Cross-Cohort Mixture Analysis: A Data Integration Approach With Applications on Gestational Age and DNA-Methylation-Derived Gestational Age Acceleration Metrics

Abstract

Data integration of multiple studies can provide enhanced exposure contrast and statistical power to examine associations between environmental exposure mixtures and health outcomes. Extant research has combined populations and identified an overall mixture–outcome association, without accounting for differences across studies. We extended the Bayesian Weighted Quantile Sum (BWQS) regression to a hierarchical framework to analyze mixtures across cohorts. The hierarchical BWQS (HBWQS) approach aggregates sample size of multiple cohorts to calculate an overall mixture index, thereby identifying the most harmful exposure(s) across cohorts; and provides cohort-specific associations between the overall mixture index and the outcome. We showed results from 10 simulated scenarios including four mixture components in three, eight, and ten populations, and two real-case examples on the association between prenatal metal mixture exposure—comprising arsenic, cadmium, and lead—and both gestational age and epigenetic-derived gestational age acceleration metrics. Simulated scenarios showed good empirical coverage and little bias for all HBWQS-estimated parameters. The Watanabe–Akaike information criterion showed a better average performance for the HBWQS regression than the BWQS across scenarios. HBWQS results incorporating cohorts within the national Environmental influences on Child Health Outcomes (ECHO) program from three different sites showed that the environmental mixture was negatively associated with gestational age in a single site. The HBWQS approach facilitates the combination of multiple cohorts and accounts for individual cohort differences in mixture analyses. HBWQS findings can be used to develop regulations, policies, and interventions regarding multiple co-occurring environmental exposures and it will maximize the use of extant publicly available data.