Alexander John Zapf

Compartmental models of infectious disease transmission make assumptions about human behaviors. Specifically, they parameterize interactions across population groups, assumed to have distinct epidemiologically relevant behavioral patterns, primarily through contact matrices stratified by demographic variables such as age, gender, or socioeconomic status. Although such demographic characteristics are readily measurable, they may inadequately capture the social and psychological forces that govern protective behaviors. Drawing on 20 waves of a national survey conducted throughout the COVID-19 pandemic in the United States, we show that institutional trust—particularly trust in public health agencies, physicians, and hospitals—is a dominant predictor of protective behavior adoption. For mask wearing during periods of strongest pandemic activity, institutional trust explains more behavioral difference across population groups than age, income, education, and partisan affiliation combined. In unadjusted analyses, the difference in protective behavior adoption between individuals with the highest and lowest trust in the US Centers for Disease Control and Prevention (CDC) was four- to six-fold larger than the corresponding differences by age, income, or education, and exceeded the difference between Democratic and Republican respondents. This association was institutionally specific (the relationship attenuates for trust in banks), and behaviorally specific (trust in the CDC is associated with protective behaviors but not visiting a doctor). The latter suggests that trust modifies voluntary compliance with public health recommendations rather than healthcare access or use. These findings identify institutional trust as a candidate stratifying variable for models of behavior-coupled transmission dynamics, and as a measurable target for public health surveillance.

Early during emerging infectious disease outbreaks, case-based surveillance is constrained by limited test availability, diagnostic delays, and low clinical suspicion. Novel pathogens that mimic symptoms of established conditions may generate detectable outbreak signals in routine testing data, as infected individuals seek testing for known conditions and test negative. We developed analytic and simulation frameworks using Poisson and negative binomial models to evaluate whether total and negative testing volumes for a clinically similar (“mimicking”) condition can provide timely outbreak warning. We systematically assessed detection performance across variations in baseline test counts, epidemic growth rates, testing fractions among cases, detection threshold stringency, and overdispersion. Thresholds based on negative tests consistently outperformed those based on total tests, achieving earlier detection with about one-third fewer cumulative cases while maintaining comparable false-positive rates. However, reliable detection required stringent conditions: high testing fractions among epidemic cases, low baseline test volumes, and low overdispersion. When testing fractions fell below 5%, epidemic size estimates offered little practical value; when above 30%, precision improved substantially. These findings support prioritizing access to disaggregated test result data but caution that this detection approach is best positioned as a resource-efficient complement within integrated surveillance portfolios rather than a standalone early warning system.

Compartmental models of infectious disease transmission make assumptions about human behaviors. Specifically, they parameterize interactions across population groups, assumed to have distinct epidemiologically-relevant behavioral patterns, primarily through contact matrices stratified by demographic variables such as age, gender, or socioeconomic status. Although such demographic characteristics are readily measurable, they may inadequately capture the social and psychological forces that govern protective behaviors. Drawing on 20 waves of a national survey conducted throughout the COVID-19 pandemic in the United States, we show that institutional trust — particularly trust in public health agencies, physicians, and hospitals — is a dominant predictor of protective behavior adoption. For mask wearing during periods of strongest pandemic activity, for example, institutional trust explains more behavioral variance across population groups than age, income, education, and partisan affiliation combined. In unadjusted analyses, the difference in protective behavior adoption between individuals with the highest and lowest trust in the CDC was four- to six-fold larger than the corresponding differences by age, income, or educational attainment, and exceeded the difference between Democratic and Republican respondents. This association was institutionally specific (e.g., the relationship attenuates for trust in banks), and behaviorally specific (e.g., trust in the CDC is associated with protective behaviors but not visiting a doctor). The latter suggests that trust modifies voluntary compliance with public health recommendations rather than access to or use of healthcare. We conclude that compartmental models of disease transmission would be substantially improved by incorporating institutional trust as a stratifying variable. We additionally offer a trust-integrated mathematical modeling framework and recommendations for the data infrastructure needed for its implementation.

Public health monitoring traditionally relies on active reporting from diverse data sources, including clinical and administrative data, disease registries, and population-based surveys. Yet these surveillance methods often face challenges such as incomplete reporting, time lags, and variable population coverage. Meanwhile, diagnostic laboratories routinely generate vast volumes of operational data that are currently untapped for public health monitoring. As these data are not collected for scientific inquiry or population-level surveillance, they often lack formal validation and may contain sensitive information. We developed a Bayesian hierarchical model to decompose aggregated laboratory assay volume data for 1.1 billion clinician-ordered assays across the U.S. from October 2019 to March 2023 into interpretable epidemiological and health system signals. The signals generated by these models were compared with known perturbances to health systems, such as the COVID-19 pandemic. The method does not rely on assay outcomes or individual-level data, providing quantitative signals of epidemiological trends and health system responses while protecting both the privacy of patients and commercially sensitive information. Temporal analysis reveals qualitatively different responses of assay volumes to major public health events, identifying assays whose use paralleled surges in hospitalization rates during the COVID-19 pandemic documented through traditional public health reporting structures. This framework suggests that routine operational data can be used to augment traditional surveillance by identifying anomalous patterns for expert epidemiological investigation. To be truly effective, data from multiple vendors must be integrated to create a comprehensive real-time national or supranational public health surveillance platform.

Early during emerging infectious disease outbreaks, case-based surveillance is constrained by limited test availability, diagnostic delays, and low clinical suspicion. Novel pathogens that mimic symptoms of established conditions may generate detectable outbreak signals in routine testing data, as infected individuals seek testing for known conditions and test negative. We developed analytic and simulation frameworks using Poisson and negative binomial models to evaluate whether total and negative testing volumes for a clinically similar (“mimicking”) condition can provide timely outbreak warning. We systematically assessed detection performance across variations in baseline test counts, epidemic growth rates, testing fractions among cases, detection threshold stringency, and overdispersion. Detection thresholds based on negative tests consistently outperformed total test thresholds, achieving earlier detection with approximately one-third fewer cumulative cases while maintaining comparable false positive rates. However, reliable detection required stringent conditions: high testing fractions among epidemic cases, low baseline test volumes, and low overdispersion. When testing fractions fell below 5%, epidemic size estimates provided little practical information; above 30%, precision markedly improved. These findings support prioritizing access to disaggregated test result data but caution that this detection approach is best positioned as a resource-efficient complement within integrated surveillance portfolios rather than a standalone early warning system.