Bias reduction methods for propensity scores estimated from error-prone EHR-derived covariates