Matching, weighting, or regression? Evidence from a comprehensive simulation study of Stata treatment effect estimators

Estimating treatment effects with cross-sectional data is one of the most widespread approaches in empirical research. Provided that researchers are able to measure all relevant control variables, it is possible to approximate unbiased (causal) treatment effects. To this end, Stata offers a wide range of standard and user-written commands. Naturally, the question remains which of these methods is most robust for producing unbiased point estimates and valid inference. We address this question by evaluating 14 different commands in a comprehensive simulation study. Using three different settings (unbiased, biased, and incorrect functional form), we analyze a variety of empirically relevant scenarios. Our results indicate that linear (OLS) regression exhibits the lowest bias, the smallest standard errors, and the most accurate coverage in almost all simulation specifications. Entropy balancing and some matching approaches offer advantages when nonlinearities are incorrectly specified. Surprisingly, several methods deviate substantially from the target estimands, even in unbiased "best-case" scenarios.