| Abstract: | It is widely recognized that many of the samples we use for statistical analysis in international politics are the result of some selection process. Not surprisingly, selection models are becoming increasingly popular. At the same time, the role of strategic interaction has begun to play a more important role in statistical analyses. However, it has not been clear how statistical strategic models and selection models relate to each other, or what the effects are of employing one when the other is the more appropriate model. In this article, I 1) clarify why international relations scholars cannot shield themselves from selection bias simply by assuming their results are limited to a given sample; 2) show how recent statistical strategic models relate to traditional selection models and generalize the two sets of models by deriving a correlated strategic model; and 3) examine the effects of misspecifying either correlated errors or strategic interaction. My results indicate that failure to model the strategic interaction produces worse specification error than failure to account for correlated disturbances. In fact, traditional bivariate probit models appear to be superior only when states are almost completely uncertain about each others' preferences. |
