Gentrification and Changes in the Spatial Structure of Labor Demand

Transcription

1 Gentrification and Changes in the Spatial Structure of Labor Demand Nathaniel Baum-Snow, Brown University Daniel Hartley, Federal Reserve Bank of Chicago September 30, 2015 [Preliminary and incomplete. Please do not cite without permission from the authors] The views expressed are those of the authors and do not necessarily represent the views of the Federal Reserve Bank of Chicago or of the Board of Governors of the Federal Reserve System or its staff. 1

2 1 Introduction In the decades following WWII, the central regions of most U.S. metropolitan areas were in decline. Between 1960 and 2000, the aggregate central city population share in the 100 largest metropolitan areas fell from 49 percent to 24 percent while employment share declined from 0.61 to 0.34 (Baum- Snow, 2014). Sometime after 1980, however, the populations of many large cities began to stabilize and central areas of large cities began to rebound. Since 2000, the downtown areas of most central cities have experienced remarkable demographic change. Central neighborhoods in most cities have experienced rapid income growth and new residential housing construction. Areas near central business districts in many cities have experienced faster than average population growth since 1980 (Lee & Lin, 2014). This paper systematically characterizes the roles of various aspects of changes in demographic supply and neighborhood demand conditions to explain the changing fortunes of central neighborhoods in cities. Using the structure of a standard model of neighborhood choice which we will estimate with highly disaggregated decennial census tabulations, we will assess the welfare consequences of gentrification for various demographic groups, with a particular focus on assessing such consequences for incumbents in gentrifying neighborhoods. Table 1 reports statistics describing neighborhood change within 5 km of central business districts of 118 large U.S. metropolitan areas since Column 1 shows that the fraction of the population in these central areas declined in each decade after 1970, though the rate of decline has slowed and central area population levels have declined only slightly in each decade since Despite these continued population declines, evidence in the remaining columns indicate that downward demand shifts for central neighborhoods reversed sometime between 1980 and Each column shows the fraction of central area population living within a census tract within the top tercile of the CBSA distribution of the variable indicated in the header, with terciles recalculated in each year. That all fractions are below one-third indicates that these central areas have remained more disadvantaged than average CBSA neighborhoods during our entire study period. However, results in Columns 3-4 indicate that central neighborhoods experienced rapid relative growth in their white and college educated populations after Results in Column 5 indicate that income declines from the 1970s reversed during the 1980s, with very strong relative income growth in central neighborhoods since The higher levels in Column 4 than Column 3 indicate that central neighborhoods have always been relatively attractive to more educated minorities. Column 6 shows results for an equally weighted index of socioeconomic status (SES), constructed using standardized versions of the variables used to construct Columns With low housing supply elasticity, greater housing and land demand for central neighborhoods by higher SES individuals 1 While greater fraction white does not directly indicate higher socioeconomic status, it has been found to indirect indicate this. Altonji, Duraszelski and Segal (2000) review evidence that blacks have lower wealth than whites conditional on income. Neal & Johnson (1996) find that blacks have lower levels of pre-labor market preparation conditional on education. 2

3 can manifest itself mostly as population turnover rather than overall population growth. Evidence in Table 1 represents a benchmark against which we compare counterfactual neighborhood compositions that remove changes in various components of residential demand for central neighborhoods since These results indicate that central neighborhood population and SES status would have declined even more rapidly had demand by the college educated, whites and higher income households not increased after Increases in the fraction of the population college educated explain most of the shifts in the racial composition of central area populations and a bit less than half of the change in education levels of these areas, with the rest explained by changing neighborhood choices by the college educated. Our joint examination of race and income yields strong evidence that income growth in these central neighborhoods has primarily been driven by increases in demand by higher income households to livere there. Future revisions of this paper will additionally investigate the roles of shifts in family compositions and the age structure of the population, along with shifts in demand amongst narrowly defined demographic groups within these broad classifications. The remarkable post-2000 demographic change in central neighborhoods comes in the context of convergence in racial composition and income across neighborhoods in most CBSAs since However, consistent with evidence in Chetty et al. (2014) using individual level data, there exists considerable variation across CBSAs in the prevalence of such convergence. In particular, we provide evidence that more rapidly growing metropolitan areas experienced more rapid neighborhood convergence in the 1980s and period. Such neighborhood convergence, and downtown gentrification in particular, may actually harm poor residents on average, as they are more likely to be renters and thus be faced with financing the capital gains of their landlords as land values increase. Busso, Gregory & Kline (2013) discuss such a possibility in the similar case of neighborhood improvements that come from government investment. Below we develop a theory that will be used in future versions of this paper to evaluate such welfare gains. A better understanding of the drivers of neighborhood demographic change may also provide clues about some reasons for the growth in income inequality nationwide since Gould, Lavy & Paserman (2011) and Damm & Dustmann (forthcoming) provide independent evidence of the effects of neighborhoods on long-run outcomes. To the extent that neighborhood quality influences outcomes, it is important to better isolate the mechanisms that have driven changes in neighborhood inequality. In particular, it will be important to understand the extent to which gentrifying neighborhoods retain incumbent residents (who can benefit from positive spillovers) or price them out. A complete characterization of changes in demand for neighborhoods requires accounting for housing supply conditions in addition to neighborhood demand conditions. As residential demand for a neighborhood increases, housing supply conditions regulate the price increases which determine changes in population quantities. Therefore, isolating the magnitudes of neighborhood demand 3

4 shifts requires a model that explicitly incorporates price adjustments. Future versions of the paper will adapt standard models of neighborhood choice developed in Bayer, Ferreira & McMillan (2007), Galiani, Murphy & Pantano (2014) and Bayer et al. (2015) to aggregate data to recover structural estimates of demand parameters for each neighborhood by various demographic groups. Critical to this exercise is the consideration of general equilibrium effects in which demand shifts partially determine home price changes, which end up redistributing some population to closely substitutible neighborhoods. For our purposes, this is an important extension to the more partial equilibrium treatments of neighborhood choice in the existing literature. As is laid out in Couture & Handbury (2015), outward demand shifts for central neighborhoods may be driven by some combination of increases in their consumer amenities and nearby labor demand. A common narrative describes how post-wwii declines in central city amenities more valued by higher income residents including safety, public school quality (Baum-Snow & Lutz, 2011) and public capital may have ceased or reversed in many cities by In addition, changing racial attitudes and a slow reversal of the Great Migration, which initially promoted white flight (Boustan, 2010), may have be important. On the labor demand side, industries with more educated workers have remained relatively centralized and have experienced greater than average national growth rates. Services, transportation, communications and public utilities, public administration and finance, insurance and real estate are the 1-digit industries that are more centralized than average. Besides public administration, these were the fastest growing 1-digit industries nationally These same industries also enjoy the highest productivity benefits of density (Baum- Snow, 2014). These facts are in line with Kain s (1992) classic analysis indicating that the shifting composition of central city labor demand toward skilled workers had already begun in the 1970s, thereby exacerbating the potential mismatch between the locations of unskilled jobs and unskilled workers. Our results indicate that cities with more robust overall demand growth experienced greater neighborhood convergence and more prevalent recent downtown gentrification. While overall CBSA demand growth has driven some downtown revitalization, cities with downtown employment specialized in growing industries have experienced particularly strong residential demand growth. Concurrent with increases in the importance of skill intensive industries more likely to locate in urban areas, the composition of the working age population has changed markedly since Women s labor force participation has increased, families are having fewer children, and the age of women at childbirth is increasing. All of these demographic changes have pushed up households value of commuting time relative to their demand for living space and may have represented quantitatively important forces for residential centralization given the greater centralization of jobs than residences. Moreover, the rising age at first marriage means that a greater share of the population is less constrained about living closer to work. Future versions of this paper will investigate the potential importance of all of these channels for driving recent downtown revitalization. This paper proceeds as follows. Section 2 presents our gentrification measures, the data, and 4

