Propensity score analysis with partially observed covariates: How should multiple imputation be used?