Chapter 3 Regression-based approaches
The previous section described a set of unsupervised techniques for the analysis of environmental mixtures. These are used to process the complex data before further analyses and to address well defined research questions related to the identification of common patterns of exposures or clustering of individuals based on exposure profiles. In the context of environmental health studies, however, the ultimate goal is often to investigate whether exposure to mixtures of environmental factors are associated with a given health outcome, and possibly whether these associations represent causal effects. Epidemiologists are usually trained to address these questions using regression-based techniques such as generalized linear models, for binary and continuous outcomes, or parametric and semi-parametric regression techniques for survival data, for time-to-event outcomes. Nevertheless, environmental exposures often present complex settings that require handling regression with care. The goal of this section is to present the use of classical regression techniques (i.e. ordinary least squares (OLS)) in mixtures modeling, describe its limitations, and introduce some important extensions of OLS that allow overcoming these shortcomings.