5 some additional descriptive evidence on the changing fortunes of downtown areas and trends in neighborhood inequality. Section 3 evaluates potential explanations for changes in neighborhood inequality and examines the roles of central neighborhoods. Section 4 lays out a methodology for constructing counterfactual neighborhood compositions and explores the results of various counterfactual exercises. Section 5 discusses a full neighborhood choice model that will be estimated in future versions of the paper, allowing us to more carefuly quantify changes in demand for neighborhoods by different demographic groups and the consequences of gentrification for incumbent residents. Finally, Section 6 concludes. 2 Characterizing Neighborhood Change We aim to construct intuitive measures of neighborhood change that capture demand shifts for living in each neighborhood going back at least to 1980 and can be constructed using data available at the census tract level. Depending on the local elasticity of housing supply, outward demand shifts must manifest themselves as some combination of increases in housing prices, population and the socioeconomic status of residents. Evidence in Table 1 shows that most central neighborhoods, whose changes we are most interested in understanding, did not experience population growth but did experience marked improvements in the socioeconomic status of their residents. We thus view changes in the fraction white, fraction of adults with a college degree and mean household income as useful indicators of changes in residential demand for a neighborhood. 2 While far from an ideal measure, fraction white proves particularly useful as available data allows us to construct tract racial compositions under many different counterfactual scenarios, as is explored in detail below in Section 4. While measures constructed from tract home values and rents represent a valuable additional source of information about neighborhood change, they are not as straightforward to construct consistently or interpret. We thus primarily delay our discussion of such measures to Section 4, after having developed a conceptual framework that shows how to extract useful information from them. We construct a summary measure of neighborhood change using the three individual demographic measures discussed above. This summary measure for tract i is the average number of standard deviations tract i is away from its mean in each year for each of these components. We call this equally weighted tract z-score the socioeconomic status (SES) index. For tract i in CBSA j in year t and variables indexed by k the SES index is calculated as SES ijt = 1 yijt k yk jt, 3 2 Future versions of the paper will consider median individual income or earnings rather than mean household income. k σ k jt 5

6 where y k jt and σjt k are calculated with tract population weights. While we will also experiment with using the first principal component of these same three underlying variables, we prefer the equally weighted z-score approach as it doesn t mechanically assign more weight to a variable only because it has more variation and we think that all three measures indexed by k are roughly equally important indicators of neighborhood status. We also use our three demographic measures and the SES index to generate summary measures of changes in neighborhood inequality for each CBSA since The process for doing so resembles that in Chetty et al. s (2014) summary measures of intergenerational income mobility, but as applied to census tracts over time instead of parent-child pairs. In particular, we calculate correlations between CBSA demeaned outcomes between year t and t-10 or 1980, applying tract population weights in the base year. Correlations of 1 indicate no change in neighborhood inequality on average while correlations of less than 1 indicate neighborhood convergence. Chetty et al. (2014) and Lee & Lin (2014) use percentile ranks in each year rather than demeaned outcomes as a basis for describing intergenerational mobility and neighborhood popoulation change respectively. However, our analysis benefits from distinguishing neighborhoods that experienced small changes from those with large changes in their outcomes (relative to CBSA means), even if they had the same changes in rank. 3 Figure 1 depicts four measures of neighborhood change in the Chicago CBSA between 1980 and 2010, allowing for visualization of trends in neighborhood inequality. We calculate demeaned share white (Panel A), college graduate share (Panel B) log household income (Panel C) in each tract in 1980 and 2010, weighting by tract population. These demand indicators are graphed against each other in a scatterplot, with 45 degree and regression lines indicated. Both of these lines pass through (0,0) in each panel by construction. Green dots represent tracts within 5 km of the CBD whereas orange dots represent other Chicago CBSA tracts. Panel D shows results using our overall SES index of neighborhood quality. Regression slopes of less than 1, seen for log mean tract household income, tract share white and the composite SES index, indicate that Chicago neighborhoods have experienced convergence in these dimensions. The slopes of these regression lines are our neighborhood change measures for Chicago. Points above a regression line that are far to the left of a 1980 mean represent gentrifying census tracts. Figure 1 reveals much heterogeneity in neighborhood change in Chicago , with our three SES status measures clearly capturing distinct things. Chicago tracts experienced divergence on average in fraction college but convergence in the other measures. Results for share white in Panel A are closest to evidence that has been documented in the literature to date. The masses of points at the bottom left and top right of Panel A represent large concentrations of stable minority and white census tracts respectively. The relatively large number of tracts along the right edge 3 We experimented with additional summary measures of CBSA neighborhood change that are allowed to depend on initial tract conditions. 6

