Coping with Uncertainty in Adversarial Risk Analysis

Morgan, Henrion, and Small (1990) present a useful taxonomy of types of uncertainty. They note that uncertainties can arise from: (1) statistical variations (e.g., “random error in direct measurements of a quantity”); (2) systematic errors (e.g., “biases in the measuring apparatus and experimental procedure”); (3) subjective judgments (e.g., for quantities where empirical data are largely unavailable); (4) linguistic imprecision (e.g., translation of verbal phrases into numerical probabilities); (5) variability (e.g., quantities that vary over time or space, or from one person to another); (6) inherent randomness or unpredictability (which cannot be reduced by further research); (7) disagreement (e.g., among multiple experts); (8) approximation (e.g., due to limits in the spatial resolution of a model); and (9) model uncertainty (i.e., uncertainty about the most appropriate model to represent some phenomenon).

In the context of terrorism risk, statistical variation might arise from the sparseness of data on actual terrorist attacks; systematic error could reflect the fact that many terrorist plots are likely to go unobserved; and uncertainty about the most appropriate model (e.g., game theory vs. decision theory vs. risk analysis) is likely to be widespread. Even within a single model, experts may disagree about the relative importance of different terrorist objectives (e.g., fatalities vs. economic damage). These importance weights might also vary over time, and from one terrorist group to another. Thus, good strategies for dealing with uncertainty—both statistical methods for drawing reliable conclusions from sparse data sets (e.g., about whether the frequency or severity of terrorist events is increasing over time; Mohtadi and Murshid 2009), and also methods for dealing with “deep uncertainty” (e.g., about whether terrorist groups are optimizing vs. satisficing; Shan and Zhuang 2013)—are crucially important.