Co-exposure confounding and amplification of bias—an exploration with practical interpretation and perspectives

Epidemiologists increasingly estimate associations between mixtures of correlated co-exposures and health outcomes. We propose an exploratory approach to help differentiate co-exposure confounding from co-exposure amplification of bias (CAB) for studies considering correlated co-exposures using a case example and simulations based on observed relationships. Our example uses associations observed in published studies between co-occurring per- and polyfluoroalkyl substance (PFAS) biomarkers and tetanus antibody concentrations but could apply to any context where statistical modeling is performed with and without inclusion of correlated co-exposures. Using directed acyclic graphs (DAGs), to define an assumed causal structure, and simulations of three PFAS predicting antibody titers, we show that if researchers can assume one component in the exposure mixture is correlated with the other components, but does not cause the outcome, then CAB can be detected by testing whether the estimate for the non-causal component in a multi-pollutant model is different from zero. In our simulation, the estimate for the non-causal component was zero (95%CI, -0.22, 0.25) when CAB was absent and -0.25 (95%CI, -0.48, 0.00) when present. These results illustrate a possible method to identify CAB and select statistical models that may produce less biased effect estimates if our assumed DAG is correct.