7 of the graph at almost 100 percent white in 1980 and ending up less than 70 percent white may have experienced tipping (Card, Mas & Rothstein, 2008). But a handful of tracts went in the other direction between 1980 and 2010, seen in the upper left area of the graph. These largely minority tracts in 1980 that gained white share much faster than the typical Chicago tract are almost exclusively within 5 km of the CBD. Indeed, all but 3 of the tracts within 5 km of the CBD that were less than 80 percent white in 1980 experienced increases in white share between 1980 and 2010, even though share white decreased on average. Such downtown area gentrification is clearly visible for the other measures as well in Figure 1, with green dots clustered in the upper left area of each panel. Figure 2 contains analogous graphs depicting changes in Chicago tract SES indexes over each decade of our study period. It shows that Chicago experienced a small amount of neighborhood convergence in each decade Green dots clustered on both sides of the regression line to the left of 0 in Panels A and B but only above the line in Panels C and D indicate the gentrification that began during the 1990s in Chicago. In Section 3, we document statistically that such patterns of neighborhood change near CBDs applies not just to Chicago, but is pervasive across medium and large metropolitan areas, and that poor tracts near CBDs turned around on average only after Data Construction We primarily use decennial census data and the American Community Survey (ACS) data tabulated to the tract level for this analysis. Central to our investigation is the need for joint distributions of population by race, education, income, age and family structure across census tracts. To recover as many of these joint distributions in the most disaggregated form possible, we make use of both summary tape file (STF) 3 and 4 census tabulations. We also use information about family structure and age by race from STF1 data from the 2010 census. (Because the 2010 census did not collect information about income or education, we must rely in the 5 year ACS data for these tract distributions.) The STF4 data is necessary to incorporate trends in income distribution and family structure by race into the analysis. All census tracts are normalized to year 2000 geographies using census bureau reported allocation factors. We include regions of all year 2008 definition metropolitan areas (CBSAs) that were tracted in 1970 and had a population of at least 250,000, leading to a sample of 118 CBSAs. 4 Much of the analysis weights each tract such that each CBSA is weighted equally. Future versions of the paper will supplement this information with census tabulated commuting patterns from 2000 and panel data from Equifax describing changes in the composition and creditworthiness of households in census tracts. The Data Appendix provides more detail about data construction % of the 2000 definition tract must have been tracted in 1970 to be in our sample. 7

8 Figure 3a shows a map of the 118 CBSAs in sample. Figure 3a shades the CBSAs by the share of central area residents living in a top tercile census tract in 1980, for which tracts are ordered by the SES index. Notice that the preponderance of metro areas are shaded dark green, indicating that their central areas are lagging far behind the rest of the CBSA. The two notable exceptions are New York and Santa Barbara which had central areas that were much more prosperous than their CBSA remainders. Figure 3b shows the same map but for Here we see many CBSAs that moved across the 0.33 threshold, meaning their central areas have a higher SES index than the average neighborhood for the CBSA. These include many large educated cities we think of as having experienced rapid recent gentrification, but with some additional cities as well. Figure 4 provides a more detailed picture of the geography of our central areas by only showing the area of the Midwest near Chicago. In Figure 4 the tracts in our CBSA sample that are greater than 5km from the CBD are shaded blue while those within 5km of the CBD are shaded red. 2.2 Patterns of Neighborhood Change Table 2 shows the fraction of the population within 5 km of the CBD in a typical CBSA living in tracts moving more than 20 percentile points or 1/2 a standard deviation of the CBSA tract distributions of our four baseline measures of neighborhood socioeconomic status. These numbers are calculated weighting by tract share of CBSA population in the base year. Commensurate with evidence in Table 1, each of our four measures indicate that central area tracts were on balance in decline during the 1970s. Results in Panel D show that central neighborhoods declines reversed sometime in the 1990s, when 3.1 percent of the central area population moved up at least 20 percentile points, relative to 2.5 percent in rapidly declining central tracts. Similarly, 5.3 percent of this population lived in tracts moving up at least 1/2 a standard deviation relative to 3.4 percent living in tracts moving down this much. This increase in the SES index of central tracts during the 1990s was mostly driven by income gains. As in Table 1, evidence in Table 2 shows that the resurgance of central areas really took off between 2000 and During this period, 7.0 percent of central population lived in tracts moving up 20 percentile points relative to only 1.9 percent living in tracts moving down in the typical CBSA. Figure 5 summarizes neighborhood changes in household incomes across all 118 CBSAs in our sample over each decade in our sample period. It reports the 10th, 50th and 90th percentiles of decadal changes in each 5 percentile bin of base year household income distributions. Slopes of regression lines indicate the average amount of neighborhood household income convergence across CBSAs over the indicated decade. Figure 6 presents analogous plots using the SES index. Evidence in Figures 3 and 4 indicates that neighborhoods in the typical CBSA experienced convergence in both household income and the SES index in each decade since 1970 except household income for the period. However, there is a lot of variation across neighborhoods, with the largest 8

9 variation experienced by neighborhoods in the tails of the base year distributions. That is, the most unstable neighborhoods are the ones that are either very distressed or very well off. Examination of individual dots to the left of each panel provides information about how much poor neighborhoods have rebounded in each decade. 3 Drivers of Neighborhood Change We now characterize variation in neighborhood convergence across CBSAs and asses the extent to which these differences are explained by housing and residential demand conditions. Using our CBSA indices of changes in neighborhood inequality, we investigate which types of cities experienced greater neighborhood convergence in each decade since Table 3 presents information about the distribution of our indices of neighborhood convergence across CBSAs. Each column for each period reports results from a CBSA level regression of the neighborhood change index using the measure at top on a constant, various demeaned base year CBSA characteristics listed in the notes and the change in CBSA log employment over the decade expressed in standard deviation units for some decades. Coefficients on control variables are not statistically significant in all but a few cases. Normalizations of control variables and ln(employment) to be mean 0 allows for interpretation of the coefficient on the constant to be the average index of neighborhood convergence across CBSAs in the indicated decade. instrument for ln(employment) with a Bartik (1991) type industry shift-share measure. This instrument is constructed by interacting the 1-digit industrial composition of employment in each CBSA in the relevant base year with national employment growth rates in each industry to generate a predicted change in employment for each CBSA. 5 The idea is to isolate demand shocks for living in CBSA that are driven by national trends in industry composition rather than factors that could be correlated with unobservables driving neighborhood change. Decades in which ln(employment) is not included in regressions exhibit insufficient first stage power. 6 Table 3 features a number of interesting results. We Neighborhood racial composition has been converging more and more rapidly over time. The 1970s experienced stable neighborhood racial composition, despite the fact that this decade experienced rapid white flight from cities to suburbs. The average white share neighborhood change index was 0.98 in the 1980s and 0.87 in the 1990s. This evidence is consistent with that in Cutler, Glaeser & Vigdor (1999) and Glaeser & Vigdor (2012) who document that racial segregation peaked in 1970 and has declined in every decade since. Negative coefficients on ln(employment) indicate that more rapidly growing CBSAs ex- 5 That is, we construct the Bartik instrument for CBSA j that applies to the period t 1 to t as: Bartik jt = k S jkt 1 ln(emp kt /emp kt 1 ), where S jkt 1 is the fraction of employment in CBSA j that is in industry k at time t 1 and emp kt is national employment in industry k at time t. 6 We also experimented with using an alternative Bartik style instrument constructed using national trends in wages by industry. This instrument yields similar results. The Data Appendix further explains instrument construction. 9

