New methods for estimating propensity scores following multiple imputation

Propensity score (PS) analysis can address the issue of obtaining biased estimates of treatment effect due to the influence of confounding variables in an observational study. One of the primary challenges encountered in this analysis is the presence of observations with missing covariates. In such a case, several potential solutions based on the Multiple Imputation (MI) technique were proposed in the literature. These methods combine treatment effect estimates (MIte), or PSs (MIps), or parameter estimates in the PS model (MIpar) in order to estimate treatment effect. Furthermore, the augmenting MIps method has been proposed as a r-fold repetition of the MIps method. This thesis proposes two new methods, augmenting MIte and augmenting MIpar, for combining estimates obtained with the MI method in PS weighting. The aim is to determine the most effective method. In the comparison of the methods, different simulation scenarios were considered to account for the effect of sample size (n=550 and 1100), the effect of the number of imputations, and the r (4 different scenarios as m×r=100), and the effect of different treatment effect values for TOTE and OTE (3 different scenarios). The treatment effect estimate, mean absolute difference, and bootstrap variance of the treatment effect estimate were employed as criteria for evaluating the performance of the methods. The results of the simulation study indicated that the augmenting MIte method demonstrated superior performance in TOTE estimation, while the augmenting MIte and augmenting MIps method exhibited superior performance in OTE estimation. The addition of interaction terms to the imputation model affected the interpretations made according to the performance criteria. However, the results weren't affected by the different sample sizes and numbers of imputation scenarios. Consequently, the augmenting MIte method is recommended in the case of a homogeneous treatment effect in PS weighting, while the augmenting MIte method, in which interaction terms are added to the imputation model, is recommended in the case of a heterogeneous treatment effect.