A tale of two pandemics: evolutionary psychology, urbanism, and the biology of disease spread deepen sociopolitical divides in the U.S.

The paper addresses why COVID-19 spread and responses differed across the U.S., examining statistical drivers in the general population rather than only victim risk factors. As of May 11, 2020, the U.S. accounted for about 30% of global COVID-19 deaths, amid widening social and political divides over non-pharmaceutical interventions (e.g., masks, distancing). The study’s purpose is to identify county-level drivers of per capita COVID-19 deaths and to explain divergent public behaviors using social learning and evolutionary psychology. The author posits two contrasting U.S. experiences—one centered on the New York metropolitan area with strong, obvious signals of danger, and another across the rest of the country with weaker, more variable signals—shaping behavior through evolved social learning mechanisms such as in-group information sharing, imitation, memory, and costly punishment.

The paper situates its analysis within social learning theory and the evolution of cooperation. It references work on imitation and norm enforcement (Flinn, 1997; McElreath et al., 2010; Hoel et al., 2019; Whiten, 2017), mechanisms promoting cooperation including costly punishment, contingent cooperation, external forcing, and memory/imitation (Alvard, 2003; Boyd and Mathew, 2007; Gurven, 2006; Huang et al., 2018; Liu et al., 2019), and the strengthening of social norms under threat (Roos et al., 2015) with historical parallels in epidemics (Dutta and Rao, 2015). One-shot learning is highlighted as an evolved cognitive capacity intensified by uncertainty (Biederman, 1987; Lee et al., 2015). Epidemiological and socioeconomic risk literature is referenced regarding comorbidities, pollution, and disparities (CDC COVID-19 Response Team, 2020; Millett et al., 2020; White and Nafilyan, 2020). This body of work frames expectations that local experiences and social networks inform adaptive yet divergent responses to COVID-19.

Unit of analysis: U.S. counties (n=3,143). COVID-19 deaths were obtained from USAFacts. Covariates were assembled from U.S. public data sources for each county: population density; urban status (Rural-Urban Continuum Codes); historical foreign travel; median income; per capita ICU beds; percent uninsured; median male age; comorbidity indicators (percent smokers, obesity, diabetes, extreme alcohol use); pollution (particulate matter); percent non-Hispanic Black; and percent of labor force in occupations affected by shutdowns. Due to multicollinearity (e.g., income correlated with urbanism and comorbidity; uninsured correlated with income and urbanism; percent Black correlated with pollution, comorbidity, and income), the percent non-Hispanic Black variable was dropped after factor analysis and stepwise regression indicated it primarily proxies other risks. The dataset was partitioned between the New York Combined Statistical Area (NYCSA; 31 counties) and the remaining 3,112 counties, based on initial modeling that showed ~80% of explained variance driven by the NYCSA. A linear model related per capita COVID-19 deaths to the independent variables with standardized coefficients, using stepwise regression to mitigate multicollinearity and retain only variables improving model fit. Adjusted R-squared was reported. Principal components analysis (PCA) with varimax rotation supplied orthogonal factor scores as an alternative approach. PCA yielded four interpretable components: PC1 (Comorbidity: high comorbidity, low income; weak rural association), PC2 (Density/Travel: high population density and foreign travel), PC3 (Young Urban: younger median male age, with weaker urban and pollution loadings), and PC4 (Insured: health insurance). Variance explained by PCs: 0.20, 0.18, 0.17, 0.12 (cumulative 0.67). Models were evaluated over multiple dates from March to October 8, 2020 to assess temporal dynamics. Coefficients and model R² were tested for statistical significance (typically p<0.0001).

