Causal Inference Methods for Evaluation of Large-Scale Environmental Policy Effects under Complex Treatment Interference

Abstract

The evaluation of large-scale environmental policy effects is often complicated by the intricate mechanisms through which environmental exposures affect populations. In causal inference, these treatment structures result in methodological challenges for effect estimation and generating relevant policy insights. This dissertation discusses considerations for addressing these challenges and proposes novel approaches for causal inference under complex interference settings. In Chapter 1, we examine the wide-reaching effects of the COVID-19 lockdowns in the United States on concentrations of fine particulate matter (PM2.5). We employ a two-stage procedure in which we obtain robust lockdown-attributable effect estimates using a synthetic controls approach, and then investigate the potential geographical, mobility, and socioeconomic drivers associated with the estimated effects. This study found that the effect of COVID-19 lockdowns varied dramatically across the country, with evidence that environmental policies aimed at limiting individual-level behaviors may not alone be sufficient to consequentially lower particulate matter exposure. In Chapter 2, we propose an approach for robust estimation of heterogeneous treatment effects under a complex interference scenario known as bipartite network interference (BNI), in which the units a treatment is imposed on are disjoint from those that outcomes are measured on. In environmental policy, treatment effect heterogeneity is an important consideration given the heightened susceptibility of minority and marginalized groups to pollution exposure-related health impacts. We design a novel empirical Monte Carlo simulation approach to evaluate the performance of estimators in this setting, and demonstrate how our proposed estimators can be used in conjunction with subgroup discovery methodology to identify subgroups whose effects differ from the population average without \textit{a priori} specification. Through extensive simulations, we empirically assess our proposed estimators under various outcome and misspecification scenarios. Then, we apply these approaches to investigate the effect of emissions control interventions installed on coal-fired power plants on ischemic heart disease hospitalizations among older Americans. In Chapter 3, we expand the scope of BNI methodology to introduce an approach for quasi-experimental studies with panel data under the BNI setting. Motivated by the need for robust estimates of the health impacts of emissions control technologies on power plants over time, we propose a causal inference framework for difference-in-differences (DiD) analysis under BNI with staggered treatment adoption. Using a data reconfiguration and mapping strategy, we define estimands and conduct analyses intuitively at the intervention unit level, eliminating the need to arbitrarily define outcome unit-level treatments. We also propose an approach for reframing estimates to the outcome unit level, maintaining the ability to obtain more policy-relevant insights and interpretations of treatment effects. We apply the proposed framework to study the impact of power plant flue gas desulfurization technology installations over the period 2003-2014 on coronary heart disease hospitalizations among Medicare beneficiaries.