Missing data interpolation in integrative multi-cohort analysis with disparate covariate information

Integrative analysis of datasets generated by multiple cohorts is a widely-used approach for increasing sample size, precision of population estimators, and generalizability of analysis results in epidemiological studies. However, often each individual cohort dataset does not have all variables of interest for an integrative analysis collected as a part of an original study. Such cohort-level missingness poses methodological challenges to the integrative analysis since missing variables have traditionally: (1) been removed from the data for complete case analysis; or (2) been completed by missing data interpolation techniques using data with the same covariate distribution from other studies. In most integrative-analysis studies, neither approach is optimal as it leads to either loosing the majority of study covariates or challenges in specifying the cohorts following the same distributions. We propose a novel approach to identify the studies with same distributions that could be used for completing the cohort-level missing information. Our methodology relies on (1) identifying sub-groups of cohorts with similar covariate distributions using cohort identity random forest prediction models followed by clustering; and then (2) applying a recursive pairwise distribution test for high dimensional data to these sub-groups. Extensive simulation studies show that cohorts with the same distribution are correctly grouped together in almost all simulation settings. Our methods' application to two ECHO-wide Cohort Studies reveals that the cohorts grouped together reflect the similarities in study design. The methods are implemented in R software package relate.