The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
Federal Reserve Bank of Chicago The Long-Run Effects of Neighborhood Change on Incumbent Families Nathaniel Baum-Snow, Daniel Hartley, and Kwan Ok Lee March 18, 2019 WP 2019-02 https://doi.org/10.21033/wp-2019-02 Working papers are not edited, and all opinions and errors are the responsibility of the author(s). The views expressed do not necessarily reflect the views of the Federal Reserve Bank of Chicago or the Federal Reserve System. * The Long-Run E¤ects of Neighborhood Change on Incumbent Families Nathaniel Baum-Snow, University of Toronto Daniel Hartley, Federal Reserve Bank of Chicago Kwan Ok Lee, National University of Singapore March 18, 2019 Abstract A number of prominent studies examine the long-run e¤ects of neighborhood attributes on children by leveraging variation in neighborhood exposure through household moves. However, much neighborhood change comes in place rather than through moving. Using an urban economic geography model as a basis, this paper estimates the causal e¤ects of changes in neighborhood attributes on long-run outcomes for incumbent children and households. For identi…cation, we make use of quasi-random variation in 1990-2000 and 2000-2005 skill-speci…c labor demand shocks hitting each residential metro area census tract in the U.S. Our results indicate that children in suburban neighborhoods with a one standard deviation greater increase in the share of resident adults with a college degree experienced a 0.4 to 0.7 standard deviation improvement in credit outcomes 12-17 years later. Since parental outcomes are not a¤ected, we interpret these results as operating through neighborhood e¤ects. Finally, we provide evidence that most of the estimated e¤ects operate through public schools. 1 Introduction There is considerable empirical evidence that neighborhood and school environments are important determinants of human capital accumulation and long-run life outcomes. Chetty et al. (2018) show the existence of considerable heterogeneity in rates of intergenerational mobility across census tracts in which children grow up, even those within a few miles of each other. Chetty et al. (2016), Chyn (2016), Chetty & Hendren (2018a, b) and Laliberté (2018) show that children with longer We thank Raj Chetty, Gilles Duranton, Remi Jedwab, Santiago Pinto, Tony Yezer and various seminar participants for valuable comments on earlier drafts of the paper. The views expressed are those of the authors and do not necessarily represent the views of the Federal Reserve Bank of Chicago or the Federal Reserve System. 1 exposure to lower poverty neighborhoods and areas with higher rates of intergenerational income mobility have higher earnings and educational attainment in adulthood. Gould et al. (2011) and Damm & Dustmann (2014) show that attributes of the neighborhoods to which immigrant children are randomly assigned have large e¤ects on incomes and the propensity to commit crimes in adulthood. These and other existing studies identify causal e¤ects of neighborhoods through quasi-randomization achieved by observing household moves to new neighborhoods. While evidence on neighborhood e¤ects through household moves comes with appealing identi…cation properties, Aliprantis & Richter (2018) and Chyn (2018) demonstrate wide treatment e¤ect heterogeneity, even among the public housing population that is their focus. The low takeup rate of housing voucher o¤ers among public housing residents observed in the Moving to Opportunity experiment analyzed by Chetty et al. (2016) and Aliprantis & Richter (2018) suggests that, absent forced moves through demolitions, the largest e¤ects of neighborhood change may occur in place rather than through household moves. Despite considerable policy interest about the impacts of gentrifying neighborhoods on incumbent residents, there is relatively little evidence in the literature about these e¤ects. This paper provides estimates of causal e¤ects of neighborhood change on long-run outcomes for parents and children in incumbent households. We isolate variation in changes in residential neighborhood demographic composition using skill-oriented labor demand shocks to potential commuting destinations. This allows us to separate out the e¤ects of neighborhood change on children that run through neighborhoods from those that run through wealth e¤ects of the children’s parents. Our key treatment variable is "Resident Market Access" (RMA), a commuting time discounted aggregate of employment accessible from each residential census tract that also incorporates competition e¤ects in labor supply from other residential locations. Tsivanidis (2018) shows that RMA is a conceptually appealing measure, as it exhibits iso-elastic equilibrium relationships with income