No-harm calibration for generalized Oaxaca–Blinder estimators

Summary In randomized experiments, adjusting for observed features when estimating treatment effects has been proposed as a way to improve asymptotic efficiency. However, among parametric methods, only linear regression proven form an estimate of the average effect that is asymptotically no less efficient than treated-minus-control difference in means regardless true data generating process. Randomized assignment provides this do-no-harm property, with neither truth model nor generative outcomes being required. We present general calibration method confers same no-harm property onto estimators leveraging broad class nonlinear models. This recovers usual regression-adjusted estimator ordinary least squares used, and further noninferior using methods such logistic Poisson regression. The resulting are both difference-in-means have not undergone calibration. show our equivalent inverse-probability-weighted logit link predicted potential covariates. simulation study, we demonstrate common without procedure may perform markedly worse calibrated unadjusted means.