Keywords: Hierarchical Data; Non-parametric Bootstrapping; Unbalanced data; Bootstrap Re-sampling Techniques, Clinical Trials, Pre-Stratified Sampling Procedures, Binomial Sampling
Introduction: In clinical trials, we often compare two treatments w.r.t. survival rate using non-parametric bootstrap (NPB) techniques. Simple NPB is likely inappropriate sampling technique as it treats all the observations as independent. However, clinical trials data often include multiple sources of variation, due to the pre-stratified sampling scheme with several strata, such as, hospitals in multi-center study, or different studies in meta-analysis. Adjusting for such variations is vital in deciding which sampling schemes one must use for NPB to avoid bias. Methods: Using a randomized trial data and simulations, we compare two re-sampling strategies, namely, a) bootstrapping on the highest level and b) bootstrapping on the lower levels to determine which of the two performs better.
Results: The randomized trial comparing two chemotherapy treatments, as well as, the simulation study, indicates that bootstrapping on the highest level is better than the one on lower level.
Discussion: Our findings suggest that bootstrapping on higher level is more robust to model misspecifications and may be versatile for wider applications.