Approaches to identify the contribution of understudied toxicants in complex metal mixtures

Abstract

Background and Aim. Exposures to complex mixtures are increasingly acknowledged to contribute to adverse health outcomes through complex joint or interactive effects. Emerging advanced statistical approaches facilitate appropriate analyses of complex mixtures by incorporating high correlations, nonlinearities, and interactions among individual exposures. Selection of appropriate analytical approaches for complex mixture analysis, and the interpretation of results, is dependent on the specific scientific goals. This presentation discusses and compares the use of complementary and combined approaches in analyses of metal mixture exposures in the Navajo Birth Cohort Study (NBCS). Methods. We compare results and interpretation from 1) Clustering methods that assess the study participant’s co-occurring exposure patterns; 2) Regression methods with shrinkage and variable selection that identify a subset of exposure mixtures associated with the outcome of interest; 3) Kernel machine regression approaches that model mixture effects by incorporating interaction and nonlinear relationships; 4) Ensemble learning methods that provide outcome prediction by ranking each variable’s predictive power, thereby handling a large number of variables; and 5) Causal analysis methods that provide inference that mimics interventional effects to control for confounders, reducing biased estimates that may limit generalizability. Results. Different approaches used have helped to identify not only the patterns of exposures in the population, but also understudied metals that, in mixtures, are significant contributors to birth outcomes and developmental delays in NBCS. The results obtained from different approaches have affected the interpretation and opened new avenues of investigation for understudied toxicants. Conclusions. Using complex analytic strategies to parse effects of complex exposures supports both understanding of population-specific exposure-response relationships, and also the specific contributors to those responses and associated mechanisms that are generalizable to populations where metal-mixture mineralogic composition, chemistry, doses, and patterns of exposure will vary. Keywords clustering, penalized regression, kernel machine regression, ensemble learning, causal analysis, mixtures, metals