net of commuting cost, housing prices and population in an urban economic geography model similar to that developed in Ahlfeldt et al. (2015), based on Eaton & Kortum (2002). While our key treatment variables are theoretically founded, they are also highly correlated with intuitive measures of commuting time discounted aggregates of employment by skill. As such, our analysis does not depend on model structure to be informative. For identi…cation, we isolate (conditionally) exogenous variation in RMA growth rates through Bartik (1991) type skill-oriented labor demand shocks to employment locations within short commuting times of residential locations. We build instruments by constructing counterfactual post1990 RMA for each census tract using 1990 employment shares by industry in each tract and national industry growth rates excluding the metropolitan region in question. To strengthen identi…cation and limit the potential for trends in local consumer amenities to in‡uence our results, we exclude labor demand shocks in the census tracts of residence when constructing the instruments, condition on neighborhood attributes that may be correlated with industry composition in nearby 2 employment locations and make use of variation within metro areas interacted with 2-2.5 km wide distance bands from central business districts. We show that these three measures eliminate correlations between instruments and pre-treatment trends of demographic variables of interest in most settings we analyze. Selection of skilled employment growth into more a- uent and educated neighborhoods, which exhibit lower than average rates of gentri…cation, means that OLS regressions tend to understate the true causal e¤ects of nearby employment growth on neighborhood change. Severen (2019) uses a similar strategy to structurally estimate parameters governing the Los Angeles area economic geography, facilitating a welfare analysis of the Los Angeles Metro Rail construction during the 1990s. We measure outcomes for four separate samples of individuals who were treated with neighborhood change in the 2000-2005 or 1990-2000 periods. Our primary analysis uses panel data on about 10,000 children in the 1985-1989 birth cohorts of the Federal Reserve Bank of New York Consumer Credit Panel / Equifax (CCP) data, and their parents. For this group, we observe information on individuals’credit records (credit score, credit card limits, loan delinquencies, mortgages, etc.) plus census block of residence in years 2000 through 2017. We also examine outcomes of 1990-2000 neighborhood change on about 1,500 children born 1972-1989 in the Panel Study of Income Dynamics (PSID) and their parents. Each outcome data set has its advantages and drawbacks. The PSID data include more informative outcome measures, allow us to look at younger children at treatment, and allow us to follow people for a longer time period after treatment. However, its smaller sample size results in wider con…dence bands and less scope for investigation of heterogeneity in treatment e¤ects. Moreover, because 1990 microgeographic information is used to build the instruments, the 1990-2000 period presents more identi…cation challenges. The CCP has larger sample sizes and exists for a time frame with arguably better identi…cation, but only allows us to see proxies for income and examine the e¤ects of neighborhood change for at most 17 years. Our sample region includes the 254 metropolitan areas for which data on 1990 employment by industry exists at the microgeographic level from the Census Transportation Planning Package (CTPP) and for which census tract-level data existed in 1970. Our results indicate that labor demand shocks oriented toward those with a college degree cause neighborhood change, measured as increases in either the share of residents in the neighborhood with a college degree or as increases in a composite index of neighborhood quality, conditional on labor demand conditions for those with less than a college degree in both the 1990-2000 and 2000-2007 periods. For the later period, a one standard deviation greater increase in collegedegree oriented RMA (henceforth "skilled RMA") is estimated to cause a neighborhood to move up its metropolitan area’s distributions of growth in college fraction by 23-38 percent of a standard deviation and growth in neighborhood quality by 14-24 percent of a standard deviation. While these estimates are more precise for suburban areas, the point estimates are similar for central and suburban regions. 