Assessing metal mixture effects on neuropsychological development: A trade-off between complexity and interpretability

Several studies have explored the joint effect of mixtures of metals on health outcomes by using sophisticated statistical techniques such as Bayesian Kernel Machine Regression (BKMR). Although these appealing tools can detect complex relationships between metals, their interpretation is not straightforward. Indeed, BKMR is frequently used jointly with other methods, and final conclusions are simplified in terms of increase/decrease of the risk function at fixed values for individual metals. In this paper, we explore the feasibility and interpretability of such techniques to assess the effect of metal mixture exposures on neuropsychological development in early childhood. This is a cross-sectional study. Initially, we use Principal Components Analysis (PCA) to detect main latent neurodevelopment domains. The scores derived by the BKMR exposure-response function were computed and compared with those provided by traditional linear regression models (LRM). Pearson's correlation coefficients of the estimations obtained by BKMR and LRM accounting for pairwise interactions between metals (lead, molybdenum, and selenium) were 0.95, 0.95 and 0.92, for the three identified latent domains: executive functions, motor functions and visual and verbal functions, respectively. The observed differences between both estimations mainly occurred in participants having low concentrations of lead and low and high values of selenium, which suggest linearity in the associations and not high order interactions between metals. We concluded that, in this case, employing linear regression to model the impact of mixtures on targeted outcomes led to similar results than more complex statistical techniques, allowing a better understanding of both individual and interaction effects between metals.