A1015
Title: Unbiased multigroup comparisons by integrating multiple observational studies: A new concordant population approach
Authors: Subharup Guha - University of Florida (United States) [presenting]
David Christiani - Harvard TH Chan School of Public Health (United States)
Yi Li - University of Michigan (United States)
Abstract: The effective synthesis of information from multiple observational studies to make meta-analytic comparisons of multiple group responses is a challenging problem. Existing weighting and matching techniques cannot incorporate domain knowledge or directly analyze multiple cohorts with three or more groups (e.g., races), preventing the generalizability of the results to a natural population. We propose a new class of generalized balancing weights that incorporate known attributes of a larger population of interest into the target population, adjusting for under-sampled groups. Optimizing over any unknown attributes, we obtain the concordant target population. For censored outcomes, we propose balance-weighted Kaplan-Meier estimators to calculate confidence intervals and quantiles of the marginal survival curves of multiple groups. We devise small-sample procedures for uncertainty quantification and assess the performance through simulation studies. We apply the method to compare race-specific cancer survival among several TCGA glioblastoma multiforme (GBM) patient cohorts by adapting to the known racial decomposition of GBM in the U.S. population, and find that Blacks are more vulnerable and endure significantly worse prognoses.
