Advancing Causal Inference in Climate Disaster Epidemiology: Evidence from Wildfires and Power Outages

Abstract

As climate change accelerates and reshapes the ways in which our environment affects human health, the frequency and severity of wildfires have been steadily increasing. Unfortunately, climate disasters such as wildfires cannot be fully prevented, so we must turn to mitigation strategies instead. To target such policies and interventions appropriately, we must weigh the full range of impacts of climate disasters and tradeoffs we make with different mitigation strategies — economic and logistical, but also human health and societal. Just as important is understanding the direct health and healthcare utilization effects of a climate disaster as it unfolds, which is critical to mounting a timely, well-resourced response -- for example, anticipating surges in emergency department and inpatient care in the days following a wildfire. Often, strategies intended to mitigate the effects of a devastating climate disaster not only have the potential to themselves cause harm, but end up coinciding with the very disaster they were initially intended to prevent. This is often the case with public safety power shutoffs (PSPS) meant to prevent wildfires. Understanding the health effects in these cases -- of the climate disaster itself, its mitigation strategies, and the co-occurrence of the two -- as well as developing the methodological tools needed to rigorously study them, is essential. However, challenges arise in studying health effects of climate disasters. Climate disasters cannot be randomized, there is often no obvious comparison population, and associational studies, while important, are often insufficient to drive policy change. An association alone does not establish that an exposure causes an outcome, and intervening on an exposure that is merely associated with, rather than one that causes, an outcome can result in inefficient resource allocation and limited health improvements. In response, this dissertation applies quasi-experimental, self-controlled study designs -- which build a counterfactual from an exposed population's own data rather than a separate comparison group -- to estimate the causal effects of wildfires and PSPS events on healthcare utilization, and translates these methods into open-source, accessible software. The first aim used a time-stratified case-crossover design to estimate the independent and joint effects of PSPS events and wildfire fine particulate matter ( PM2.5) on acute healthcare utilization in California from 2013--2019. Leveraging circuit-level outage data and modeled wildfire PM2.5 estimates, this analysis found that severe PSPS events were associated with elevated odds of same-day respiratory and chronic obstructive pulmonary disease (COPD)-specific emergency department and hospital encounters, with evidence of multiplicative interaction between PSPS events and wildfire PM2.5 for these outcomes. Associations were strongest among individuals with COPD and adults over 65 years. The second aim used a two-stage interrupted time series (ITS) approach with machine learning to estimate the effects of the January 2025 Los Angeles fires on healthcare utilization among 3.7 million Kaiser Permanente Southern California members. Using National Oceanic and Atmospheric Administration HYSPLIT trajectories and evacuation zone data to separate PM2.5 and evacuation exposure, this analysis found that cardiovascular and neuropsychiatric emergency and inpatient visits increased in heavily smoke-exposed areas in the week following ignition, while respiratory visits declined, and overall care-seeking fell in evacuated areas, with greater reduction in care-seeking among those living in the Eaton vs. Palisades evacuation zones. The third aim generalized the two-stage ITS with machine learning approach into its2s, an open-source Python package. its2s provides built-in model architectures, configurable defaults, and a reproducible workflow -- including versioning, seed management, and parallelization -- for constructing counterfactuals in the absence of comparison populations. The utility of the package and its features are showcased in two case studies, one using a simulated policy intervention and another examining excess injury mortality following the 2021 Pacific Northwest heat dome. This package is intended to lower the technical barrier to applying this method broadly across environmental epidemiology and other disciplines. Together, these three aims demonstrate that self-controlled quasi-experimental designs can generate causal evidence on the health effects of wildfires and power outages even where no randomized trial is possible or a comparison population is available. Wildfires themselves drive substantial cardiorespiratory and neuropsychiatric morbidity, though these effects can diverge by outcome, timing, and exposure pathway -- smoke and evacuation do not affect all forms of care-seeking the same way or on the same timeline. PSPS events -- despite being designed to prevent wildfires -- carry their own distinct health harms when the power goes out, impacts that may be compounded when PSPS events co-occur with wildfires nonetheless. This evidence points to specific, actionable gaps in current mitigation policy -- the need to weigh the health harms of PSPS events against the wildfire risk they are designed to prevent (aim 1), the need to target protections toward populations who are most vulnerable (e.g., individuals reliant on electricity-dependent durable medical equipment) (aim 1), and the need to prepare healthcare systems for these shifting cardiorespiratory and neuropsychiatric care needs in the wake of wildfires (aim 2). To do this, we need both the empirical evidence and the methodological tools to target such efforts appropriately (aims 1, 2, and 3). As wildfires, co-occurring power outages, and other climate disasters continue to intensify and expand geographically, this dissertation's methodological and software contributions provide tools for generating the causal evidence needed to guide policy to protect population health.