Correlated confounders in mixtures analysis context
On this page:
Background: Environmental exposures, like air pollutants, are often correlated with each other, and may also be correlated with factors that require analytic control, like social determinants of health, that may act as confounders. Performance of methods for analysis of correlated exposures, or exposure mixtures, with correlated confounders is poorly understood. We characterized performance of two mixtures analysis methods, weighted quantile sums (WQS) and quantile g-computation (QGC), in the context of correlated confounders.
Methods: We simulated data emulating preterm birth and 8 ambient pollutants (benzene, cadmium, lead, toluene, xylenes, phenols, ethylbenzene, hexane) based on the empirical correlation matrix from the 2011 National Air Toxics Assessment for North Carolina. We also simulated correlated years of education and annual income. Simulated log-odds ratios of pollutants on preterm birth were all either null or positive. We fit crude (pollutant), partially adjusted (+ education), and fully adjusted (+ education + income) WQS and QGC models. We compared true values to observed estimates across 500 simulations of N=5000 to estimate mean absolute error (MAE) and mean squared error (MSE).
Results: Simulated pollutants ranged from highly correlated (xylenes:ethylbenzene r = 0.99) to un-correlated (phenol:cadmium r = 0.02). Income was moderately correlated with education (r = 0.45) and was weakly correlated with phenol (r = -0.11). QGC quantile estimates consistently showed lower MSE than WQS on a per-model basis (crude QGC MSE: 0.0059 vs. crude WQS MSE: 0.0085). WQS showed lower MAE than QGC in adjusted and semi-adjusted models (adjusted QGC MAE: 0.0611 vs. adjusted WQS MSE: 0.0286) but not in crude models (crude QGC MAE: 0.0595 vs. crude WQS MSE: 0.0800).
Conclusion: QGC MAE was less variable with adjustment than WQS, indicating that, with adjustment, QGC produces less biased, but more variable, mixture estimates than does WQS.