10 perienced more rapid convergence in neighborhood racial compositions. In contrast to racial compositions, most CBSAs experienced neighborhood divergence in fraction college graduate on average. However, this divergence did abate over the course of our sample period. Neighborhood convergence in fraction college was also somewhat stronger in more rapidly growing CBSAs. In contrast to the other two measures, neighborhood income convergence was strongest in the 1970s and has been declining since. In the 1970s, the index of household income convergence was 0.88 on average across CBSAs, rising to 1.01 in the period. But as with the other measures, more rapidly growing CBSAs experienced more rapid neighborhood convergence. Indeed, in each decade the effect of overall growth on neighborhood convergence is strongest when such convergence is measured in terms of household income. Taken together, the SES index results in the final column of Table 3 are roughly an average of the results for components in the other three columns. Evidence in Table 4 documents that demand growth in areas near CBDs only became established during the 1990s and is concentrated in the poorest neighborhoods. Moreover, it shows that downtown areas disproportionately benefitted when CBSAs experienced stronger economic growth overall, and particularly so if downtown industries were doing well. Each column in each panel of Table 4 presents results from a separate tract level regression of the change in the SES index on 5 km CBD distance ring indicators cbddis d ij, log distances to natural amenities ln(amendis m ) indexed by m and indicators for distances from top quartile SES index tracts in 1970 topdis d ij, with each tract weighted by its CBSA population share. We control for distance to natural amenities to account for the possibility that CBDs are more likely to be located near such anchors of high income neighborhoods (Lee & Lin, 2013). We control for distance to top quartile tracts in order to exclude the possibility that tracts near CBDs gentrified simply because of expansions of nearby high income neighborhoods (Guerrieri, Hartley & Hurst, 2013). Because we want to distinguish reasons for which poor tracts gentrify from reasons for which richer tracts change, we run these regressions separately for each 1970 defined CBSA tercile of the SES index. We weight by the tract s fraction of CBSA population. The estimation equation is as follows: S ij = ρ j + 4 α d cbddis d ij + αb 1cbddis 1 ij Bartik j + α1cbddis s 1 ij Spatbartik j d=1 + 4 β d topdis d ij + δ m ln(amendis m ij ) + ε ij d=1 m To understand heterogeneity in neighborhood change for those within the inner CBD distance ring of 0-5 km, we interact this indicator with the same Bartik variable used in Table 3 and an analogous version built only using the industrial composition of employment within 5km of the CBD as of year 2000 explained below, both standardized into separate z-scores. Because we do not observe the change in employment within 5 km of CBDs, we cannot use it as a regressor directly, for which we 10

11 would instrument with the spatial Bartik instrument. For this reason and to maintain consistency across the two Bartik demand shifters, we estimate the reduced form described by (??). 7 The goal of the spatial Bartik variable Spatbartik j is to isolate labor demand shocks that hit the CBD area harder than other areas of a CBSA. The overall fraction of employment near the CBD at F emp nationwide is given by F emp = k Semp k f emp k, where S emp k is the share of overall employment in industry k and f emp k is the fraction of industry k employment near the CBD. We think of as being driven by fundamental attributes of the production process like the importance of f emp k agglomeration spillovers to TFP. Therefore, we predict the change in the fraction of employment near the CBD to be Spatbartik jt = k S emp jkt 1 f emp k2000 ln(emp kt/emp kt 1 ). Results in Table 4 demonstrate the reversal of fortunes experienced by many low SES tracts after Panel A shows that on average bottom tercile tracts near the CBD were declining during the 1970s and 1980s, with a clear reversal of this decline during the 1990s and in the period which was unique to low SES tracts. Results in Panel B show that middle and high tercile tracts near CBDs were consistently on the decline or stable respectively after As we explore further throughout the remainder of the paper, this downtown gentrification could have been driven changes in consumer amenity conditions or labor demand conditions. Interactions with the two Bartik variables explore the potential importance of shifts in labor demand. Evidence in Table 4 indicates that bottom tercile neighborhoods near CBDs in CBSAs with positive employment growth overall did better in the 1990s but worse in the 1970s. That is, CBSA economic growth actually hurt central areas of cities during the 1970s. Overall CBSA growth had no statistically significant effect for other types of central neighborhoods in any time period studied. In contrast, central areas of CBSAs with an industry composition that was more likely to be near CBDs did better in lots of cases, and never worse. Amongst all types of census tracts, central tracts did better in the period in CBSAs with CBD oriented industry compositions, with middle tercile tracts in such CBSAs also improving more during the 1980s. This is evidence that the types of production activities taking place in central areas of CBSAs have been driving at least some of the residential resurgance in these areas. Table 5 presents regressions analogous to those in Table 4, except that an index of tract housing value growth rates is used as the dependent variable. In particular, the dependent variable in Table 5 is calculated as the residuals from a regression of log mean tract housing value on various characteristics of owner occupied housing and CBSA fixed effects. Because positive demand shifts for neighborhoods will be reflected as some combination of increases in quantities of residents, 7 We are forced to maintain year 2000 industrial composition for the Spatial Bartik instrument because of limits to data availability in other years. We use the census journey to work tabulations to calculate this industrial composition. 11

12 potential income of residents and housing prices, we view evidence of house value growth and/or increase in SES index for neighborhoods as signs of outward demand shifts. Indeed, CBSAs with high housing supply elasticities (Saiz, 2010) may have had some neighborhoods with large outward demand shifts but experienced only small relative changes in housing costs in such neighborhoods. However, because they have the smallest availability of developable land, central areas of cities are likely to have such supply elasticities that are amongst the lowest in any given CBSA. Evidence in Table 5 shows increases in home values in bottom tercile tracts near CBDs but declines in home values in middle and upper tercile tracts near CBDs. It is well known that housing prices are negatively serially correlated at decadal time scales (Baum-Snow & Marion, 2009), so the results for the bottom and top terciles are not surprising. Coefficients on the Bartik interaction are not significant in any instance. However, coefficients on the spatial Bartik interaction are positive and significant for the bottom tercile in , the middle tercile in and the top tercile for and That is, home prices did increase in some decades in central areas of CBSAs with CBD oriented employment mixes. Overall, evidence in Table 5 is broadly consistent with the evidence in Table 4, that poor central neighorhoods have seen a resurgance, and especially those in CBSAs with CBD oriented employment mixes. 4 Counterfactual Neighborhood Compositions To separate out the roles of CBSA level demographic change from changes in individual groups neighborhood demands, we carry out decompositions of the sources of neighborhood change along the lines proposed by DiNardo, Fortin & Lemieux (1996). In particular, we calculate indices of neighborhood change in two types of counterfactual environments. First, we hold the shares of CBSA population in various demographic groups fixed over time, but allow neighborhood choices of these demographic groups to shift as in equilibrium to evaluate the extent to which changes in the overall demographics of CBSAs have driven neighborhood change. That is, these counterfactuals adjust the composition of CBSA demographic shares to resemble those in 1980, but not allocations of people in any particular observed demographic group across tracts. Second, we allow the CBSA demographic shares to evolve as they have in the data since 1980, but allocate these quantities to each census tract using the fractions that each demographic group made up of the tract in This second exercise allows us to determine the extent to which changes in the demand for each census tract by each observed demographic groups, rather than overall demographic shifts, have driven neighborhood change for different locations. 12

