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

Implicit biases, automatic differential attitudes towards individuals from various groups, significantly hinder equity. These biases affect numerous life aspects, including healthcare, education, employment, policing, and mental/physical health. While research shows implicit biases are socially driven, the systematic impact of urban self-organization remains unexplored. This study uses urban scaling theory, a framework that models cities as complex systems, to predict how city characteristics influence implicit biases. This approach links city-wide infrastructure to individual psychology, hypothesizing that larger, more diverse, and less segregated cities will exhibit lower levels of implicit bias. The ubiquity of implicit biases and their detrimental effects necessitate a theoretical understanding of their relationship with urban environments, potentially leading to interventions for less biased urban areas. Earlier research highlighted the early development and stability of implicit racial biases, their association with diverse school populations, and neurobiological evidence linking lower bias with more automatic out-group processing. However, more recent work demonstrated that interventions increasing out-group exposure reduce implicit biases, suggesting that biases reflect ongoing social environment predictions. The enduring effects of historical factors like slavery and segregation in the US highlight the lasting influence of structural factors on social contexts and implicit biases. This research aims to establish a quantitative link between city properties and implicit biases through a mathematical model.

Early research indicated that implicit racial biases develop early in life, persist into adulthood, and are less prevalent in diverse schools. Neurobiological studies confirmed this, showing that individuals with lower bias process out-group stimuli more automatically, suggesting early exposure to diverse populations is key for developing out-group expertise and reducing biases. However, later studies showed that interventions with older individuals also reduce biases, although effects diminish without continued intervention. This highlights that individual biases reflect ongoing predictions about the social environment, meaning stable biases reflect stable social contexts, not fixed individual cognition. The long-lasting effects of historical events such as slavery and segregation in the US exemplify this, with areas having larger slave populations in 1860 exhibiting higher current implicit racial biases. This demonstrates how structural influences on social contexts can perpetuate biases across generations.

This study utilizes a mathematical model grounded in urban scaling theory, which describes cities as social networks structured by their hierarchical infrastructure. This model accounts for individuals belonging to distinct groups and group-identity-biased connections. The model focuses on inter-group interactions, relating them to implicit bias levels. The model predicts that larger, more diverse, and less segregated cities have lower implicit racial biases. This prediction is tested using data from the racial Implicit Association Test (IAT) spanning 2010-2020, encompassing approximately 2.7 million individuals. The IAT measures bias as a difference in response times when pairing images of White versus Black faces with positive or negative words. Data is linked with U.S. Census demographics and population data for Combined Statistical Areas (CBSAs), which capture the extended social networks of cities. While the sample isn't nationally representative (younger, more educated, higher percentage female), racial demographics across cities are strongly correlated, making it suitable for relative comparisons. The model's equation (Equation 3) is broken down into three multiplicative terms: a scaling relationship, a diversity adjustment, and a segregation adjustment. Four distinct measures of residential racial segregation are used, and the model is tested across these measures. To control for individual demographics, IAT responses are transformed into an indicator for positive bias towards White faces, and logistic regressions are performed including city-level (population, diversity, segregation) and individual-level variables (race, education, birth sex). A noise ceiling analysis is employed to assess the model's performance given the inherent noise in the IAT measure. Furthermore, a Granger causality analysis is used on data from 43 cities with 10 years of data to assess the temporal precedence between changes in population size, diversity, segregation, and implicit biases.

The study found strong support for the model's predictions: 1. Larger cities show lower average levels of implicit racial bias. The scaling relationship between implicit bias and city size (log-log) revealed a negative relationship. 2. More diverse cities have lower levels of implicit bias, reflecting a diversity adjustment term. 3. Less segregated cities have lower levels of implicit bias, captured by the segregation adjustment term. These three factors (city size, diversity, segregation) explain a median of 33.6% of the variance in implicit racial bias across cities (2015-2020) with a noise-corrected R² range of [0.38, 0.93]. The diversity effect accounts for more variance than segregation. These structural factors predict individual IAT responses even when controlling for individual demographics, suggesting large-scale structural characteristics influence individual biases. The independent estimates of the learning rate (α), derived from the scaling exponent and the diversity adjustment, converge, supporting a shared mechanism (learning curve based on out-group exposure). These estimates are consistent with experimental interventions simulating inter-group contact. Granger causality analysis revealed that changes in city size, diversity, and segregation precede changes in implicit biases at short timescales (1-2 years). However, at longer timescales (3 years), the relationship becomes bidirectional, suggesting feedback loops where implicit biases also influence city structure.

The findings demonstrate that relatively simple considerations of heterogeneous mixing within social groups can explain a substantial portion of the variation in implicit racial biases across US cities. The fact that only three factors account for so much variance supports the notion that implicit racial biases are strongly shaped by social contexts rather than individual differences. The model generates testable predictions, including the short-term precedence of structural factors over bias changes and the presence of feedback loops at longer timescales. The findings align with the urban scaling literature demonstrating relationships between city size and other social and psychological phenomena. The results imply that fostering inter-group contact in cities is beneficial for reducing implicit racial bias. Future research should investigate the mechanisms underlying the long-term bidirectional relationship between biases and structural factors (e.g., selective migration).

This study successfully demonstrates a quantitative relationship between city-level characteristics (size, diversity, segregation) and implicit racial biases. The findings support a model suggesting that implicit biases emerge from the interplay between large-scale structural factors shaping social contexts and individual learning from these contexts. The model's predictions, supported by empirical data, provide a framework for understanding and potentially mitigating implicit biases in urban environments. Future work should explore the detailed mechanisms of the feedback loop at longer timescales and the role of other factors contributing to the remaining variance in implicit biases across cities.

The study's reliance on IAT data, while extensive, has limitations. The sample is not nationally representative and may underestimate overall bias levels. While the model accounts for a significant portion of variance, other factors not included could still influence implicit biases. The Granger causality analysis is correlative, not providing direct causal evidence of the temporal precedence between variables.