- NYCSA models: Approximately 80% of variance in per capita deaths was explained. The most consistent and strongest positive association was with historical foreign travel. Unexpectedly, population density showed a negative association within NYCSA samples at certain times. Median male age showed negative associations, interpreted as counties with younger populations having higher per capita deaths, suggesting younger individuals as vectors. These relationships remained largely unchanged as of October 8, 2020. - Non-NYCSA models: Consistent predictors were urbanism and foreign travel; more urban counties with greater foreign travel histories experienced more deaths. However, explained variance was much lower than NYCSA, indicating higher variability and weaker signals. By October 8, 2020, additional significant associations emerged: negative with median income and median male age; positive with comorbidity and pollution. The variance explained remained low (e.g., R² around 0.166), underscoring continued heterogeneity outside NYCSA. - PCA factor models: For NYCSA, the Density/Travel factor was the dominant predictor of per capita deaths (e.g., coefficient ~0.79; Adj R² up to ~0.61 on 5/2/20). By mid-May, the Insured factor also became significant with a negative relationship to deaths (higher insurance associated with lower deaths). For non-NYCSA, Density/Travel and Urban factors were strongest; Comorbidity and Insured factors were initially opposite-signed before flipping by mid-summer to expected directions (as of Oct 8, 2020: Comorbidity positive ~0.264; Insured negative ~-0.144). - Temporal and geographic disparities: As of May 8, 2020, NYCSA average per capita death rate was 0.00113 versus 0.00007 for non-NYCSA (ratio ~15.3). By Oct 8, 2020, NYCSA averaged 0.00168 versus 0.00045 for non-NYCSA (ratio ~3.7). Extremely rural counties (largest category) exhibited death rates over 20 times lower than NYCSA on May 8, narrowing to ~3.8 by Oct 8. These figures highlight two distinct experiences of the pandemic with markedly stronger danger signals in NYCSA and weaker, variable signals elsewhere. - Political correlation: Per capita deaths were negatively correlated with county-level Trump 2016 vote share (May 2, 2020, r = -0.230), aligning with urban-rural and socioeconomic divides. - Behavioral inference: The consistent negative association with median male age suggests younger populations likely acted as key transmission vectors, consistent with observed risky behaviors among youth and lower individual risk of severe outcomes.

The findings confirm that logistical transmission factors—historical foreign travel, urbanism, and population density—primarily drove early COVID-19 mortality patterns, but their predictive strength varied sharply by region. NYCSA exhibited strong, coherent signals consistent with a concentrated outbreak seeded by travel and amplified by density; the rest of the country experienced weaker and more heterogeneous dynamics, with socioeconomic, comorbidity, and pollution factors becoming more salient over time. These differential experiences shaped behavior through evolved social learning mechanisms. In high-signal urban areas, direct experiences and one-shot learning reinforced protective behaviors (masking, distancing), further supported by in-group imitation and norm enforcement. In low-signal rural areas, limited direct exposure and trusted in-group information networks fostered skepticism or resistance to interventions, sometimes enforced via costly punishment within groups. These socially learned adaptations were individually rational given local risk and economic trade-offs but collectively deepened political divides, aligning with existing urban-rural cleavages and partisan identities. The study suggests that acknowledging these cognitive and social dynamics is crucial for designing effective, equitable public health messaging and interventions.

The study demonstrates two distinct U.S. pandemic experiences: a highly impacted NYCSA with strong, consistent signals of risk and a more variable, lower-signal experience elsewhere. County-level statistical modeling shows that the logistics of transmission—foreign travel and urbanism—dominated early mortality patterns, with comorbidity, pollution, and insurance coverage gaining importance as the pandemic diffused. Interpreting these patterns through evolutionary psychology and social learning theory explains divergent behaviors and politicization: people adapt via in-group learning, imitation, and norm enforcement based on local signals and economic trade-offs. Practically, public health responses should leverage social learning by tailoring messages to specific audiences, using credible in-group messengers and channels, and applying frameworks like APEASE to ensure acceptability, practicability, effectiveness, affordability, minimal negative spillovers, and equity. Such targeted communication may help mitigate further spread and reduce sociopolitical polarization.

- Multicollinearity among key covariates required stepwise regression and PCA-based factor modeling; interpretation of factors involves judgment. The percent non-Hispanic Black variable was dropped as it acted as a proxy for other risks (income, comorbidity, pollution). - The negative association of population density within NYCSA was unexpected and restricted to that sample, warranting caution in generalization. - Outside NYCSA, models had low explained variance, indicating high heterogeneity and weaker, inconsistent signals across counties. - Temporal dynamics: relationships evolved over time (e.g., signs of comorbidity and insurance effects flipped), reflecting changing spread patterns and potentially policy or behavior changes. - County-level aggregation and averaging across counties (unweighted) limit causal inference; averaged figures cannot be interpreted as direct individual-level probabilities of death. - Observational, correlational design cannot establish causality.