Testing Theories with Qualitative and Quantitative Predictions

Researchers in the social sciences as well as other disciplines rely on statistical models to develop and test theories. The standard approach applies multivariate analysis—usually linear regression—and statistical hypothesis testing to observational data. The researcher may have several independent variables in mind as candidate predictors of the dependent variable; those reaching statistical significance compose the final model. In other situations, a theory is assumed to be correct and regression analysis is used to estimate parameters. Though experiments with random assignment of subjects are recognized as the “gold standard” for research, these are rarely possible in political science. Instead we rely on statistical controls to overcome problems inherent in using nonexperimental data. This paper suggests that our confidence in using statistical methods to construct theories is misplaced and that theory testing is more productive when we combine definitive theory-generated predictions with statistical methods. I begin with a review of problems in the current approach to statistical analysis, then give several examples of how prediction can be improved.