Explaining Interstate Conflict and War: What Should Be Controlled for?

Most multivariate models designed by analysts of intemational conflict focus on one key explanatory factor and include several control variables. There are prominent norms or customs in the subfield of international politics regarding the construction of multivariate models and the selection of control variables. Several of these norms or customs may make the results of multivariate analyses confusing and difficult to interpret. Analysts typically do not, for example, distinguish between confounding and intervening variables even though the implications and impacts of such variables are substantially different. Most researchers also fail to distinguish between confounding variables and variables that have an impact on interstate conflict that is complementary to that of the key explanatory factor. Commonly, control variables are included in a model for no other reason than that they also have an impact on interstate conflict or some other outcome variable. In some recent analyses, "independent" variables are included that are related by definition to the key explanatory variable, or to each other. This practice introduces into multivariate models artifactual, misleading degrees of statistical association between variables related to each other by definition with tautological relationships masquerading as empirical causal connections that complicate the interpretation of results. Finally, the construction of pooled cross-sectional, time series analyses is consistently based on the assumption that the key explanatory factor, as well as the control variables, have essentially identical impacts on interstate conflict across space, and over time. Substantial evidence, some of which is provided in this paper, suggests that this assumption is unwarranted. This paper provides five guidelines for the construction of multivariate models that address these issues in a manner aimed at making the results of multivariate analyses more intelligible and credible.