5.1. First Stage Estimation Results
Using the dataset created above, this study explores the correlation between TRI toxic emissions and racial segregation, and how tastes for endogenous demographics blur the role of exogenous public goods on segregation. Primary results focus on estimations using 0.25-mile radius buffer around TRI facilities. The model regressed in this study is a two-stage equilibrium sorting model, in which the first stage estimation recovers the household-specific taste parameters as well as a vector of mean indirect utilities for each neighborhood. In specifying the model, a limited set of interactions is included between household characteristics and neighborhood attributes to conserve the estimation’s degrees of freedom and limit potential problems of collinearity. The first stage estimation results using the specified interactions of household characteristics with housing and neighborhood attributes are shown in
Table 2.
Coefficients of interactions from the first stage indicate variation in the households’ preferences for specific housing and neighborhood characteristics of their housing choice. The interest of the first-stage estimation is the signs rather than the magnitude of the coefficients. Signs for Coefficients of interactions are all as expected, and statistically significant. More specifically, the a priori expectations concerning interactions of household specific characteristics with levels of TRI releases, and with neighborhood demographic composition are as expected. A positive coefficient for interactions of blacks and other races with TRI emissions indicates that households of black and other races have an increased likelihood of choosing neighborhoods with increased TRI emissions than whites. The negative coefficient for the interaction of household income with emission level is interpreted as, if households’ observed income increases, there is a decreasing likelihood of those households to choose housing in neighborhoods exposed to pollution. These results are consistent with previous studies [
45,
46] that non-white households are more likely to sort into neighborhoods with higher toxic release levels, and relatively wealthy households are more able to afford access to environmentally superior neighborhoods, confirming that TRI emissions do enter households’ residential location decision making [
5]. As communities exist with various levels of TRI emissions, interactions from the first-stage estimation describe households’ heterogeneous preferences for neighborhood environmental quality. The interactions of household race with neighborhood demographic composition (race “White” is dropped in the regression to avoid collinearity) implies that households prefer living in a closer spatial proximity to households of the same race as themselves. The negative coefficient for the interaction between black households and “Percent Census Block Group other races” indicates that, compared with neighbors of other races, black households are more prone to live with white neighbors. Similarly, compared with black neighbors, households of other races are more likely to select white neighbors. Looking at interactions of household income with neighborhood demographic composition, the negative coefficients imply that, while controlling for all the other factors, an increase in household income will decrease the likelihood of choosing a neighborhood with a higher proportion of non-white population. The coefficient of interaction between householder’s educational attainment and neighborhood demographic composition shows that households with higher educational attainment are less likely to select neighborhoods with higher percentage of non-white population. The first-stage estimation confirms the heterogeneity in preferences for both neighborhood public goods and demographic composition.
Other interactions of household characteristics with housing and neighborhood characteristics include the interaction between household size and number of bedrooms, the interaction of whether have children in family with school-district ranking score, and interaction of householders’ educational attainment with the school-district ranking score. In accordance with previous studies [
15,
47], larger households prefer larger houses. The coefficient for the interaction of households have children with the school-district ranking score is positive, indicating households with children care more about school quality. A positive coefficient for interactions of householder’s educational attainment with school-district ranking score indicates that highly educated households are more inclined to choose neighborhoods with higher school quality.
5.2. Second Stage Estimation Results
Using the estimation results of the mean utility from the first stage estimation as dependent variables, the second stage estimation can be implemented. As shown in Equation (10), there may be a correlation between housing prices and unobserved housing/neighborhood characteristics in the second stage estimation. For example, two identical houses in neighborhood of identical quality may have different prices, depending on how they are situated compared with other houses in nearby communities. To solve this endogeneity problem, following Bayer et al. [
32], an instrumental variable is introduced in the second stage, which is created based on Equations (11) and (12). The auxiliary regression includes the same variables as those estimated in the second stage for the 1-, 2-, 3-, 4-, and 5-mile rings around each neighborhood centroid. With this instrument in place an IV regression of Equation (10) is run and the results are reported in
Table 3.
Parameters estimated in the second stage returned the mean preferences for housing and neighborhood characteristics.
