Enhancing Validity in Observational Settings When Replication is Not Possible

Abstract

We argue that political sciexntists can provide additional evidence for the predictive validity of observational and quasi-experimental research designs by minimizing the expected prediction error or generalization error of their empirical models. For observational and quasi-experimental data not generated by a stochastic mechanism under the researcher’s control, the reproduction of statistical analyses is possible but replication of the data-generating procedures is not. Estimating the generalization error of a model for this type of data and then adjusting the model to minimize this estimate—regularization—provides evidence for the predictive validity of the study by decreasing the risk of overfitting. Estimating generalization error also allows for model comparisons that highlight underfitting: when a model generalizes poorly due to missing systematic features of the data-generating process. Thus, minimizing generalization error provides a principled method for modeling relationships between variables that are measured but whose relationships with the outcome(s) are left unspecified by a deductively valid theory. Overall, the minimization of generalization error is important because it quantifies the expected reliability of predictions in a way that is similar to external validity, consequently increasing the validity of the study’s conclusions.