Implicit racial biases are lower in more populous more diverse and less segregated US cities

The study investigates whether and how the organization of people in cities systematically influences implicit racial biases. While urban environments exhibit regular structural patterns that shape social interactions, implicit biases—automatic differential attitudes toward racial out-groups—are known to drive inequities across health, education, employment, and policing. Prior work suggests implicit biases reflect social contexts and may be stable due to stable environments rather than immutable individual traits. The authors propose that urban population size, demographic diversity, and residential racial segregation structure opportunities for inter-group contact, which in turn affects implicit racial bias through learning. The research aims to develop a theoretical model grounded in urban scaling theory linking city-scale properties to inter-group interaction rates and, via learning curves, to implicit bias levels, and to test these predictions with large-scale U.S. data.

Early research showed implicit racial biases develop early in life, can appear stable into adulthood, and are reduced in more diverse school settings. Neurobiological studies link lower implicit bias to more automatic processing of out-group stimuli, consistent with expertise effects. More recent interventions with older children and adults demonstrate that increased exposure to out-group individuals can reduce implicit biases, though effects decay without continued exposure, supporting a view of biases as ongoing predictions about social context. Historical structural factors such as slavery and long-standing segregation have enduring effects on social networks and are associated with present-day regional implicit biases. Urban scaling theory provides a framework for modeling social interactions enabled by urban infrastructure, and prior work has incorporated group heterogeneity, homophily, and segregation to explain socio-economic outcomes. A substantial literature indicates inter-group contact reduces prejudice and implicit bias (e.g., Allport’s contact hypothesis and subsequent empirical work), and there is evidence linking homophily to reduced between-group ties. Together, these strands motivate a model where city organization shapes inter-group contact opportunities that drive implicit bias through learning.

Theory: The authors extend urban scaling models to heterogeneous groups with potential biased mixing. They decompose average per-capita interactions into within-group and between-group components, with reduced between-group interaction rates parameterized by h. They posit that implicit bias b reflects a cognitive processing cost that decreases with inter-group exposure following a power-law learning curve: b ~ k_inter^(-α). For two-group cities (White and Black), they derive a simplified expression predicting that bias decreases with city size (scaling), increases with lower diversity, and increases with higher segregation. The model predicts: (1) b scales with population size (N) with a negative exponent; (2) deviations from the mean scaling relate negatively to diversity; (3) deviations relate positively to segregation. Data: Implicit racial bias was measured using Project Implicit’s racial IAT (Dhiep metric) for U.S. participants 2010–2020. Approximately 2.7 million responses with geographic identifiers were aggregated to Combined Statistical Areas (CBSAs). Cities were included per year if they had ≥500 IAT responses (robustness checks at ≥250 and ≥1000). U.S. Census five-year estimates provided CBSA populations (B01003) and tract-level racial demographics (B02001). Residential racial segregation between White and Black populations was computed at the tract level using four metrics: mean deviance, normalized segregation index, Gini coefficient, and exposure/correlation ratio (η^2), then mapped to CBSAs. Because h is not directly observed, segregation terms in the theory were proxied by these residential segregation measures under a linear dependence assumption. Statistical analyses: (1) Population scaling: OLS regression ln(b) = C + β1 ln(N) + ε to estimate the scaling exponent (β1) for each year. (2) Diversity and segregation adjustments: Using city-specific residuals ε as the dependent variable, OLS assessed contributions of diversity ln(N1 N2 / N^2) and segregation (via the proxy measures) to deviations from scaling. Variance inflation factors were checked. Robustness across years and segregation measures was evaluated. (3) Individual-level models: Logistic regressions on individual IAT outcomes (indicator Dhiep > 0) included city-level ln(N), diversity, and segregation covariates, plus individual controls (race, birth sex, education). (4) Learning rate estimation: The learning rate α was independently inferred from the population scaling exponent and from the diversity adjustment term, and compared to upper bounds from 18 experimental interventions simulating additional inter-group contact. (5) Noise ceiling: Split-half correlations of city-level IAT scores provided lower and upper bounds on explainable variance to contextualize R^2 values. (6) Temporal precedence: For 43 cities with full 10-year data, Granger causality tests assessed whether changes in population, diversity, and segregation precede changes in bias (and vice versa) using lags of 1–3 years. Statistical significance was evaluated via a χ^2 test on summed squared residuals. Multiple robustness checks, alternative segregation measures, and different IAT response thresholds were conducted.

