A comparison of different techniques for handling missing covariate values in propensity score methods

In observational studies, confounders may bias the estimates of the causal effects. Propensity score adjustment is commonly used to correct these estimates for such bias. The presence of missing data is another problem that often characterises the statistical analysis. The aim of our contribution is to discuss the performance of the four main methods for dealing with confounding based on the propensity score when combined with different techniques for imputing missing values according to the Multiple Imputation approach. The discussion will be based on the insight achieved from an extensive simulation study, which embraces multiple scenarios, on the estimation of the ATT