Randomized controlled trials (RCTs) provide strong internal validity compared with observational studies. However, selection bias threatens the external validity of randomized trials. Thus, naive RCT results may not apply to either broad general public or narrow populations, such as specific insurance pools. This dissertation concentrates on using propensity scores (PSs) to generalize results from an RCT to a target population.

We are interested in two types of target populations: finite and infinite populations. We propose a simulation framework that imitates the environment composing a randomized trial and target population, i.e., the single-layer simulation construction for an infinite target population and the double-layer construction for a finite target population.

We develop a model-free inverse probability weighted estimator (IPWE) to estimate the average treatment effect in a target population and propose several variance estimation methods, including parametric estimation methods and bootstrap-based methods for different types of target populations. In general, we show via simulation study that the variance methods perform equally well, but the bootstrap-based methods have more robust performance under extreme cases. Moreover, the variance estimation methods always perform better for their designated target population under extreme cases.

In addition to IPWEs, we study model-based estimators including regression-based and survey regression-based estimators. With weights, the survey design-based estimators perform similarly to the regression-based estimators. However, the performance of the model-based average treatment effect estimators is vulnerable to the specification of outcome analysis models. Thus, IPWEs are more robust, especially under extreme cases.

We also propose estimating PSs with separate models for the treatment and control arms of the trial, rather than estimating PSs together (treatment and control groups in trials combined). We find that for many common situations, IPWEs with weights estimated separately perform better than those with weights estimated together (with less bias and less variance) if the trial is randomized.