EFFECTS OF MISSPECIFICATION OF THE PROPENSITY SCORE ON ESTIMATORS OF TREATMENT EFFECT

A comparison was carried out between the propensity score and prognostic models in estimating treatment effects from observational studies. One issue investigated was the effect of estimating the propensity score on estimators of treatment effect. It was found that estimating the propensity score introduced no additional bias. A second question addressed comparisons of the propensity score and prognostic approach when a confounder is omitted. The results indicate that biases due to omitted covariates are large and of the same magnitude. Third, misspecifications of the propensity score were compared to misspecified response models. Here it was found that estimators obtained from incorrect response models had much larger biases than estimators from incorrectly estimated propensity scores. In all cases there were two types of models, one involving a continuous and one a binary response. Least squares estimators for the continuous response were compared to stratified mean differences between treatment groups. For the binary response maximum likelihood estimators of the odds ratio were compared to Mantel-Haenszel estimators. In both cases the strata were based on the quintiles of the true and estimated propensity scores.