Abstract
Estimating the joint effect of a multivariate, continuous exposure is crucial, particularly in environmental health where interest lies in simultaneously evaluating the impact of multiple environmental pollutants on health. We develop novel methodology that addresses two key issues for estimation of treatment effects of multivariate, continuous exposures. We use nonparametric Bayesian methodology that is flexible to ensure our approach can capture a wide range of data generating processes. Additionally, we allow the effect of the exposures to be heterogeneous with respect to covariates. Treatment effect heterogeneity has not been well explored in the causal inference literature for multivariate, continuous exposures, and therefore we introduce novel estimands that summarize the nature and extent of the heterogeneity, and propose estimation procedures for new estimands related to treatment effect heterogeneity. We provide theoretical support for the proposed models in the form of posterior contraction rates and show that it works well in simulated examples both with and without heterogeneity. Our approach is motivated by a study of the health effects of simultaneous exposure to the components of PM$_{2.5}$, where we find that the negative health effects of exposure to environmental pollutants are exacerbated by low socioeconomic status, race and age.
Introduction
In environmental health research, individuals are rarely exposed to isolated chemicals; rather, they exist within a complex "mixture" of pollutants. While causal inference has made significant strides in handling binary treatments (e.g., the presence or absence of a medical intervention), it frequently struggles with multivariate continuous treatments. This challenge is twofold: first, the mathematical complexity of estimating joint effects across multiple continuous variables often leads to the "curse of dimensionality"; and second, treatment effects are rarely uniform across a population. A specific pollutant mixture may pose a significantly higher risk to an elderly individual in a low-income neighborhood than to a younger individual in a more affluent area.
Existing methodologies often rely on the assumption of homogeneity—that the treatment effect is the same for everyone—or fail to provide a rigorous metric to rank which demographic or socioeconomic factors drive differences in risk. This paper addresses these gaps by introducing a flexible Bayesian framework designed to estimate joint effects and quantify treatment effect heterogeneity in multivariate, continuous settings.
Methodology
The Bayesian Framework
The proposed approach utilizes a nonparametric Bayesian model that decomposes the outcome regression into three distinct components:
- Main Effects of Covariates: Capturing how demographic and socioeconomic factors independently influence health outcomes.
- Main Effects of Exposures: Estimating the average impact of the pollutant mixture across the population.
- Interaction Terms (Heterogeneity): Modeling how specific covariates modify the impact of the pollutants.
To implement this decomposition, the authors employ advanced variants of Bayesian Additive Regression Trees (BART). Specifically, SoftBART is used for main effects to ensure smooth, realistic dose-response curves. For the interaction terms, Targeted Smoothing BART (tsBART) is utilized, allowing the model to isolate how individual covariates modify the joint treatment effect.
Multivariate Treatment Effect Variable Importance Measure (MTE-VIM)
A central contribution of this work is the development of the MTE-VIM (denoted as $\psi_j$). This metric represents the proportion of total treatment effect heterogeneity explained by a specific covariate. By providing a standardized ranking, MTE-VIM allows researchers to move beyond the observation that "effects vary" to a precise quantification of which factors—such as race or income—are the primary drivers of vulnerability.
Computational Scalability and Identifiability
To handle the massive computational requirements of nationwide datasets, the authors implemented a blocking scheme to partition data into manageable segments without compromising accuracy. Furthermore, to ensure the model correctly attributes effects, a "shifting" strategy was developed during the Markov Chain Monte Carlo (MCMC) training process. This ensures that interaction terms capture true heterogeneity rather than absorbing average treatment effects.
Data and Application
The methodology was validated using both simulated data and a massive real-world dataset:
- Medicare Population Data: Coverage of all US Medicare beneficiaries (age 65+) across 38,702 zip codes from 2000 to 2016.
- Environmental Data: High-resolution grids of $PM_{2.5}$ components, including ammonium, nitrates, sulfate, elemental carbon, and organic carbon, alongside ozone levels.
- Covariates: A combination of individual-level data (age, race, sex, Medicaid eligibility) and zip-code-level socioeconomic indicators (BMI, smoking rates, income, education, and population density).
Key Results
Simulation Performance
In simulations across scenarios ranging from "no interaction" to "strong heterogeneity," the proposed "SepBART" method outperformed standard models like Bayesian Kernel Machine Regression (BKMR). SepBART achieved the nominal 95% coverage and the lowest Root Mean Squared Error (RMSE), particularly in cases where strong heterogeneity was present—a scenario where traditional models often fail.
Environmental Health Insights
The application to Medicare data yielded several critical findings:
- Consistent Harm: Across the 17-year study period, the model estimated a positive Average Treatment Effect (ATE), confirming that increased exposure to $PM_{2.5}$ mixtures consistently correlates with increased mortality.
- Primary Drivers of Risk: The MTE-VIM identified race, dual eligibility for Medicaid (a proxy for low socioeconomic status), and age as the most significant modifiers of the treatment effect.
- Disparity Amplification: The analysis revealed that the negative health impacts of air pollution are significantly exacerbated in areas with lower socioeconomic status and higher proportions of racial minorities, highlighting a critical intersection between environmental exposure and social inequity.
Discussion and Limitations
While the framework provides a robust tool for environmental epidemiology, certain limitations remain. The MTE-VIM requires the selection of a "reference level" for exposure; while results in this study were robust, poor baseline choices in data-sparse regions could lead to extrapolation errors. Additionally, the model assumes that while covariates modify the treatment effect, they do so additively—meaning it does not explicitly model complex interactions between the modifiers themselves (e.g., the combined effect of being both elderly and low-income).
Future research may extend this framework to binary outcomes (such as hospitalization rates) or count data, further broadening the utility of these importance measures in public health policy and environmental regulation.
Summary updated