Weighted quantile sum (WQS) regression approaches for mixture analyses in environmental epidemiology

Epidemiological studies of environmental exposures face a fundamental problem in identifying specific bad actors in wide arrays of potential toxicants. The present study examined how 34 types of dioxin, furan, and polychlorinated biphenyls (PCBs) affect leukocyte telomere length (LTL), a biomarker of chronic disease, particularly cancer. We developed two methods of weighted quantile sum (WQS) regression to identify the effects of overall mixtures as well as individual chemicals’ effects. Methods: Two ensemble approaches were implemented for WQS analyses of the 2001-2002 NHANES dataset, which included 1,330 subjects with PCB exposures and LTL data. The first approach, hereby WQSB, estimated the weights of individual chemicals with a bootstrap step that averaged across 100 bootstrap samples including the full array of 34 PCBs, dioxins, and furans. We contrasted this method with a subsetting approach (WQSS) that estimated individual weights from 1000 sample subsets of 6 chemicals randomly drawn from ~1.3 million possible subsets, given 34 chemicals overall. Results: With WQSB we found a significant positive mixture effect (p<0.0001) with increasing exposure concentrations and LTL. This method further allowed us to identify the proportional contribution of each chemical to the overall effect, revealing that 74.04% of mixture effects were attributable to only 6 chemicals, and of these 3 were furans.. The WQSS approach also revealed a mixture effect (p<0.0001), but only one of the furans strongly contributed to this. Discussion: Two ensemble approaches to WQS regression revealed significant, positive associations between PCB, dioxin, and furan exposures and LTL. The WQSB technique revealed that most of these effects were driven by a small subset of chemicals, particularly furans. The WQSS approach may be less sensitive to multicollinearity in datasets with a large array of components, given that furan exposures were highly correlated