The possibility that actors strategically condition their behavior on
partially unobservable factors poses a grave challenge to causal inference,
particularly if only some of the actors whose behavior we analyze are at
risk of experiencing the outcome of interest. We present a crisis bargaining
model that indicates that targets can generally prevent war by arming. We
then create a simulated data set where the model is assumed to perfectly
describe interactions for those states engaged in crisis bargaining, which
we assume most pairs of states are not. We further assume
researchers cannot observe which states are engaged in crisis bargaining,
although observable variables might serve as proxies. We demonstrate that a
naïve design would falsely (and unsurprisingly) indicate a positive
relationship between arming and war. More importantly, we then evaluate the
performance of matching, instrumental variables, and statistical backwards
induction. The latter two show some promise, but matching fares poorly.