In observational studies, propensity scores are commonly estimated by maximum likelihood but may fail to balance high-dimensional pretreatment covariates even after specification search. We introduce ...