Estimating Heterogeneous Causal Effects under Bipartite Network Interference: An Application to Air Pollution Regulatory Policy

Marginalized groups are exposed to disproportionately high levels of air pollution. In this context, robust evaluations of the heterogeneous health impacts of air pollution regulations are key to justifying and designing maximally protective future interventions. Such evaluations are complicated by two key issues: 1) much of air pollution regulatory policy is focused on intervening on large emissions generators while resulting health impacts are measured in exposed populations; 2) due to air pollution transport, an intervention on one emissions generator can impact geographically distant communities. In causal inference, such a scenario has been described as that of bipartite network interference (BNI). To our knowledge, no literature to date has considered how to estimate heterogeneous causal effects with BNI. First, we propose, implement, and evaluate causal estimators for subgroup-specific treatment effects via augmented inverse propensity weighting and G-computation methods in the context of BNI. Second, we design and implement an empirical Monte Carlo simulation approach for BNI through which we evaluate the performance of the proposed estimators. Third, we use the proposed methods to estimate the causal effects of flue gas desulfurization scrubber installations on coal-fired power plants on ischemic heart disease hospitalizations among 27,312,190 Medicare beneficiaries residing across 29,034 U.S. ZIP codes. While we find no statistically significant effect of scrubbers in the full population, we do find protective effects in marginalized groups. For high-poverty and predominantly non-white ZIP codes, scrubber installations at their most influential power plants, when less-influential plants are untreated, are found to result in statistically significant decreases in IHD hospitalizations, with reduction ranging from 6.4 to 43.1 hospitalizations per 10,000 person-years.