Grant: Travel Awards - Institute Community Support

Primary findings in randomized controlled trials (RCTs) in cancer are typically reported with confidence intervals (CIs). However, when the subpopulations in the RCTs differs from each other to an extensive degree, CIs shrink to almost zero width and the results of that are not very useful. In contrast, prediction intervals (PIs), as already established in reporting meta analysis findings, do properly reflect effect variation across subgroups. We propose using Bayesian PIs to summarize heterogenous subpopulation effects in RCTs. Utility of the approach is demonstrated using simulation studies and application to trial data from the German Breast Cancer Group (GBSG), the 'GBSG Study' (n=686). With stratification variables tumor grade (I, II, III) and menopausal status (yes/no) in the GBSG Study, the Bayesian PI indicates more useful reporting and results than CI. Bayesian PIs should be routinely used when reporting treatment effect estimates to properly reflect treatment effect heterogeneity in pre-specified subpopulations in RCTs.