Matching as Nonparametric Preprocessing for Reducing Model Dependence in Parametric Causal Inference

The paper addresses the problem that parametric causal estimates in published work often depend heavily on discretionary modeling choices such as which control variables to include and which functional forms to use, choices that readers rarely see and that can produce widely varying results. The authors review matching methods, which promise causal inference under fewer assumptions but are commonly misinterpreted, and propose a unified two-step approach in which data are first preprocessed nonparametrically with matching and then analyzed with the parametric techniques a researcher would have applied anyway.

By separating the design stage of matching from the analysis stage of parametric estimation, the procedure makes the resulting causal inferences more accurate and considerably less dependent on arbitrary modeling decisions. The authors also provide easy-to-use software, giving applied researchers a practical way to reduce model dependence while still using familiar parametric tools.