3 Our evidence from the CCP data indicates that neighborhoods that change in the direction of higher educational attainment or higher quality have positive long-run e¤ects on incumbent children growing up in these neighborhoods in suburban areas only. The e¤ect of these shocks appears to run through neighborhood or school channels rather than through parent wealth e¤ects. In particular, we …nd that 11-15 year old children that experience a one standard deviation higher skilled labor demand shock in suburban neighborhoods have about an 18 point gain in their credit scores, a $2,000 higher credit limit and are 7 percentage points more likely to have a mortgage (a proxy for owning a home) 17 years later. These numbers represent about 15 percent of a (cross-sectional) standard deviation for this cohort. The estimates grow from about zero at ages 21-25 and are typically slightly greater for children growing up in the least educated neighborhoods. We present compelling evidence that exposure to improved school quality rather than other types of neighborhood e¤ects primarily drives our results. Conditioning on school district …xed e¤ects reduces the estimated e¤ects on children to about zero. That is, variation in gentri…cation across neighborhoods within the same school district has no estimated e¤ect on long-run outcomes of resident children. This evidence matches that of Laliberté (2018), who uses unique data from Montreal to estimate that about 80% of "neighborhood e¤ects" run through school quality. Using variation between school districts, we …nd that the e¤ects are larger for children growing up in higher quality school districts. This may re‡ect the e¤ect of school quality on the propensity for college educated parents to send their children to the local public schools. We show that there are no e¤ects of shocks to skilled RMA on credit outcomes for the parents of children in our sample, indicating no evidence that parental wealth e¤ects are driving the results. Moreover, we …nd that once their children leave the home, parents whose neighborhoods gentri…ed after 2000 choose to move to neighborhoods with educational attainment compositions looking much like the ones in which they lived in 2000 (pre-gentri…cation). This comes despite the fact that their children, after a period of living in less educated neighborhoods in their 20s, upgrade their neighborhoods through migration to neighborhoods that have a higher share of residents with a college degree (henceforth, "college share" or "fraction college") by about 1-2 percentage points above the direct e¤ects of the shocks on their 2000 neighborhoods. These extra e¤ects are greater for those that grow up in the least educated neighborhoods. Exposure to higher quality neighborhoods in youth leads to the choice of living in higher educated neighborhoods in adulthood. In summary, our CCP results indicate that a one standard deviation increase in neighborhood college share leads to 40-70 percent of a standard deviation improvement in incumbent child outcomes, on average, in the suburbs. For those growing up in the most disadvantaged neighborhoods, the e¤ect is on the higher end of this range. School quality appears to be the main driver of these e¤ects, with larger e¤ects in higher quality school districts. Our results for the 1990-2000 period corroborate those for the later period but come with additional empirical challenges. Higher 1990-2000 skilled RMA growth predicts higher test scores 4 for young adults and higher 2015 employment rates and family incomes. It is di¢ cult to use the PSID to recover the mechanisms that drive the reduced form 1990-2000 treatment e¤ects for several reasons. First, we …nd evidence that parent’s incomes are a¤ected in addition to child outcomes. Second, we estimate large con…dence intervals for some outcomes of interest. Finally, there are limits to the possibility of breaking out the e¤ects by parental education due to a lack of statistical power. The positive wealth e¤ects that we measure for the parents are likely due to the higher conditional correlation of skilled and unskilled RMA shocks during this period making it di¢ cult to isolate a shock that makes the neighborhood better but has no e¤ect on the parents. Our evidence on the e¤ects of neighborhood change highlights a potentially important force driving increased income inequality and reduced intergenerational mobility. More educated households have been disproportionately exposed to improvements in neighborhood quality in recent decades. Figure 1 shows kernel density graphs of 1990-2000 and 2000-2010 changes in the share of one’s Census tract with a college degree for resident children whose parents are imputed to have either less than a high school degree or a college degree or higher educational attainment. In both decades, but especially in the 1990s, there is clear evidence that the distribution of neighborhood upgrades for college graduate residents has more mass on the right side than the distribution for residents without a high school degree. This trend reinforces the 1990 