Population-Based Assessment of the Health Effects of Climate Exposure Using Hyperlocal Predictive Models

We aim to develop a daily temperature model on 30m road segments for all roads in the contiguous United States from 2000 to 2020. We will use satellite data, land use features, elevation, etc. as predictors in a stacked machine learning approach which will be trained on tens of thousands of weather stations to estimate hyperlocal predictions. We will in turn use this model to identify urban heat islands, hot and cold spots, and identify populations most vulnerable to suboptimal temperature exposure. This model will be made publicly available to all stakeholders including community members, researchers, and policy makers to help examine and address structural environmental injustice in exposure to temperature and its health effects. Next, we will use this exposure data in epidemiological studies to see whether short- and long-term temperature exposures are associated with health outcomes using causal modeling and mixtures approaches. We will conduct a case-crossover analysis looking at short-term exposures and geocoded deaths in thirteen US states (to residence in most states and census block groups in the rest). For long-term exposure, we will use a variation of a causal difference in-difference model which can account for both measured and unmeasured confounding between spatial units over time to assess how changes in area-level temperature affect changes in rates of morbidity and mortality. This will be followed up by a grouped weighted quantile sum approach which can group correlated exposures and identify their joint effects as a mixture as well as the contribution of each individual component. We will look at several simultaneous mixtures: climate exposures, air pollution exposures, and socioeconomic exposures. Finally, we will conduct extensive sensitivity and subgroup analyses. We will use negative controls assure that associations are not due to confounding. We will correct for measurement error using regression calibration, stratified by state and season. We will examine how the associations are modified by race/ethnicity, poverty, measures of social vulnerability and deprivation, and land use characteristics. Our morbidity data includes all state in-patient data and emergency department visits from several states across years. This allows us to examine the effects of temperature on non-fatal outcomes across the life span. Our large population-based datasets ensure that we will have enough power to detect associations in our subgroup and stratified analyses. This project will greatly enhance our understanding of the full health effects of exposure to temperature in numerous population groups using causal methodology and mixtures approaches.