Machine Learning remedies to unmeasured confounding biases in environmental mixture studies

SUMMARY To draw causal interpretation and policy recommendations from epidemiological investigations of environmental mixtures, it is key to quantify the impact of residual unmeasured confounding. Rigorous evaluation of this potential threat to causal inference faces several challenges. This application, in response to

PA-20-195, will accomplish the following goals: (1) develop sensitivity analyses approaches to detect and assess the impact of unmeasured confounding bias in the estimation of the overall environmental mixture effect, single pollutant main effect and interactions among mixture components accounting for uncertainty on

unmeasured confounding structure and strength; (2) apply these approaches in the analysis of a birth cohort of Bangladeshi mother-infant pairs to characterize the joint effect of metal exposures on child development; (3) develop and disseminate computationally efficient and user-friendly software for widespread application of the

methods in environmental epidemiology. The proposed work will address methodologic gaps in the causal investigation of health effects. First, most of observational environmental mixture studies are plagued by unmeasured confounding bias. Approaches for quantitative bias analysis to investigate the impact of this bias

either assume a single exposure or a single unmeasured confouder and are therefore inadequate. Second, multicollinearity, skewness of exposures, complex exposure-response relationships and intermediate or time- dependent confounding challenge valid estimation and inference. We propose to fill these methodological gaps

by developing and applying probabilistic sensitivity analyses and negative control exposures approaches for quantifying the causal health effects of environmental mixtures under the counterfactual framework in the presence of unmeasured confounding. We will develop and apply the new methods to estimate the mixture

effect and individual pollutant effect that incorporate quantitative bias analysis for multiple unmeasured confounders and allow for complex exposure-response relationships of metals and dimension reduction adopting the Bayesian multiple index model (Aim 1(a)). The approach will accommodate multiple potential

unmeasured confounding structures and incorporate uncertainty on unmeasured confounding structure and strength (Aim 1(b)). We complement this approach with the negative control exposures approach, which allows to detect unmeasured confounding leveraging alternative assumptions on the confounding structure (Aim 2).

We will investigate the effect of metal mixtures on birth length and child neurodevelopment in a Bangladeshi cohort evaluating the impact of unmeasured confounding bias (Aim 3(a)). We will develop user-friendly and efficient R packages that implement the proposed methods and R shiny apps to allow interactive visualizations

of the results under different hypothesized unmeasured confounding scenarios (Aim 3(b)). Our work has great potential to have broad impact on environmental epidemiology research and beyond.