A New Integrative Method for Multigroup Comparisons of Censored Survival Outcomes in Multiple Observational Studies

While comparing multiple groups of subjects, covariate imbalance generates confounding and results in biased comparisons of outcomes. Weighting approaches based on propensity scores facilitate unconfounded comparisons in an implicit target population. However, existing approaches cannot (i) directly or efficiently analyze multiple observational studies with three or more groups, (ii) provide meaningful answers, because their target populations may differ considerably from the population of interest, or (iii) deliver precise inferences using censored survival outcomes for a wide variety of estimands. We design generalized balancing weights and construct target populations that incorporate researcher-specified characteristics of the larger population of interest, synthesize information from the different cohorts of subjects, and appropriately compensate for any over- or under-represented groups to achieve covariate balance. The constructed target population, termed concordant target population is agnostic to specific estimators, estimands, and outcomes because it maximizes the effective sample size (ESS) to deliver inferences for different estimands in a realistic natural population. For finding the concordant population, theoretical results identify the global maximum of ESS for a conditional target density, allowing the remaining iterative procedure to involve straightforward parametric optimization. Simulation studies and analyses of TCGA databases illustrate the proposed strategy's practical advantages relative to existing weighting techniques