|
Abstract:
|
In observational studies, the confounding variables among different treatment groups are usually not balanced. The generalized propensity score method has been used to balance covariates, including confounding variables, so that the treatment effect could be appropriately estimated. The generalized propensity scores are usually estimated by a parametric method such as multinomial regression, or a nonparametric method such as classification and regression tree (CART), pruned CART, random forest, and boosting method. There is no general consensus on which method is the optimal in balancing all the covariates, thus leading to appropriate estimate for treatment effect. The optimal method is more likely data dependent. We propose using the rank aggregation method to obtain the optimal method for estimating the generalized propensity scores in terms of balancing all the covariates. After the generalized propensity scores are estimated, the stratification method and inverse probability weighted method are applied to estimate the treatment effect. Extensive simulations are carried out to examine the performance of these methods.
|
