What do we mean by the effect of a mixture as a whole?

Abstract

BACKGROUND AND AIM: The analysis of health effects of exposure mixtures is a critically important issue in human epidemiology. One element of mixtures epidemiology that we may want to know is the causal effect of a mixture as a whole. While seemingly a straightforward question, there are nuances that complicate it. Here we present issues that arise in forming causal questions about mixtures as a whole. METHODS: We use Directed Acyclic Graphs (DAGs) and the math of linear regression to discuss the conceptualization of a mixture as a whole and implications for causal inference both when assumptions about residual bias are met and when they are not. RESULTS:How a mixture as a whole is conceptualized and thus represented in a regression model modifies the causal question being addressed. For example, the concentration of each of several correlated toxicants in blood (component parts mixture) vs. the level of the (typically external) source of the toxicants that is responsible for their correlation (whole mixture). Under relatively simple data structures with no confounding these two conceptualizations are similar, but still not the same. When residual confounding of the individual component parts may be present, analyzing the external whole mixture effect has advantages over analyzing the sum of the component parts; in turn interpreting the sum of the component parts has advantages over interpreting the components separately. However, some confounding structures for the components of a mixture are intractable to typical statistical approaches to mixtures analyses, although extensions of g-computation may help. CONCLUSIONS:How one conceptualizes a mixture has implications for the interpretation of analytic results and identifying causal effects that could underlie interventions. Careful consideration of potential biased pathways related to the components of a mixture is critical and has implications for analyzing the data and interpreting the results. KEYWORDS: mixtures, causal inference, exposome, interventions, statistical analysis