Empirical Tests: Does the Model Match Real Segregation?

A model that produces realistic-looking segregation in simulation is not automatically an explanation of real segregation. The Schelling model must be tested against data: do the mechanisms it proposes actually operate in real housing markets? Do the dynamics it predicts match observed patterns of neighborhood change? Where it fails, what does it miss?

The Calibration Problem

Testing the Schelling model against real data faces a fundamental obstacle: the model’s key parameter --- the tolerance threshold f --- is not directly observable. We cannot open a household’s head and read its threshold. The model predicts behavior conditional on preferences, but the preferences must be inferred.

Two approaches are used.

Stated preference surveys ask respondents to evaluate hypothetical neighborhoods of varying composition. Farley, Schuman, Bianchi, Colasanto, and Hatchett (1978) and subsequent Detroit Area Studies found that most white respondents would accept neighborhoods with up to 30 percent Black residents, but the fraction willing to enter declined sharply beyond that point. Black respondents consistently preferred mixed neighborhoods (50-50 or similar). These surveys have been repeated over decades, showing gradually increasing tolerance among white respondents --- higher fractions willing to enter more diverse neighborhoods --- but persistent asymmetry between groups.

Revealed preference methods infer thresholds from actual housing choices. If a household moves out of a neighborhood that is diversifying, and the move is not explained by income, housing quality, or job location, the residual may indicate a composition preference. Bruch and Mare (2006) used residential mobility data from Los Angeles County and found that the probability of moving increases as the fraction of different-race neighbors increases, consistent with the Schelling mechanism --- but the effect is modest compared to income and housing cost factors.

The calibration problem does not invalidate the model. It means that testing requires careful identification strategies that isolate the preference effect from confounding factors.

Card, Mas, and Rothstein (2008): The Tipping Point Test

The most influential empirical investigation of Schelling tipping dynamics is Card, Mas, and Rothstein’s 2008 paper in the Quarterly Journal of Economics. The authors tested whether U.S. Census tracts exhibit tipping behavior --- a non-linear relationship between initial minority fraction and subsequent change in composition.

Their method: identify Census tracts that were near a candidate tipping point in one Census year and track the change in white population over the following decade. If Schelling dynamics operate, tracts just above the tipping point should experience white population decline (cascade toward segregation), while tracts just below should remain stable (mixed composition persists).

The results supported the tipping hypothesis. The authors identified tipping points in the range of 5 to 20 percent minority share, varying across metropolitan areas. Tracts that crossed the tipping point experienced significantly faster white population decline than tracts that remained below it, controlling for observable neighborhood characteristics.

Two caveats. First, the estimated tipping points are lower than what stated preference surveys would predict --- surveys suggest tolerance for 20 to 30 percent minority composition, but the observed tipping points occur at 5 to 20 percent. This discrepancy may reflect the difference between stated and actual preferences, the role of expectations about future neighborhood change, or the influence of factors the model does not include (school quality, property values, social networks). Second, the identification strategy cannot fully separate preference-driven departure from economically driven departure, because minority in-migration may correlate with neighborhood economic changes.

What the Model Gets Right

The Schelling model’s predictions match several well-documented features of real residential dynamics.

Persistence without active exclusion. U.S. residential segregation declined only modestly between 1970 and 2020, even as explicit legal discrimination was dismantled and stated racial attitudes shifted substantially (Glaeser and Vigdor, 2012). The model predicts exactly this: once segregation is established as an equilibrium, it is self-sustaining even if the preferences that produced it weaken, because the segregated state satisfies any threshold above zero. This matches the empirical observation that de jure desegregation did not produce de facto integration.

Dynamics of neighborhood transition. When neighborhoods do transition from one demographic majority to another, the process is typically rapid --- taking one to two decades --- and unidirectional. This matches the cascade dynamics of the model: once tipping begins, the feedback is self-reinforcing and produces a fast, complete transition rather than gradual convergence to a new mix.

Comparative statics across cities. Cities with more moderate racial attitudes (as measured by survey data) tend to have somewhat lower segregation, consistent with the model’s prediction that lower f produces less segregation. Similarly, cities with higher vacancy rates tend to have higher segregation, consistent with the model’s prediction that more mobility options accelerate sorting. These cross-city patterns are rough but directionally correct (Cutler, Glaeser, and Vigdor, 1999).

Speed of change once tipping begins. Bruch and Mare (2006) found that in Los Angeles County, the speed of demographic change in transitioning neighborhoods matched the model’s predicted dynamics: rapid departure of the departing group concentrated in the first years after the tipping point is crossed, followed by deceleration as the neighborhood approaches homogeneity and remaining residents are satisfied.

What the Model Gets Wrong

The model omits several forces that are central to real-world segregation, and relying on it exclusively misrepresents the phenomenon.

Income sorting. Residential location is heavily determined by housing cost, and housing cost varies spatially. Because racial income gaps are persistent and large --- the median Black household wealth in the U.S. is approximately one-seventh of median white household wealth (Darity, Hamilton, Paul, Aja, Price, Moore, and Chiopris, 2018) --- income sorting independently produces racial segregation even in the absence of any racial preference. The Schelling model, by treating agents as identical except for type, cannot distinguish preference-driven segregation from income-driven segregation.

Institutional discrimination. Historical and ongoing discrimination in mortgage lending, real estate steering, zoning, and covenant enforcement has produced and sustained segregation through mechanisms that operate independently of individual preferences. Redlining maps drawn by the Home Owners’ Loan Corporation in the 1930s predict neighborhood racial composition decades later (Aaronson, Hartley, and Mazumder, 2021). The Schelling model contains no institutions.

School district boundaries and public services. Household location choice is heavily influenced by school quality, which in the U.S. is tied to geographic boundaries. School district lines create discontinuities in neighborhood composition that the Schelling model’s smooth grid cannot represent. Bayer, Ferreira, and McMillan (2007) showed that household sorting by race is substantially driven by sorting by school district.

Asymmetric preferences. The standard Schelling model assigns the same threshold to both types. Empirical evidence suggests strong asymmetry: in stated preference surveys, Black respondents consistently prefer mixed neighborhoods (50-50), while white respondents prefer white-majority neighborhoods. This asymmetry means the model’s symmetric version mischaracterizes the operative dynamics. Clark and Fossett (2008) showed that incorporating asymmetric preferences into the model changes the predicted equilibria substantially.

The naturalization risk. The most dangerous misapplication of the model is using it to argue that segregation is “just preferences” and therefore either natural or beyond policy reach. This framing is empirically incomplete: institutional forces, income inequality, and discriminatory practices contribute to segregation through mechanisms entirely absent from the model. The model shows that mild preferences are sufficient for segregation; it does not show that they are the operative cause in any specific real city.

Krysan and Crowder’s Cycle of Segregation (2017) provides the most comprehensive account of how preferences, information, social networks, and institutional barriers interact in real housing search --- a process far more complex than the Schelling model’s random relocation to an empty cell.

What Remains Valuable

The model’s contribution is not descriptive but diagnostic. It severs the inference from observed pattern to inferred intention. When a city’s neighborhoods are segregated, the natural interpretation is that individuals or institutions chose segregation. Schelling showed that this inference is invalid: mild preferences --- preferences that would be described as tolerant in any survey --- are sufficient. The policy implication is not that preferences are the only cause, but that even after all institutional barriers are removed, the dynamics may sustain existing segregation unless the interaction structure itself is changed.