Vector Autoregression and the Study of Politics

In many respects political scientists agree about how best to model political processes. But we disagree about how to translate our theories into structural equations; each of us seems to have our own structural equation model of the same theory. This disagreement is a serious impediment to theory building. Vector autoregression (VAR) is a means of circumventing this problem. We explain the logic of this alternative modeling strategy and examine its relative virtues. In particular, VAR and the more familiar structural equation (SEQ) approaches are compared in terms of their epistemological underpinnings, empirical power, and usefulness in policy analysis. This comparison shows that the two modeling strategies are based on different conceptions of theory and of theory building and that, for the four-six variable systems we usually study, the choice between VAR and SEQ models presents a trade-off between accuracy of causal inference and quantitative precision, respectively. In addition, VAR models have the disadvantage of being unable to incorporate multiplicative and nonlinear relationships as easily as SEQ models. But VAR models have the advantage of providing a more complete treatment of policy endogeneity than SEQ models. These and other contrasts in the two modeling strategies are illustrated in a reanalysis of Alt and Chrystal's (1983) permanent income model of government expenditure.