13 4.1 Construction of Counterfactual Neighborhoods We observe the joint population distribution f jt (i, r, e) of race, r, and education, e, and joint household distribution f jt (i, r, y) of race and income, y, across census tracts i in CBSA j in year t. Given the structure of available data, we are forced to evaluate counterfactual distributions in only two demographic dimensions at a time. 8 To clarify notation, we present how we calculate counterfactual tract populations based on joint distributions of race (white, black and, other) indexed by r and education (youth, less than high school, high school, some college, college or more) indexed by e across tracts. Denote Njt re as the total number of people in group re in CBSA j at time t and N jt as the total population of CBSA j at time t. The density function of CBSA demographics is f jt (r, e) = if jt (i, r, e). Crucially, we treat CBSA-level allocations f jt (r, e) and populations N jt as exogenous. It is important to note that while aggregate population does not influence conclusions drawn from these mechanical counterfactuals, it will matter once we incorporate housing supply Adjustments for Changes in Demographic Shares For the first set of counterfactuals, we revert the share of the CBSA population in each r and e cell back to 1980, according to the following equation: f jt (i, r, e) = f jt (i r, e)f j8 (r, e). (1) This is akin to quantity re-weighting in the DiNardo, Fortin & Lemieux (1996) setup. fjt (i, r, e) measures the fraction of people of each race/education combination that would live in each tract if the overall CBSA fraction of each race/education combination did not change after One difficulty when considering the counterfactual, fjt (i, r, e), is that it does not separately reveal the extent to which changes were due to shifts in the racial or education composition of the population. To better understand which element matters more, we also carry out such shares adjustments for race only conditional on education. For example, the counterfactual distribution allowing education shares to evolve but holding CBSA race shares constant is f jt(i, r r, e) = f jt (i r, e)f j8 (r e)f jt (e). (2) Comparing statistics about neighborhood changes generated using f jt r (i, r, e) and f jt (i, r, e) sheds light on the extent to which changes in the shares of the population in each education group influenced neighborhood composition while holding overall racial shares constant. 8 Future versions of the paper will additionally make use of information about family and age structure of tract populations by race. 13

14 4.1.2 Adjustments for Changes in Neighborhood Choices The other type of counterfactual distributions we calculate tells us the extent to which changes in neighborhood choices by particular demographic groups have driven gentrification. The most general version of this counterfactual is to reallocate year t total population in each demographic group to 1980 locations (using the fraction of that each demographic group contributed to the population of the census tract in 1980) but maintaining the contemporaneous CBSA-level demographic shares. f jt (i, r, e) = f jt (r, e)f j8 (i r, e). (3) Simultaneous implementation of demographic shares and neighborhood choices adjustments described in (1) and (3) yields f j8 (i, r, e), the actual density function from Therefore, we can think of these counterfactual experiments as together fully describing changes in neighborhood compositions between 1980 and year t Implementation for Race-Income In order to calculate counterfactual distributions of race and income, data limitations force us to focus on households and families rather than individuals. We observe the number of families by race in various income bins in 1980 and households by race in income bins in later census years. To make income categorizations consistent over time, we calculate quintiles of the national family income distribution in 1980 and the national household distributions in later decades using the 5% census micro data public use sample. Using this information, we calculate the fraction of families or households by race in each quintile of the national distribution in each year. That is, f jt (i, r, y) is the fraction of households in CBSA j and year t who are of race r, in national income quintile y and living in tract i for t > For t = 1980, this is the fraction of families instead. We have assembled the necessary data, and we are currently switching to using households for the 1980 income measures instead of families. In addition, future versions of the paper will also analyze roles of changes in the joint distributions of race-family structure and race-age in understanding central area neighborhood change. 4.2 Counterfactual Results Table 6 shows how the share of CBSA population living within 5km of CBDs would have changed under the various counterfactual scenarios laid out in the prior sub-section. These numbers are calculated analogously to those in Table 1 Panel B Column 2, except using counterfactual data sets. Panel A uses data with tract level education-race joint distributions whereas Panel B uses data with tract household income-race joint distributions to construct counterfactual data sets. Column 1 in Panel A repeats data on the evolution of the fraction of CBSA populations and 14

15 households living within 5 km of CBDs from Table 1 Panel B Column 2. These numbers are a benchmark against which we compare counterfactuals using population data. Column 1 in Panel B shows an analogous benchmark calculated using data on families in 1980 and households in other years. Because the way in which income data is tabulated, we are forced to use households or families rather than individuals, and thus, this is the appropriate benchmark. The reason that the first entry in Panel B Column 1 is positive is that non-family households are more likely to be located near CBDs and these units are excluded from the family counts. Future versions of the paper will resolve this problem by using household income throughout. Overall, adjusting CBSA shares to resemble those in 1980 has a sufficiently small effect on the results so as not to be quantitatively important for understanding changes in shares of people and households living within 5 km of CBDs. Adjusting CBSA population shares for race conditional on education (Column 2 in Panel A) yields slightly more rapid declines in central areas as these areas are less white than average. That is, the increasing fraction of minority populations has resulted in a small boost for populations in central neighborhoods of CBSAs. However, maintaining 1980 education-race CBSA shares yields almost no difference between central area counterfactual and actual population growth. The rise in shares of CBSA populations that are college graduates has counteracted the rise in minority share to cancel out. This means that changing neighborhood choices must have driven the patterns seen in Panel A Column 1. Indeed, results in Column 4 confirm that holding neighborhood choices constant for 1980 and allowing shares to change over time generates slight counterfactual growth in central area populations. That is, the decline in the white population and increase in the college educated population approximately offset to keep central neighborhoods stable in this counterfactual world in which neighborhood choices don t change. When examining counterfactual race-income population distributions (Table 6 Panel B), similar results hold, though it is clear that the rise in the fractions of both high and low income households has helped support downtown populations. Absent the shifts in the income distribution, the fraction of families/households living in downtowns would have declined by 3.8 percentage points rather than the 2.4 percentage points they actually declined between 1980 and 2010 (last row, Panel B Column 3). Counterfactual allocations holding neighborhood choices constant in Column 4 reveal an even starker pattern. Here we see a counterfactual increase of relative to an actual decline of Evidence in Table 8 discussed below indicates that this reflects reductions in the propensity of nonwhites and the poor to live in downtown areas. Had these groups remained in downtown areas, their populations would have grown robustly - because they are a growing fraction of the overall population. Table 7 examines downtown neighborhood racial compositions in the same counterfactual environments as in Table 6. Benchmark statistics in the data are listed in Column 1 with analogous statistics under various counterfactual scenarios in the remaining columns. All statistics for counter- 15

