecasting Maritime Piracy

This paper extends systematic analyses of maritime piracy by verifying the robustness of empirical results and examining the forecasting ability of empirical models. Recent research by Ward, Greenhill and Bakke (2010) finds that statistically significant relationships frequently offer poor guidance when it comes to anticipating the inception of civil war. We assess the predictive ability of purported causal factors of piracy using evaluative statistical tools such as receiver-operating characteristic plots, out-of-sample predictions, and outlier analysis. Statistical results for in-sample and out-of-sample tests show that while factors such as military capacity, population size, coastline length, and trade volumes are statistically related to piracy, state fragility has by far the strongest predictive effect despite only being moderately statistically significant in the models. Outlier analysis demonstrates that while several countries experience higher numbers of piracy incidents than predicted, empirical models are generally robust to the presence of outliers. For policymakers, the findings suggest that counter-piracy efforts focused on capacity-building measures have the greatest potential for reducing the piracy threat.

With 80% of global trade carried by cargo ships plying the world's oceans, piracy presents a serious threat to the international economy. So far, in 2012, there have been more than 250 pirate attacks worldwide, many occurring near strategic chokepoints, such as the Gulf of Aden, the Malacca Straits, and the Arabian Sea. As the international community has begun to grapple with this hazard, the magnitude of the problem has become evident. The oceans of the world are large and difficult to police. Further, the tens of thousands of merchant ships now active cannot reasonably be monitored by the limited naval and coastguard forces currently available. While extensive research has been directed at the causes and correlates of terrorist violence, only a handful of studies have been published that assess the drivers of maritime piracy. Not only is there a need to document the conditions under which piracy occurs, but policymakers require assessments that go beyond statistical significance to provide guidance in anticipating and preventing pirate attacks.

Existing research on piracy points to state weakness, economic opportunity, and geography correlating with pirate incidents, but it is unclear whether such empirical findings by themselves provide policymakers with much help in anticipating the extent or location of maritime violence. This is in part a consequence of probabilistic models ignoring "out-of-sample predictive heuristics" (Ward et al. 2010:364) to help guide forecasting, as well as neglecting a sophisticated analysis of potentially high-leverage outliers.1 Admittedly, social scientists have recognized the risk in relying on measures of statistical significance to accurately predict future events. However, few studies actually assess the robustness of models and individual variables using both in-sample and out-of-sample data. Models built using one set of data should be expected to explain variance in those data fairly well since that is typically what researchers are looking for. It is quite a different challenge for the same model and variables to correlate as strongly with the dependent variable using an entirely different set of cases. "Results that are drawn from robust models," as Ward et al. (2010) write, "have a better chance of being correct." And models that correctly capture the data-generating process should offer policymakers a more trustworthy guide in forecasting future events.

This paper is a part of a larger effort to ensure that models of maritime violence are both robust to model specification and generalizable across time and space. Our analyses here assess three types of model sensitivity. First, we examine the robustness of existing results presented in systematic research on piracy. Second, we evaluate the predictive ability of factors currently found to drive piracy. Third, we consider the extent to which individual cases and covariate patterns influence the observed correlational relationships. The paper proceeds as follows. We first provide a brief historical overview of piracy and the subsequent response by powerful states. We then review research on the drivers of piracy to inform our sensitivity analysis evaluating major explanatory concepts proposed in piracy research. Our predictive analyses use receiver operator characteristic (ROC) plots to assess the forecasting ability of a standard model of piracy. We then examine the influence of individual cases and covariate patterns to determine the generalizability of our results. Our analyses show that existing explanations of piracy succeed in generating useful predictions on patterns of piracy. We also find that among these explanatory factors, measures of institutional strength have the strongest predictive power. Finally, our examination of outliers indicates that the coefficient estimates remain largely unaffected by individual observations and covariate patterns. We conclude with some thoughts on the usefulness of such sensitivity analyses for policymakers.