Table 3 shows that the price coefficient is negative and statistically significant, which means that houses with higher price result in lower utility ceteris paribus. Houses with more bedrooms and equipped with fireplaces and central air conditioning are more preferred, while older houses provide lower utility. Of particular interest for this study are the coefficients for TRI emission variables and neighborhood demographic composition (Percent Census Block Group black and Percent Census Block Group other races) in the second stage. The expected negative and significant coefficients for TRI emissions are obtained, which is consistent with previous literature finding that TRI facilities lower nearby housing values [
44]. The negative coefficients for the variables of “Percent Census Block Group Black” and “Percent Census Block Group Other Races” indicating that neighborhoods with higher percentages of non-white population are less attractive than neighborhoods composed mostly of white residents. Sensitivity analysis is used to check the robustness of results from
Table 2 and
Table 3. An alternative definition of the toxic release exposure variable is tested using 0.5-mile radius buffer around each TRI reported facility instead of 0.25-mile. The estimation process is then replicated. As shown in
Table A1, though the magnitude of the coefficients related to TRI emissions from the sensitivity analysis is bigger, the qualitative nature of the results does not change with the new specification.
5.3. Simulation Results
The results of our sorting model reveal that preferences for environmental quality and own-race neighbors might drive residential segregation among households according to race. The disutility associated with environmental hazards is likely to cause households to move to neighborhoods with lower levels of environmental hazards, resulting in excess housing supply in neighborhoods with higher levels of environmental hazard exposure, with an overall decrease of housing prices. Thus, households (such as low-income black households) that put priority of house price over environmental safety would move to neighborhoods with increased environmental hazards, resulting in potential housing segregation and public good (e.g., clean air) inequality. In addition, sorting over endogenous demographics could also drive segregation. Persistent racial residential segregation is often considered as the result of whites preferring to live with whites while blacks wish to live near many other blacks [
48].
To explore the relationship between residential segregation and household preferences for environmental quality and neighborhood demographics, a counterfactual simulation is conducted by switching off heterogeneous preference for TRI emissions and preferences for self-segregation. In the simulation, this study assumes that households’ preferences over TRI emissions and neighborhood demographic composition do not change with household race and thus taste parameters of interactions between neighborhood TRI release and neighborhood demographic composition with household race are turned off. While the probabilities that each neighborhood is selected as a result of households change in response to the two counterfactual scenarios, the corresponding predicted demographic composition is calculated to replace the initial composition applying these probabilities. Since the sorting model itself does not perfectly predict the location choices each household makes, it is important to point out that the neighborhood sociodemographic measures initially estimated by the model will not match the observed sociodemographic composition of each neighborhood. Therefore, before calculating the predicted demographic compositions for each simulation scenario, this study first solved for the initial estimation error associated with each neighborhood, and added this initial prediction error to the sociodemographic measures calculated in each counterfactual scenario. Using the new neighborhood demographic composition, the degree of neighborhood racial segregation, which is measured by own-race exposure rate following the definition of Bayer et al. (2004), is calculated. To construct own-race exposure rate, calculations by neighborhood are made by determining the fraction of households in the three different race categories that reside in the same neighborhood as the household of interest, and averages are then created for these neighborhoods over all households of a given race. The sorting model results show that households with different income differ in preferences for TRI emissions and neighborhood demographic composition. Therefore, the own-race over-exposure rate by income quantiles summarized to investigate the different effects of TRI emissions and demographic composition on racial segregation among different income classes.
The counterfactual simulation results summarized in
Table 4 describe effects of preferences for exogenous TRI emission and endogenous demographic composition on shaping the extent of neighborhood residential segregation by reporting three sets of exposure rate: (1) observed own-race exposure rate; (2) simulated own-race exposure rate by switching off only heterogeneous preference for TRI emissions with respect to race; and (3) the simulated own-race exposure rate by switching off both TRI emissions and same-race neighbor preferences. The main purpose of the simulation analysis is the comparison between the observed and both simulated segregation patterns. The overall neighborhood racial composition of Franklin County is 67.35% white, 23.92% black and 8.73% other races.