baseline in which less educated children already live in predominately less educated neighborhoods (seen in Figure A1). Such exposure to educated neighbors can have important long-run impacts. Indeed, Fogli & Guerrieri (2018) calibrate an OLG model with neighborhood choice to show that magnitudes of neighborhood e¤ects in line with those estimated in this paper and Chetty & Hendren (2018a, b) interacted with a shock to the distribution of skill prices generate changes in the distribution of skill quantities that increase income inequality by 25-40 percent beyond the impact of the skill price shock alone. With high returns to neighborhood quality, the logic of revealed preference would indicate that people should migrate toward more educated neighborhoods, all else equal. Figure 2 shows evidence to this e¤ect, but also that the choice to migrate does not depend as much on future neighborhood change as it does on initial neighborhood quality. Panel A shows distributions of 2000 fraction college by the choice to migrate to a di¤erent tract by 2017 for parents in our CCP sample (left side) and their children (right side). The left graph shows clear selection of moving parents from the least educated neighborhoods, relative to stayers (the red dashed line is to the left of the blue solid line). In the CCP sample 69 percent of the parents end up moving to a new tract by 2017. These movers are much more likely to come from less educated neighborhoods, perhaps in order to invest more in their children. However, the right side shows that the same pattern is less pronounced for their children. Given that 85 percent of children have migrated to a new tract by 2017, mostly to establish their own households, it is not surprising that the ones that remain in the same neighborhood do so for reasons other than neighborhood attributes. The evidence in Panel B of Figure 2 shows that the educational composition in migration 5 destination neighborhoods indeed suggests that households value having more educated neighbors. However, the distribution of neighborhood change from which migrants depart looks very similar to that for households that do not move. Panel B depicts three distributions of changes in fraction college for parents and their children. The solid blue lines are the distributions of 2000-2007 changes in fraction college for the year 2000 tract of residence for those that are living in the same tract in 2017 as in 2000. The short-dashed red lines depict the distributions of 2007 fraction college in 2017 tract of residence minus 2000 fraction college for the 2000 tract of residence amongst movers only. The long-dashed green lines are the distributions of 2000-2007 changes in fraction college in movers’ 2000 tracts of residence. Comparing the red and green lines in the two graphs indicates that the children use mobility to upgrade their neighborhood quality more than their parents (red distribution has more mass to the right for the children). However, the green and blue lines coincide very closely in both Panel B graphs, indicating little selection of movers on the subsequent changes in neighborhood quality of their initial neighborhoods. Beyond using identifying variation from labor demand shocks in commuting destinations, our evidence of more selection on levels-of than changes-in neighborhood college fraction further supports our empirical strategy of using variation in neighborhood change for identi…cation. The results shown in Figure 2 are similar when broken out by the educational attainment of the household. This paper complements the existing literature on neighborhood e¤ects by presenting estimates that apply to a broad population, including those who choose not to move, and to more local neighborhoods relative to many of the best identi…ed estimates in the literature to date. Chetty et al. (2018a, b) make causal statements about children in households who choose to move across commuting zones or county boundaries and can measure "neighborhood e¤ects" down to the county level. Laliberté (2018) corroborates their estimates for movers within Montréal. While they can estimate the e¤ects of census tract attributes, Chyn (2018) and a series of papers about the Moving to Opportunity program including Chetty et al. (2016) and Aliprantis & Richter (2018) are limited to estimating e¤ects for public housing residents who may not be representative of the broader population. Altonji & Mans…eld (2018) use assumptions about the ability to invert the local amenity vector into observable characteristics of neighborhood residents to identify lower bounds on neighborhood e¤ects. Using restricted access census data, Brummet and Reed’s (2019) empirical setting is perhaps most similar to ours; both papers …nd consistent evidence that high rates of household mobility