Racial and ethnic price differentials in a small urban housing market

Abstract

This article examines whether the presence of blacks and Hispanics has a negative impact on prices in a small urban housing market in the US. To alleviate estimation biases associated with unobserved neighborhood heterogeneity, we focus on housing price differences across micro-neighborhoods in the small and relatively homogeneous city of Kingston, New York, introduce GIS-based spatial amenity variables as controls, and account for clustered errors, neighborhood fixed effects, spatial errors and spatial lags. Our results, with the exception of the spatial error model, conform with the consensus reached primarily from studies of large cities that the presence of blacks in a neighborhood is associated with lower housing prices and that the impact of the presence of Hispanics is considerably weaker. The spatial error model yields weaker and statistically insignificant results for blacks, providing some evidence that price discounts in relatively black neighborhoods may be caused not by preferences for segregation but by the correlation of race and the quality of neighborhood amenities.

Keywords:

• housing
• prices
• race
• ethnicity
• neighborhood

Acknowledgement

We acknowledge research support from the Bard Summer Research Institute (BSRI) of Bard College.

Notes

¹In a recent survey of the literature, Zabel (2008) defines racial prejudice as a preference for neighbors of the same race.

²Neighborhood fixed effects have been used in several hedonic models that focus on other issues such as school quality. See for example the cross-sectional estimates in Brasington and Haurin (2006) and both cross-sectional and panel estimates in Clapp, Nanda and Ross (2008).

³The data does not allow us to answer the first question; to establish discrimination, we need to identify the race and ethnicity of the homeowner. Home sales data does not identify such information of buyers and sellers.

⁴The population estimate is from US Census (2010). See Table 1 for earlier data.

⁵Census tracts are “small, relatively permanent statistical subdivisions of a county.” Census tracts “usually have between 2,500 and 8,000 persons and, when first delineated, are designed to be homogeneous with respect to population characteristics, economic status, and living conditions.” (US Census Bureau, Geography Division, April 19, 2000.) A census block group (BG) is “a cluster of census blocks having the same first digit of their four-digit identifying numbers within a census tract. BGs generally contain between 600 and 3000 people, with an optimum size of 1500 people” (U.S. Census Bureau, Geography Division, Cartographic Products Management Branch, July 18, 2001).

⁶ where x and y are number of individuals of two racial-ethnic groups in the neighborhood, X and Y are the corresponding numbers at the city level, and t is the total population of the neighborhood. See Population Studies Center, University of Michigan (2009) for details.

⁷We experimented with several other spatial amenity variables, such as the distance to the two business districts, but the high multicollinearity among the spatial amenities prevented us from obtaining isolating the effect of the distance to one amenity holding the distance to the other amenity constant. As a result, we chose to limit the spatial amenity variables to two, and chose two variables that are relatively uncorrelated with each other.

⁸The Schwartz Information Criterion (BIC) suggests that the third model (without neighborhood controls) is most appropriate; however, we choose the fourth model because the other two measures favor it, the BIC is not substantially different, and because the inclusion of neighborhood attributes reduces bias in the race-ethnicity coefficients and, perhaps most importantly, is theoretically more appropriate.

⁹The heteroskedasticity test performed are the normal and non-normal versions of the Breusch-Pagan (1979) and Cook–Weisberg (1983) tests (using hettest, hettest, iid and hettest, fstat in STATA) and the White (1980) test (using imtest, white in STATA). All test fail to reject homoskedasticity with p = 0.00. For normality, we examined the normal-probability plots, kernel density function of the residual and performed a Skewness and Kurtosis Test (using sktest in STATA) of D'Agostino, Balanger, and D'Agostino (1990) that was rejected at p = 0.00.

¹⁰The Variance Inflation Factors (VIF) for the continuous (non-dummy) variables are at most 8.12. For B and H, the two key variables of interest, they are between 3 and 5.

¹¹The Stata command used is the vce(robust) option in regress.

¹²Without controlling for square footage, an additional bedroom increases housing price by 4.5 percent and a bathroom increases housing price by 10.8 percent (results not reported in table).

¹³We do not report the dummy variable coefficients in Table 6. Note, however, that the coefficients reported in the fourth column of Table 5 are identical; only the standard errors are different.

¹⁴The STATA command used is the vce(cluster) option of the regress command.

¹⁵We thank an anonymous referee to for pointing out this possibility. Even though overall housing prices in Kingston did not peak until after 2007, the mean price in the three block groups with the highest proportion of blacks did reach a peak in 2006. However, when we re-estimated the model with home sales from 2000 to 2005, we obtain results that are statistically not different to those obtained with the full sample.