Table 4A reports the observed race exposure rates in the sample. Taking black households as an example, these measures imply that black households in Franklin County live in communities comprised of 41.00% white, 50.76% black and 8.23% other races on average. Comparing the measured exposure rates to the racial composition of the whole sample—67.35% white, 23.92% black and 8.73% other races—there is obvious evidence that black households live in communities with approximately two times the fraction of black households than would be found if they were uniformly distributed across the study area. The majority of the additional fraction of black households in communities in which black households live is offset by a decrease of white households. The remaining race exposure rates indicate that households of each race living with households of same race in proportion, are higher than the proportion for the entirety of Franklin County.
Table 4B shows the counterfactual exposure rates of eliminating heterogeneous tastes for TRI emissions. Differences between the simulated and observed own-race over-exposure rate in the last column of
Table 4 show that own-race over-exposure rates changed little, indicating that differences in neighborhood TRI emissions have a modest effect on neighborhood demographics. The slight increase in black residential segregation in this simulation scenario suggests that cleaning up all the dirty neighborhoods will not alleviate the residential segregation of black households. Therefore, the current residential segregation pattern is driven more by other factors (e.g., own-race neighbor preference) than the TRI emissions. To investigate further, another counterfactual simulation is conducted, during which both heterogeneous tastes for TRI emission with respect to race and own-race neighbor preferences are switched off. Results are reported in
Table 4C. When additionally turning off taste parameters of the interactions of neighborhood demographic composition with household race, the segregation (as measured by the over-exposure to households of the same race) of black household and white households decreases by 7.63% and 16.36%, respectively, while own-race over-exposure rate of other races changes slightly. Compared with
Table 4B, changes of exposure rates reported in
Table 4C are much larger than those reported in
Table 4B, indicating that segregation is driven more by demographic tastes than by tastes for exogenous environmental quality.
Intuitively, if neighborhoods segregation is driven by differences in the TRI release level, cleaning up polluted neighborhoods would reduce segregation, and our results show that sorting on TRI emissions drive little correlation between emissions and demographics. However, changes in TRI emissions could trigger sorting on demographics based on income level. Since higher income minorities do not need to join white neighborhoods to enjoy higher levels of public goods, they select to live in a closer spatial proximity to households of the same race with themselves. To test whether effects of heterogeneous tastes for TRI emissions and demographic composition on racial segregation vary according to household income level,
Table 5 reports the own-race over-exposure rate based on income quantiles.
Table 5A shows that white households are more segregated in higher income groups, while black households are more segregated in lower income groups. For households of other races, this study finds that 1st and 4th income quantiles are the most segregated groups. When turning off heterogeneous preference for TRI emissions, slight changes are found in own-race over-exposure rates for white and other other-race households of different income quantiles. However, segregation of black households in the 3rd and 4th income quantiles increases by 2.81% and 3.49%, respectively, which confirms that changes of TRI release levels could trigger segregation of black households with higher income.
Table 5C shows that the effects of heterogeneous tastes for TRI emissions with respect to race and own-race neighbor preferences on racial segregation are different according to income level. Compared with
Table 4B, the own-race over-exposure rates decrease most in the 4th income quantile for the white households, but decrease most in the 1st income quantile for black households. The segregation pattern of other-race households in all four income groups changes slightly. The analysis results reveal a more significant role of own-race neighbor preference in residential racial segregation dynamics, and this preference contributes to segregation differently according to household income.
Racial residential segregation has been a continuous issue in American society, which may attribute to the combination of sorting over public goods and preferences over neighborhood demographic composition. Segregation holds a longstanding position as one of the prime suspects in explaining the persistent inequality between blacks and whites, such as increases rates of black poverty and overall black–white income disparities [
49]. However, consistent with previous research findings that investments in low-public good communities have no effect on or can actually increase segregation [
4,
13], the simulation results show that public goods, such as neighborhood environmental quality, make little contribution to racial segregation, and current racial distribution patterns are driven more by neighborhood demographic preference. These findings imply that, to reduce racial segregation, by making investments in public goods, other factors such as own-race preference should also be considered when making segregation-reducing policies.