insulate incumbents from negative e¤ects of gentri…cation, and also …nd no long-run positive or negative e¤ects for urban children. With rapid gentri…cation occurring in the centers of many U.S. cities (Baum-Snow & Hartley, 2018; Couture & Handbury, 2017; Edlund et al., 2016), the e¤ects of gentri…cation on incumbent residents has particular current policy relevance in the United States. In addition to its contribution to the neighborhood e¤ects literature, this paper also relates to literatures about the long-run e¤ects on workers of job loss and shifting labor market opportunities plus the intergenerational e¤ects of 6 parent wealth shocks. Davis & von Wachter (2012) and Couch & Placzek (2010) show that job loss has persistent e¤ects. Heisz, Oreopolous and von Wachter (2012) show similar long run deleterious e¤ects of graduating college in a recession, especially for less able graduates. Dahl & Lochner (2012) and Hilger (2015) …nd that negative parental wealth shocks have small e¤ects on child college enrollment probabilities and long run labor market outcomes of their children. Our study shows that even if parental wealth e¤ects are small, children’s outcomes can be a¤ected in the long run through neighborhood change due to spatially correlated shocks to labor market opportunities for parents. This paper proceeds as follows. Section 2 lays out our estimation problem conceptually and shows how we separate out the e¤ects of neighborhood change that run through parents from more direct e¤ects on children. Section 3 discusses the data. Section 4 explores the neighborhood-level identifying variation in the data. Section 5 provides a theoretical framework that de…nes our key RMA predictor variables. Section 6 presents the details of our empirical implementation, including the construction of the instruments. Section 7 presents our results. Section 8 concludes. 2 Empirical Framework This section lays out the equations that we aim to estimate and shows how our estimation procedure facilitates separating out parental wealth e¤ects that are caused by the labor or housing market from neighborhood e¤ects. We face a hierarchical estimation problem within each metropolitan area. At the top of the hierarchy, a vector of demographic attributes and housing costs, ni in each neighborhood i depends on neighborhood-speci…c labor demand conditions and local amenities. Below neighborhoods are parents, whose outcomes may depend on neighborhood attributes, labor demand conditions, and some pre-determined factors like their human capital. Finally, children’s long-run outcomes depend on parental inputs when they are children, neighborhood attributes and neighborhood amenities. In general terms, our empirical approach is to focus on the variation in neighborhood change that is induced by labor demand shocks oriented toward college-educated (henceforth, "skilled") workers, while conditioning on unskilled labor demand shocks. We show below that this variation does not generate direct wealth e¤ects through the labor market for incumbent resident parents. For unskilled parents, our direct conditioning on low skilled labor demand shocks holds labor demand conditions constant. Since skilled parents typically move to a new neighborhood to take a new job, we show that incomes of incumbent skilled residents are also una¤ected by the skilled labor demand shocks hitting neighborhood i. Anticipating the discussions of the data and theoretical framework in Sections 3 and 5, we lay out our targeted estimation equations in each level of the hierarchy below. When discussing the practical identi…cation challenges in Section 6, we …ll in more details about the exact empirical 7 speci…cations used. 2.1 Neighborhoods Equation (1) describes our conceptualization of the data generating process for the change between periods t 1 and t in the vector of neighborhood demographic attributes and housing costs, ni . We denote changes in labor market opportunities for skilled workers living in tract i as t ln RM ASi and unskilled workers living in tract i as t ln RM AU i . The details of how we measure these objects are developed in Section 5. Our primary goal in the neighborhood analysis is to estimate the parameter nS in the equation below, which is the treatment e¤ect of labor market opportunities for skilled workers in and near tract i on attributes of tract i, while holding unskilled labor market opportunities constant. Xi is a set of pre-determined observed tract characteristics and amenities, conditional on which the instruments for t ln RM ASi and t ln RM AU i are exogenous. t ni = n 0 + nS t ln RM ASi + nU t ln RM AU i + Xi n + n i (1) We will estimate the parameter vector nS using instrumental variables (IV). The speci…cs of our IV strategy are described in detail in Section 6. 