16 factual populations are reported using tract tercile designations that have been recalculated using the counterfactual data. Results in Column 2 examine how racial compositions of central area tracts would look if shares are adjusted for race only conditional on education (Panel A) or income (Panel B). Results indicate that absent changes in the racial composition in the population, central areas would have experienced more rapid increases in their white populations, except during the 1980s. This reflects the fact that conditional on income, whites chose to live in these neighborhoods at higher rates over time. Comparison of results in Columns 3 and 4 in Panel A show that the full increase in fraction white can be accounted for by the fact that college graduates increased as a fraction of the population, and college graduates are more likely to be white. In particular, holding the white population at 1980 shares conditional on education generates a counterfactual 1.6 percentage point increase in the share of near-cbd residents living in the whitest third of neighborhoods within the CBSA , whereas holding both education and racial shares constant generates a counterfactual drop of 0.2 percentage points in the mweasure. This means that increases in the college educated population outweighed declines in the white population to account for its full increase. It is important to note that all of this phenomenon is driven by these demographic trends in the 1980s and 1990s. Results in Column 4 show that about 50% of the growth in the white population in central areas can be explained by whites choosing to live in these neighborhoods at higher rates than in Panel B examines the roles of race and household income. Here we see a much larger role for changes in the composition of the population across income groups than we saw for changes in racial or education composition. Evidence in Column 2 reveals a small role for shifts in the neighborhood choices of whites conditional on income group, with counterfactual declines of 1.5 percentage points relative to a benchmark of 1.9 percentage points. However, evidence in Column 3 is striking. It shows that reductions in the fraction of the population in income groups disproportionately located in central areas of CBSAs has worked strongly against growth of fraction white in these areas in each decade Moreover, results in Column 4 shows that virtually the entire change in fraction white in central neighborhoods can be explained by the greater propensity for higher income households to choose to live in these neighborhoods. Because the results showing changes in choices by race conditional on income have not been an important driver of neighborhood change, it appears that income rather than race is what really matters here. Higher income people of all races have been choosing to live in downtowns in greater numbers. Results in Table 8 examine how education and income compositions of central area tracts would look under the same set of counterfactual scenarios. Results in Panel A Column 2 indicate that about 60 percent of the increase in central area population living in the most educated areas of CBSAs is due to the fact that the fraction of college graduates in the population increased (counterfactual growth of versus 0.048). Results in Panel A Column 3 indicate that changes in neighborhood choices amongst those in race-education cells accounts for a slightly 16

17 greater fraction of the increase in the propensity for the college educated to live in downtown areas (counterfactual growth of only 0.016). Taken together, this evidence indicates roughly equally sized roles for demographic change and changes in neighborhood choices of college graduates and whites for generating more gentrified central neighborhoods. Results in Table 8 Panel B indicate that changes in the income distribution have worked against the income growth in downtown areas. Holding the race-income distribution as it was in 1980, the fraction of central area populations living in top income tercile tracts would have risen by 9.7 percentage points rather than 6.6 percentage points between 1980 and Instead, we see in Column 3 that changing neighborhood choices are driving the icnome growth. Holding these choices at their 1980 states, we see 0 counterfactual income change in central neighborhoods. 5 A Model for Interpretation Here we develop a neighborhood choice model in the spirit of Berry, Levinsohn and Pakes (1995) that will facilitate recovery of estimates of elements of changes in household demand for individual neighborhoods by household type in future drafts of the paper. The model demonstrates how to use information about the fraction of CBSA residents in various demographic categories in each tract in each year as an indication of demand by each type for these tracts. That is, the model shows how to make use of conditional choice probabilities (CCPs) as in Bayer, McMillan, Murphy & Timmins (2015). One common challenge faced in demand estimation is in how to handle the endogeneity of prices, in this case housing prices. We have the benefit of multiple years of data, which allows us to recover sufficient information about the supply side of the housing market to effectively endogenize house prices in the model. That is, the model will explicitly incorporate the existence of upward sloping housing supply (Saiz, 2010) in each neighborhood, which pushes back against population growth in neighborhoods with the most rapidly rising residential demand, redistributing these households elsewhere in the CBSA. The indirect utility of household s of type h residing in census tract i is modeled as follows: v shi = β 0h Z i + β 1h p i + β 2h d hi + δx sh + ξ i + ε shi = v hi + ε shi In this expression, Z i is observed neighobrhood characteristics, p i is the price of one unit of housing services in tract i, X sh is household characteristics, d shi is commuting distance, ξ j is unobserved neighborhood characteristics and ε shi is an i.i.d. random utility shock distributed extreme value Type I. We think of this being a long-run equilibrium where moving costs are negligible. This setup delivers the following population shares of household type h in each census tract i. π hi = exp(v hi) i exp(v hi ) 17

18 Inverting this expression, we solve for v hi up to a scale from conditional choice probabilities: v hi = ln π hi ln ( i exp(v hi ) This suggests the possibility of recovering estimates of β 0h Z i + β 2h d hi + ξ i given that we observe π hi and house prices with some structure on the data generating process for house prices. One way of restating the question of why gentrification has occurred is to ask whether we can understand changes in π ij because of changes in Z i, d hi and ξ i holding parameters constant, or whether we need to shift parameters over time to justify patterns in the data. For ease of notation, we have suppressed CBSA and time subscripts. To model housing supply, we only impose that the supply function is constant over time in each tract. That is, H i = A i p θi i. To close the model, we impose the equilibrium condition that supply equals demand in the housing market: H i = N h π hi. Because N h is exogenous, it is possible to h simultaneously recover information about supply elasticities θ i (assuming they are fixed over time) and the combination β 0h Z i + β 2h d hi + ξ i, which could change over time. As in Bayer, McMillan, Murphy & Timmins (2014), we will then be able to recover estimates of β 0h and β 2h in a second step of estimation. ) 5.1 Using the Model to Construct Counterfactuals With only information on conditional distributions like f(i r, e) and total CBSA populations by group Nj re, we have all the information we need to recover estimates of β 0hZ i + β 2h d hi + ξ i for each tract/household type combination. 5.2 The Welfare Consequences of Gentrification Estimates from the model will be used to evaluate the welfare consequences of gentrification for households in different demographic groups. The estimated model will allow us to assess the extent to which welfare changes would have been different had demand for central neighborhoods by higher income households, college educated individuals and whites not increased since Conclusions To be completed. 18

