Primary Submission Category: Multilevel Causal Inference
Estimating exposure-response curves in multi-level data: causal pooling of cluster-specific curves
Authors: Jenny Lee, Fabrizia Mealli, Francesca Dominici, Rachel Nethery,
Presenting Author: Jenny Lee*
When studying the health impacts of fine particulate matter (PM_{2.5}) at a national level, one of the key interests of policy-makers may be how different regions contribute to exposure-response function (ERF). Substantial variability on health outcomes across regions after controlling for measured confounders may be due to differences in the unmeasured regional-level variables such as cultural difference, dietary habits, and prevalence of comorbidity. In this paper, we propose pooled-ERF, a method for population-level causal ERF estimation with multi-level data, which allows for adjustment of unmeasured cluster-level confounders, enables estimation of the population-level ERF from cluster-level ERFs even when the raw data for each cluster is inaccessible (i.e. only have cluster level data), and yields an interpretable ERF that elucidates each cluster’s contribution. We compare commonly used ERF estimators across different confounding scenarios and compare performance of pooled-ERF via simulation. We apply pooled-ERF to estimate the average causal ERF between long-term PM_{2.5}exposure and all-cause mortality among $68.5$ million Medicare enrollees 2000-2016 based on the regional level ERFs across four regions in the U.S. (northeast, west, south, midwest). We demonstrate the accuracy and interpretable nature of the results using the tools and visualization techniques enabled by the proposed method, and we compare results to those of other causal inference methods.