- Scaling with city size: Across all years and segregation measures, larger cities have lower implicit racial bias. The population coefficient β1 is significantly negative with 95% CI in [-0.045, -0.031], consistent with the model’s prediction of decreasing bias with N. - Diversity and segregation effects: More diversity is associated with lower bias, while higher residential racial segregation is associated with higher bias. 95% CIs for diversity and segregation coefficients are β2 ∈ [-0.226, -0.163] and β3 ∈ [0.026, 0.066], respectively. Effects remain when accounting for collinearity (max VIF 6.31) and are generally significant post-2015 when sample sizes are larger. - Variance explained: For 2015–2020, city size scaling, diversity, and segregation together account for a median 33.6% of between-city variance in bias (range 24.2%–40.5%), corresponding to r ≈ 0.58 (range 0.49–0.64). Diversity explains more variance than segregation (diversity R^2 ≈ 0.16; segregation R^2 range ≈ [0.008, 0.082]). Noise ceiling analyses indicate the model captures a majority of the explainable variance (noise-corrected R^2 ≈ 0.38–0.93). - Individual-level relevance: When controlling for individual demographics, city-level population, diversity, and segregation remain predictive of individual IAT outcomes (β1 ∈ [-0.0404,-0.0124]; β2 ∈ [-0.1155, 0.1937], significant after 2015; β3 ∈ [0.1951, 0.7081]). - Learning rate: Independent estimates of the learning rate α from the scaling exponent and from the diversity adjustment converge and are consistent with upper bounds inferred from 18 experimental interventions simulating additional inter-group contacts, supporting a shared mechanism of learning via exposure. - Temporal precedence: Granger analyses over 43 cities show that changes in structural factors tend to precede bias changes at short lags. At 1-year lag: population → bias 73.8% ± 7.0; diversity → bias 61.9% ± 7.3; segregation → bias 69.0% ± 7.1. Reverse directions are less frequent (bias → population 19.0% ± 5.9; bias → diversity 24.4% ± 6.7; bias → segregation 19.0% ± 6.3). At 3-year lag, evidence becomes bidirectional (e.g., population → bias 85.7% ± 5.5; bias → population 76.2% ± 6.2; diversity → bias 88.7% ± 4.9; bias → diversity 84.5% ± 5.6; segregation → bias 95.2% ± 3.3; bias → segregation 81.0% ± 5.9).

Findings support the hypothesis that urban organization shapes implicit racial biases by structuring opportunities for inter-group contact. Larger, more diverse, and less segregated cities exhibit lower implicit bias, consistent with a learning-curve mechanism where exposure to out-group members reduces processing costs reflected by the IAT. The results indicate that a small set of city-scale structural factors explains a substantial portion of between-city variability, aligning with theories that implicit bias is strongly shaped by social contexts rather than stable individual dispositions. Temporal analyses suggest that changes in population, diversity, and segregation typically precede changes in bias over 1–2 years, consistent with relatively rapid psychological updating to social conditions, while over longer timescales biases and structures influence each other, potentially via mechanisms such as selective migration or evolving mixing preferences. The convergence of learning-rate estimates from distinct components of the model and their agreement with experimental intervention bounds strengthens the interpretation that exposure-driven learning mediates the link between urban structure and bias. These insights have implications for urban policy and design aimed at fostering inter-group interactions to reduce biases.

The study introduces and empirically validates a theoretical framework linking urban scaling, diversity, and segregation to implicit racial biases through an exposure-based learning mechanism. Analyses of approximately 2.7 million IAT responses across U.S. CBSAs over a decade show that larger, more diverse, and less segregated cities have lower implicit bias, with these structural factors explaining a substantial share of between-city variance within the bounds set by measurement noise. Short-term temporal precedence from structural changes to bias supports the model’s causal direction, while longer-term bidirectional influences point to feedbacks between biases and urban structure. Future work should develop explicit models for mechanisms by which biases influence urban demographics and segregation (e.g., selective migration), incorporate ambient population mixing beyond residence-based measures, refine segregation proxies, and pursue quasi-experimental or experimental designs to more directly assess causality.

- Causality: The results do not provide direct causal confirmation of the proposed mechanism; Granger precedence is not proof of causation. - Measurement: IAT scores are noisy and high-entropy; although noise ceilings were estimated, residual variance remains. - Sample representativeness: Project Implicit participants are not nationally representative (younger, more educated, more female) and likely underestimate absolute bias levels; analyses focus on relative differences across cities. - Segregation proxies: Reduced between-group interaction rates (h) are unobserved and proxied by residential segregation measures, which may not fully capture mixing in ambient or non-residential contexts. - Data thresholds and early years: Smaller sample sizes before 2015 reduced statistical power, particularly for segregation effects. - Unit of analysis: CBSAs aggregate across large areas; within-city heterogeneity in neighborhood experiences may not be fully captured. - Omitted variables: Other structural or cultural factors could contribute to remaining variance; area deprivation measures added little beyond the three structural factors but do not exhaust potential influences.