An approach to impute missing covariates in observational studies when estimating treatment effects with propensity score matching

In many observational studies, researchers estimate treatment effects using propensity score matching techniques. Estimation of propensity scores are complicated when some values of the covariates are missing. We can use multiple imputation to create completed datasets, from which propensity scores can be computed; however, we may be sensitive to the accuracy of the imputation models. We propose an iterative approach to multiply impute the missing covariates that gradually winnows the set of control units, until only the matched control units remain. This approach reduces the influence of control records far from the treated units’ region of the covariate space on the estimation of parameters in the imputation model, which can result in more plausible imputations and better balance in the true covariate distributions. In addition, this approach can be conveniently implemented together with standard multiple imputation software for missing data. We illustrate the benefits of this approach with simulations and with an observational study of the effect of breastfeeding on the child’s educational outcomes later in life.