Harnessing the Power of a Novel Synthetic U.S. Population to Uncover Public Health Insights

Abstract

Marginalized communities in the U.S. must contend with structural discrimination that has systematically placed them in under-resourced neighborhoods whose inherent social, economic, infrastructural, and environmental conditions adversely affect health and their wellbeing. Today, researchers’ understanding of these neighborhoods—characterized by substandard housing, hazardous workplaces, degraded infrastructures, close proximity to toxic sites and polluting facilities, food deserts and fast-food jungles, limited opportunities for quality education and employment, and restricted access to healthcare services and facilities—and their inhabitants rely on incomplete population data fraught with assumptions and biases symptomatic of a long history of residential segregation. Broadly, the two options for available population data are: (1) geographically-aggregated data that operates at a high spatial resolution, but only provides marginal distributions of single characteristics, or (2) individual-level microdata that provides joint distributions of multiple characteristics, but only operates at a low spatial resolution. Thus, researchers aiming to leverage available population data to devise and disperse public health interventions must forgo either an understanding of compounding risks or the ability to pinpoint localized risks. This dissertation aims to construct, validate, and demonstrate the utility of a novel synthetic U.S. population which models 330 million agents that collectively reflect the real population with respect to key social, economic, infrastructural, and environmental characteristics at the census tract level. Generated through spatial microsimulation methods that avail the strengths of both geographically-aggregated data and individual-level microdata, it affords researchers the ability to answer previously unanswerable questions about large-scale, complex public health issues across the U.S. First, we constructed the model using iterative proportional fitting (IPF)—a widely-favored static deterministic reweighting method for spatial microsimulation. We then validated it against the 2018-2022 American Community Survey (ACS), the 2020 Decennial Census, and the 2020 Residential Energy Consumption Survey (RECS) with a suite of metrics including Pearson’s R2 (mean = 0.98), standard absolute error (SAE) (mean = 0.03), and the Kullback-Leibler (KL) divergence statistic (mean 0.08), finding that our synthetic U.S. population exhibited strong overall performance in simulating the real U.S. population with variation due to the size and diversity of census tracts and the complexity of linking variables. Second, we examined the well-established compounding risk faced by children under the age of five who we know—based on our synthetic U.S. population—live in homes built before 1980 to map residential lead exposure hotspots at the census tract level. Our novel approach for identifying at-risk areas therefore employed the joint distribution of these two characteristics in direct contrast to traditional approaches that assess risk by overlapping separate marginal distributions. This enabled us to uncover up to 73% more at-risk census tracts where an estimated 3.28 million additional at-risk children reside. Additionally, our approach revealed statistically significant systematic inequities in how traditional approaches detect risk, where newly identified at-risk children were disproportionately located in census tracts with higher concentrations of uninsured, White, and Native populations. It also highlighted that relying on housing age as a proxy for residential lead-based paint exposure may neglect important disparities in housing quality. Third, we coupled our synthetic U.S. population with the 2020 Residential Energy Consumption Survey (RECS) to estimate individual probabilities of residential air conditioning (AC) access—an essential cooling solution to protect against the adverse effects of heat stress—using a machine learning model to downscale data that was previously only available with spatial information at the state level to the census tract level. Unlike traditional approaches which focus solely on AC ownership as a proxy for AC access, our measure additionally accounted for the ability to install, maintain, and operate AC, resulting in probability estimates that were 17% lower and 9% more variable. By accounting for these additional dimensions of residential AC access, we also uncovered interstate and intrastate disparities shaped by fine-scale differences in state policies, infrastructures, and practices as well as climate conditions, housing conditions, and demographic characteristics. In summary, our novel synthetic U.S. population addresses critical gaps in existing population data, uncovering limitations, assumptions, and biases that may otherwise persist as researchers aim to devise and disperse public health interventions. Beyond its utility in mapping residential lead exposure risk and residential AC access, this model can be adapted to examine a wide range of public health challenges, enabling researchers to explore how intersecting social, economic, infrastructural, and environmental forces shape outcomes at fine spatial resolutions and providing them with a powerful tool to guide more effective and equitable public health interventions across the U.S.