2.2 Parents The following equations describe the process that we conceptualize generates the data on labor market and credit outcomes for children’s parents, indexed by !, at time t who live in neighborhood 1 i in period t 1. U zi!; S zi!; = pU 0 = pS 0 + pU S + pS S t ln RM ASi + S t ln RM Ai + pU U pS U t p ln RM AU i + Xi! U t ln RM Ai + pU p pS Xi! + + pU i! pS i! (2) (3) In these equations, we condition on the same controls Xi as for the neighborhood analysis above plus some additional parent-speci…c controls in the base period (t 1). The e¤ects of nearby changes in labor demand conditions may run through the labor market, the housing market or through changing spillovers from neighbors. For example, consider pU S , the average e¤ect of skilled RMA growth on unskilled parents’ outcomes. pU S incorporates a direct e¤ect holding neighborhood demographic composition constant and an indirect e¤ect that 1 In >t the data, we follow parents as they move, thus they do not have to remain in neighborhood i in periods 1. 8 runs through its impact on neighborhood demographic composition pU S = E[ U @zi!; j @ t ln RM ASi t p ln RM AU i ; Xi! ; t ni ] + nS E[ nS . U @zi!; j @ t ni t p ln RM AU i ; Xi! ] @z U p j t ln RM AU We surmise and con…rm empirically that the direct e¤ect (E[ @ t lni!; i ; Xi! ; t ni ]) RM AS i is equal to zero since, conditional on low-skilled labor demand shocks, high skilled labor demand shocks should have no e¤ect on job opportunities for low-skilled parents. What remains is the e¤ect @z U p that runs through neighborhood characteristics ( nS E[ @ i!; j t ln RM AU i ; Xi! ]). This component t ni incorporates both housing wealth or rent e¤ects and changing spillovers from the demographic @z U p composition of tract i. Thus, given estimates of pU S , we can recover E[ @ i!; j t ln RM AU i ; Xi! ] t ni using estimates of nS from the neighborhood equation. In practice, we estimate pU S and pS S to be zero or slightly negative. As with the neighborhood equation, in the empirical implementation we instrument for t ln RM ASi and t ln RM AU i when estimating Equations (2) and (3). 2.3 Children Since we only observe outcomes for children after they become teenagers or adults, we examine U S levels of outcomes yi! or yi! of children of unskilled or skilled parents ! who lived in tract i in U period t 1. These outcomes are observed after period t. The data generating processes for yi! or S yi! are similar to those for the parent outcomes, as follow: U yi! = cU 0 + cU S S yi! = cS 0 + cS S t ln RM ASi + cU U t ln RM ASi + cS U t t c ln RM AU i + Xi! c ln RM AU i + Xi! cU cS + + cU i! (4) cS i! (5) c Here, we use a similar set of pre-determined household-speci…c controls, Xi! as for the parents. To be consistent with the literature on the e¤ects of youth environment on childrens’ human capital accumulation (e.g. Hoynes, Schanzenbach and Almond, 2018), these controls include some observed parental inputs prior to the labor demand shock treatment. For the children, we have a similar but more complicated interpretation of coe¢ cients on RMA than for the parents. For example, the average e¤ect of skilled RMA growth on children of unskilled parents is: cU S = pU S + E[ nS U @yi! j U @zi!; E[ U @yi! j @ t ni t c ln RM AU i ; Xi! ] + E[ t c ln RM AU i ; Xi! ] 9 U @yi! j @ t ln RM ASi t c ln RM AU i ; Xi! ; t ni ] The …rst term is the e¤ect of the labor demand shock that runs through parental inputs. Given that we will estimate pU S from the parents’equation and con…rm that it is about zero, we will impose that this …rst term is zero. The second term is the direct e¤ect of skilled labor demand shocks on children’s long-run outcomes, holding neighborhood and parental attributes constant. While it may be di¢ cult to imagine that this term is also not zero, we keep it in explicitly given Charles et al.’s (2016) evidence on the incentive e¤ects of labor demand conditions for teens’human capital accumulation. Finally, the last term includes the direct e¤ect of neighborhood demographic change that we are after. If the parental wealth e¤ect and the direct e¤ect of the skilled labor demand @y U c shock are zero, we can directly calculate the neighborhood e¤ect, E[ @ ti!ni j t ln RM AU i ; Xi! ] = cU S nS = . As with the neighborhood and parent analyses, in the empirical implementation we will instrument for t ln RM ASi and t ln RM AU i when estimating Equations (4) and (5). In this section, we have laid out the data generating process as a set of reduced form equations that relate changes in labor demand conditions in each neighborhood to outcomes of interest. In principle, one could set this up as a system of equations to be estimated jointly by 3SLS or GMM, allowing for one to recover estimates of neighborhood e¤ects in one step. We do not do so for