Addressing Spatially Structured Interference in Causal Analysis Using Propensity Scores

TL;DR

This paper develops propensity score methods to estimate direct and spillover causal effects in spatially structured interference scenarios, such as air pollution impacts on health, addressing violations of traditional causal assumptions.

Contribution

It adapts social network causal assumptions and proposes propensity score-based techniques to estimate interference effects in spatial epidemiology, validated through simulations and applied to air pollution data.

Findings

Propensity score pruning and matching improve effect estimation accuracy.

Detected protective direct and spillover effects of PM2.5 on lung cancer.

Methods outperform traditional approaches in spatial interference settings.

Abstract

Environmental epidemiologists are increasingly interested in establishing causality between exposures and health outcomes. A popular model for causal inference is the Rubin Causal Model (RCM), which typically seeks to estimate the average difference in study units' potential outcomes. An important assumption under RCM is no interference; that is, the potential outcomes of one unit are not affected by the exposure status of other units. The no interference assumption is violated if we expect spillover or diffusion of exposure effects based on units' proximity to other units and several other causal estimands arise. Air pollution epidemiology typically violates this assumption when we expect upwind events to affect downwind or nearby locations. This paper adapts causal assumptions from social network research to address interference and allow estimation of both direct and spillover causal effects. We use propensity score-based methods to estimate these effects when considering the effects of the Environmental Protection Agency's 2005 nonattainment designations for particulate matter with aerodynamic diameter less than 2.5 micrograms per cubic meter (PM2.5) on lung cancer incidence using county-level data obtained from the Surveillance, Epidemiology, and End Results (SEER) Program. We compare these methods in a rigorous simulation study that considers both spatially autocorrelated variables, interference, and missing confounders. We find that pruning and matching based on the propensity score produces the highest probability coverage of the true causal effects and lower mean squared error. When applied to the research question, we found protective direct and spillover causal effects.