Sparse group lasso weighted quantile sum regression for analyzing chemical mixture effects: bidirectional effect estimation with two-stage group-level evaluation.

Chemical mixture studies often require methods that account for predefined exposure groups, prioritize active components within correlated groups, and allow positive and negative associations in the same analysis. We developed Sparse Group Lasso Weighted Quantile Sum regression (SGL-WQS), an index-based method that combines WQS regression with sparse group lasso regularization. SGL-WQS estimates group- and direction-specific indices in training data and evaluates downstream associations in held-out data as exploratory conditional summaries. Evaluation included a large-scale simulation; 30-seed studies of correlation, component prioritization, sample size, dimensionality, outcome family, group-size imbalance, and signal heterogeneity; repeated split-stability analyses; benchmark comparisons; and a survey-aware urinary VOC demonstration. In the large-scale simulation, estimates were consistent with the expected positive PCB, mixed-direction metal, and null PFAS patterns. In matched Gaussian simulations, direction assignment improved with sample size and remained stable from p = 50 to 200 under baseline signals, although component attribution and validation-stage conditional informativeness declined with dimensionality. Within-scenario group-size imbalance did not produce a consistent loss of direction accuracy; weak and dense signals had a larger effect across methods. In exactly paired binary/Gaussian global-null analyses, conditional p < 0.05 proportions did not exceed 0.05 (0.025-0.045). In the survey-aware VOC demonstration, SGL-WQS provided group- and direction-specific summaries across four metabolite groups using NHANES subsample weights in the downstream refit. SGL-WQS may therefore be useful for exploratory, group-aware bidirectional prioritization rather than estimation of a single overall mixture effect or formal selective inference.