19 References Altonji, Joseph, Ulrich Doraszelski, and Lewis Segal (2000) Black/White Differences in Wealth, Economic Perspectives, Vol. 24(1), pp Bajari, Patrick, and Matthew E. Kahn (2005) Estimating Housing Demand with an Application to Explaining Racial Segregation in Cities. Journal of Business and Economic Statistics, 23(1): Bartik, Timothy (1991) Who Benefits from State and Local Economic Development Policies?, Upjohn Press. Baum-Snow, Nathaniel (2007) Did Highways Cause Suburbanization? Quarterly Journal of Economics, 122(2), Baum-Snow (2014) Urban Transport Expansions, Employment Decentralization, and the Spatial Scope of Agglomeration Economies, manuscript. Baum-Snow, Nathaniel and Byron Lutz (2011) School Desegregation, School Choice and Changes in Residential Location Patterns by Race, American Economic Review, 101(7), Baum-Snow & Marion (2009) The Effects of Low Incoma Housing Tax Credit Developments on Neighborhoods, Journal of Public Economics Bayer, Patrick, Fernando Ferreira & Robert McMillan (2007) A Unified Framework for Measuring Preferences for Schools and Neighborhoods Journal of Political Economy Bayer, Patrick, Robert McMillan, Alvin Murphy, & Christopher Timmins (2015) A Dynamic Model of Demand for Houses and Neighborhoods. manuscript. Berry, Steven, James Levinsohn and Ariel Pakes (1995) Automobile Prices in Market Equilibrium, Econometrica, 63(4): Boustan (2010) Was Postwar Suburbanization White Flight? Evidence from the Black Migration. Quarterly Journal of Economics. Busso, Matias, Jesse Gregory & Patrick Kline (2013) Assessing the Incidence and Efficiency of a Prominent Place-Based Policy American Economic Review 103(2): Card, David, Alex Mas & Jesse Rothstein (2008) Tipping and the Dynamics of Segregation, Quarterly Journal of Economics, 123 (1): Chetty, Raj, Nathaniel Hendren, Patrick Kline, and Emmanuel Saez (2014) Where is the Land of Opportunity: The Geography of Intergenerational Mobility in the United States, Quarterly Journal of Economics, forthcoming. Couture, Victor and Jessie Handbury (2015) Young Professionals and Urban Revival in America, 2000 to 2010, manuscript. Cutler, David, Edward Glaeser & Jacob Vigdor (1999) The Rise and Decline of the American Ghetto, Journal of Political Economy Damm, Anna Piil and Christian Dustmann (forthcoming) Does Growing Up in a High Crime Neighborhood Affect Youth Criminal Behavior? American Economic Review, 104(6),

20 Diamond, Rebecca (2015) The Determinants and Welfare Implications of US Workers Diverging Location Choices by Skill: manuscript. DiNardo, John, Nicole Fortin and Thomas Lemieux Labor Market Institutions, and the Distribution of Wages : A Semiparametric Approach, Econometrica, 64, Galiani, Sebastian, Alvin Murphy and Juan Pantano (2012) Estimating Neighborhood Choice Models: Lessons from a Housing Assistance Experiment. American Economic Review Glaeser, Edward and Jacob Vigdor (2012) The End of the Segregated Century Guerrieri, Veronica, Daniel Hartley and Erik Hurst (2013) Endogenous Gentrification and Housing Price Dynamics, Journal of Public Economics, 100, Gould, Eric D., Victor Lavy, and M. Daniele Paserman (2011) Sixty Years after the Magic Carpet Ride: The Long-Run Effect of the Early Childhood Environment on Social and Economic Outcomes, Review of Economic Studies, 78, Kain, John F. (1992) The Spatial Mismatch Hypothesis: Three Decades Later, Housing Policy Debate, 3:2, Lee, Sanghoon & Jeffrey Lin (2014) Natural Amenities, Neighborhood Income Dynamics, and Persistence in the Spatial Income Distribution, manuscript. Neal, Derek & William Johnson (1996) The Role of Premarket Factors in Black-White Wage Differences, Journal of Political Economy, 104(5): Saiz, Albert (2010) The Geographic Determinants of Housing Supply Quarterly Journal of Economics, 125(3):

21 Table 1: Trends in Downtown Population and Demographic Composition Population Within 5 km of a CBD Total CBSA Share Share of Pop Within 5km of CBD that Lives in a Top CBSA Tercile Tract Tercile Criterion Fraction White Frac College Ed Mean HH Income SES Index (1) (2) (3) (4) (5) (6) Panel A: Levels ,382, ,332, ,973, ,967, ,846, Panel B: Decadal Changes ,050, , , , , Note: Numbers are calculated using data at the census tract level in each year. For Columns 2-6, each tract is weighted by tract share of CBSA population, such that each CBSA is equally weighted. Similar patterns exist for those living within the largest city in each CBSA.

22 Table 2: Share of Population within 5km of CBD in Tract Changing by at Least 20 Percentile Points 1/2 Standard Deviation up down up down Panel A: Fraction White % 11.9% 11.8% 13.8% % 8.3% 6.8% 9.8% % 5.7% 9.4% 9.3% % 5.4% 13.5% 9.0% % 5.5% 25.3% 19.8% Panel B: Fraction College Educated % 8.7% 9.8% 7.8% % 5.0% 6.4% 6.7% % 4.8% 5.2% 6.3% % 3.9% 12.5% 5.3% % 3.8% 19.7% 14.4% Panel C: Average Income % 9.2% 12.9% 14.8% % 2.1% 9.4% 12.6% % 2.4% 13.8% 6.2% % 3.4% 15.5% 11.8% % 3.5% 24.1% 20.3% Panel D: SES Index % 7.2% 5.1% 8.6% % 3.2% 3.5% 3.9% % 2.5% 5.3% 3.6% % 2.0% 9.1% 3.4% % 1.9% 20.0% 13.2% Notes: Numbers are calculated analogously to those in Table 1. Each tract is weighted by its share of CBSA population.

23 Table 3: CBSA Demand Shifts and Neighborhood Inequality Inequality Criterion Fraction Fraction Mean HH SES Period White College Ed Income Index Constant (0.014) (0.008) (0.017) (0.008) Ln(Employment), standard devs. (0.032) (0.013) (0.055) (0.013) Constant (0.011) (0.005) (0.025) (0.004) Constant (0.012) (0.006) (0.006) (0.005) Ln(Employment), standard devs. (0.023) (0.011) (0.025) (0.011) Constant (0.011) (0.005) (0.009) (0.004) Ln(Employment), standard devs. (0.053) (0.026) (0.064) (0.023) Constant (0.021) (0.012) (0.026) (0.007) Notes: Each column in each block reports coefficients from a separate regression of our CBSA gentrification index on the indicated variables and share married, share of population that are children, share college and share white in the base year. ln(employment) is expressed in standard deviation units and is instrumented with a Bartik quantity instrument, as explained in the text. Only periods with sufficiently strong first stages have reported employment coefficients. Reported coefficients on the constant can be interpreted as the mean index across CBSAs. Each regression has 118 observations and are weighted by initial year tract population share of the CBSA. First stage F statistics are ( ), ( ) and ( ).

