Bayesian Hierarchical Meta-Analysis Using Individual Participant Data for Modeling Heterogeneous Dropout Patterns Across Multiple Clinical Trials

Abstract

Missing data is a pervasive issue, notably in clinical trials and prospective studies employing a longitudinal design. This problem becomes particularly pronounced when dealing with data from multi-study clinical trials. Many established research teams undertake multiple clinical trials in closely related domains, each of which may exhibit distinct patterns of patient attrition/dropout. However, due to the shared group panel and study administration, these dropout patterns are often believed to exhibit similarities across these trials. While models addressing single dropout mechanism have been extensively investigated, the analysis of heterogeneous dropout patterns remains understudied. To leverage heterogeneous data and integrate information from multiple missing mechanisms, the first part of this dissertation discusses a new meta-analysis strategy based on individual participant data (IPD) to model observational-level dropout patterns over multiple trials and improve statistical inference via a Bayesian Hierarchical Model (BHM). Extensive simulation studies were conducted to demonstrate the superiority of the new method over existing methods in terms of reduced bias, smaller estimation variability, and higher statistical power. The proposed method was also applied to 13 clinical trials for schizophrenia research exploring demographic and clinical determinants of dropout. Most clinical trials enroll adults and treatment efficacy data in children and adolescents have become scarce. The second part of this dissertation deals with network meta-analysis in examining efficacy of schizophrenia treatments in underage population. Traditional meta-analysis is restricted to direct comparison of treatments in parallel study designs. However, network meta-analysis is a special kind of meta-analysis for the comparison of multiple treatments simultaneously in a single analysis by combining direct and indirect evidence. The direct evidence is obtained from randomized control trials with direct comparison of treatments in parallel study design. The indirect evidence is obtained from comparisons of one or more common comparators by transitivity. The direct and indirect pooled evidence is the network estimate. Network meta-analysis is vital in examining the effectiveness of different treatments in randomized control trials. Accordingly, it is gaining popularity.