3.4 Advantages and limitations of regression approaches
Together with underlying some of the limitations of single and multiple regression in evaluating the effects of environmental mixtures on health outcomes, primarily due to the main problem of multicollinearity, this section has also introduced techniques that overcome such limitation while remaining embedded in a regression framework. Among these techniques, review articles and simulation studies agree in concluding that penalized regression consistently outperformed conventional approaches, and that the choice of what method to use should be selected based on one-by-one situation. A recent paper from Agier et al. (2016) provided a systematic comparison of methods based on regression in exposome-health analyses.
In practical settings, several research questions can be addressed by using multiple regression or its extensions. Nevertheless, there might be research questions that are beyond the reach of regression techniques and for which some additional methodologies should be considered.
- Assessing the overall mixture effect.
Penalized approaches addressed the issues of collinearity and high-dimension by operating some sort of variable selection. While this allows retrieving information on the actual effects for each selected component, addressing other questions such as the ones related to the overall effect of the mixture can not be evaluated. As discussed in Section 1, this is a relevant research question that is often of primary interest in environmental epidemiology. The next section will address this problem, introducing the weighted quantile sum (WQS) regression framework as a technique to evaluate the overall effect of an environmental mixture while taking into account high levels of correlation.
- Complex scenarios with several exposures and interactive mechanisms.
When the mixture of interest is composed by several exposures, it is likely that the mixture-outcome association will involve non-linear and interactive mechanisms. As the number of potential predictors gets higher, so does the complexity of the model. In such situations the performances of regression-based approaches are generally weak, and more flexible algorithms should be taken into considerations. These problems will be assessed in section 5, introducing Bayesian Kernel Machine Regression as a flexible non-parametric approach to estimate the mixture-outcome association in the presence of complex non-linear and interactive mechanisms, and then discussing techniques for the assessment of high-dimensional interactions, including machine learning algorithms based on trees modeling.