24 Table 4: Patterns in SES Index Gentrification of Tracts within 5 km of CBDs Panel A: Bottom Tercile Neighborhoods 1(< 5km to CBD) (0.023) (0.019) (0.017) (0.021) (0.038) Employment Bartik * 1(< 5km to CBD) (0.022) (0.014) (0.017) (0.019) (0.040) Spatial Employment Bartik * 1(< 5km to CBD) (0.021) (0.009) (0.017) (0.016) (0.031) N 12,004 11,992 12,245 12,236 11,979 R-Squared Panel B: Middle Tercile 1(< 5km to CBD) (0.026) (0.022) (0.017) (0.019) (0.046) Employment Bartik * 1(< 5km to CBD) (0.018) (0.017) (0.020) (0.019) (0.046) Spatial Employment Bartik * 1(< 5km to CBD) (0.018) (0.018) (0.027) (0.021) (0.041) N 11,882 11,879 12,310 12,304 11,865 R-Squared Panel C: Top Tercile 1(< 5km to CBD) (0.041) (0.023) (0.020) (0.020) (0.044) Employment Bartik * 1(< 5km to CBD) (0.041) (0.023) (0.017) (0.022) (0.049) Spatial Employment Bartik * 1(< 5km to CBD) (0.042) (0.021) (0.023) (0.028) (0.050) N 12,016 12,009 12,322 12,316 11,990 R-Squared Notes: Each column in each panel reports results from a separate regression of the tract SES gentrification index on indicated variables and indicators for 5-10, and km from the CBD and 0-5, 5-10, and km from the nearest highest quartile SES index tract as of The Bartik variables are standardized to be mean 0 and standard deviation 1. Regressions are weighted by share of 1970 tract population in 1970 CBSA population.

25 Table 5: Patterns in Regression Adjusted Owner-Occupied Housing Cost of Tracts within 5 km of CBDs Panel A: Bottom Tercile Neighborhoods 1(< 5km to CBD) (0.025) (0.019) (0.015) (0.023) (0.029) Employment Bartik * 1(< 5km to CBD) (0.018) (0.013) (0.013) (0.017) (0.028) Spatial Employment Bartik * 1(< 5km to CBD) (0.015) (0.015) (0.011) (0.017) (0.024) N 9,426 11,217 11,640 11,265 10,813 R-Squared Panel B: Middle Tercile 1(< 5km to CBD) (0.020) (0.016) (0.012) (0.014) (0.024) Employment Bartik * 1(< 5km to CBD) (0.022) (0.013) (0.012) (0.014) (0.023) Spatial Employment Bartik * 1(< 5km to CBD) (0.016) (0.012) (0.014) (0.013) (0.022) N 10,377 11,719 12,177 12,084 11,605 R-Squared Panel C: Top Tercile 1(< 5km to CBD) (0.030) (0.034) (0.035) (0.015) (0.031) Employment Bartik * 1(< 5km to CBD) (0.025) (0.014) (0.014) (0.014) (0.024) Spatial Employment Bartik * 1(< 5km to CBD) (0.019) (0.051) (0.011) (0.016) (0.053) N 10,373 11,687 12,080 12,040 11,596 R-Squared Notes: Each column in each panel reports results from a separate regression of the tract owner occupied housing cost index on indicated variables and indicators for 5-10, and km from the CBD and 0-5, 5-10, and km from the nearest highest quartile SES index tract as of The Bartik variables are standardized to be mean 0 and standard deviation 1. Regressions are weighted by share of 1970 tract population in 1970 CBSA population. The housing cost index is formed from the residuals of a regression of log mean owner occupied home value on housing unit structure characteristics (number of units in building, number of bedrooms in unit, age of building) of the tract and CBSA fixed effects.

26 Table 6: Counterfactual Changes in Fraction of Population Share within 5 km of CBDs Data 1980 Shares 1980 Choices Race X All All (1) (2) (3) (4) Panel A: X=Education Panel B: X = Income Notes: Each entry in Column (1) is constructed with actual data. Remaining entries are calculated using counterfactual data. Entries in Columns (2) maintain the 1980 CBSA fraction of the population in each cell of the race conditional on education or income distribution. Entries in Column (3) maintain the 1980 CBSA fraction in the racexeducation or racexincome joint distribution. Entries in Column (4) maintain the 1980 neighborhood choices in each racexeducation or racexincome cell. The race conditional on income distribution and racexincome joint distribution use families in 1980 and households in subsequent years as the underlying unit of observation, thereby requiring a different baseline.

27 Table 7: Counterfactual Changes in Fraction of Central Area Population Living in a Top Fraction White Tercile Tract Data 1980 Shares All 1980 Choices Race X All (1) (2) (3) (4) Panel A: X=Education Panel B: X=Income Notes: Columns are analogous to those in Table 6, except they describe the change in the fraction of the population of central area tracts which live in a top tercile fraction white neighborhood. The top block examines the role of education composition and the bottom block examines the role of income composition.

28 Table 8: Counterfactual Changes in Central Area Population Living in a Top Tercile Tract Data All 1980 Shares All 1980 Choices (1) (2) (3) Panel A: Fraction of Central Area Population within Top Education Tercile Panel B: Fraction of Central Area Population within Top Income Tercile Notes: Columns are analogous to those in Table 6, except they describe the change in the fraction of the population of central area tracts which live in a top tercile neighborhood, as defined by education composition in Panel A and income composition in Panel B.

29 Share White Relative to Mean in Panel A Share White Relative to Mean in 1980 Share with College Degree Relative to Mean in Panel B Share with College Degree Relative to Mean in 1980 Best Linear Fit Tracts within 5km of CBD 45 degree line Tracts more than 5km from CBD Best Linear Fit Tracts within 5km of CBD 45 degree line Tracts more than 5km from CBD Panel C Panel D Income Relative to Mean in Mean Z Index in Income Relative to Mean in Mean Z Index in 1980 Best Linear Fit Tracts within 5km of CBD 45 degree line Tracts more than 5km from CBD Best Linear Fit Tracts within 5km of CBD 45 degree line Tracts more than 5km from CBD Figure 1: Chicago Tract Dynamics

30 Panel A Panel B Mean Z Index in Mean Z Index in Mean Z Index in Mean Z Index in 1980 Best Linear Fit Tracts within 5km of CBD 45 degree line Tracts more than 5km from CBD Best Linear Fit Tracts within 5km of CBD 45 degree line Tracts more than 5km from CBD Panel C Panel D Mean Z Index in Mean Z Index in Mean Z Index in Mean Z Index in 2000 Best Linear Fit Tracts within 5km of CBD 45 degree line Tracts more than 5km from CBD Best Linear Fit Tracts within 5km of CBD 45 degree line Tracts more than 5km from CBD Figure 2: Chicago Tract Dynamics Z-Index. By decade.

31 Figure 3a: Sample CBSAs shaded by 1980 fraction of central area population living in a top tercile SES index tract Figure 3b: Sample CBSAs shaded by 2010 fraction of central area population living in a top tercile SES index tract