two reasons. First, this process would not accommodate separate estimation of a potential direct e¤ect of labor demand shocks on children. While we cannot separately identify such an e¤ect, we do not want to assume away its existence. Second, the various identi…cation challenges and speci…cation checks laid out in Section 6 make it more straightforward to estimate parameters in the reduced form hierarchical system and combine them afterwards. This allows for more ‡exibility in mixing and matching di¤erent parameter estimates to recover estimates of neighborhood e¤ects on the @y U c children of unskilled parents, E[ @ ti!ni j t ln RM AU i ; Xi! ] and neighborhood e¤ects on the children S @yi! j t ni of skilled parents E[ @ 3 t c ln RM AU i ; Xi! ], our main objects of interest. Data Our analysis makes use of census tract aggregate data from 1970 to 2014, commuting and place of work tabulations from the 1990 and 2000 Census Transportation Planning Packages (CTPP), Decennial Census and American Community Survey (ACS) micro data from 1990 to 2005, LODES commuting and place of work data from 2010, micro data from years 1972 forward from the Panel Study of Income Dynamics (PSID) and micro data from 2000 forward from the Federal Reserve Bank of New York Consumer Credit Panel/Equifax (CCP). We normalize all microgeographic units in these datasets to Census 2000 tract boundaries. 10 3.1 De…ning Our Study Areas The Census Bureau tabulates the 1990 and 2000 Decennial Census micro data to place of work, place of residence and directional commuting ‡ow to form the Census Transportation Planning Package (CTPP) data sets. The 1990 CTPP geography dictates how we construct our study regions. The 1990 CTPP assigns microgeographic units the size of census tracts or smaller to "regions", which roughly correspond to metropolitan areas. Total commuting ‡ows are reported for each pair of census tracts, tra¢ c analysis zones or block groups within each region, with no information reported on between-region ‡ows. Some regions overlap. Connecticut and surrounding areas and New Jersey and surrounding areas are each de…ned as one large region.2 For Connecticut and New Jersey, we de…ne new regions that each have a minimum 25 km radius around each central business district (CBD) in each state. Tracts in these CTPP regions that are beyond 25 km from all CBDs in each state are assigned to the closest CBD. The CTPP reports the mean and median home to work commute times for each pair of microgeographic units with a positive commute ‡ow. Employment in 18 industry groups by place of work are also reported, including the 6% of the employed workforce who worked at home in 1990. The reported commuting ‡ows do not include those who work at home. We map the 1990 CTPP geography to 1990 census blocks using Census Bureau reported allocation factors and use the 1990-2000 Census Block Relationship File to convert to Census 2000 tract boundaries. We use land area to form allocation weights in both conversions. We assign one CBD to each region, with its location calculated as the centroid of the set of CBD census tracts reported in the 1982 Economic Census for the region’s largest city. Those regions without a CBD in the 1982 Economic Census are assigned one based on a visual assessment of the location of city hall and the oldest bank branches in the city. Measuring the employment opportunities available in each residential census tract is central to our analysis. As we lay out in the theoretical framework in Section 5, we want to think of each region as a local labor market in isolation in which workers choose residential locations anticipating employment options available in each census tract in the region. Because we do not observe employment locations for region residents who commute beyond region borders in 1990, we organize the data to minimize the potential importance of this type of reverse commute. We measure total 1990 employment in tract j by aggregating over all commute ‡ows to j from both inside and outside the region, with a single assigned residential location outside the region. We measure the number of resident workers in tract i as the aggregate of commute ‡ows from origin i to destinations in the region only. We build all demographic, employment and commuting data described below for the 63,897 Census 2000-de…nition tracts in 306 regions.3 2 In the two cases in which overlapping regions have the same CBD (Portland, OR and Greensboro, NC), we keep only the most expansive region for the analysis. 3 We have 50,410 unique census tracts in our data, with 41,627 of these appearing in one region only. 11 Our empirical analysis relies on accurate measures of historical demographic characteristics and viable employment opportunities within commuting range. To this end, our analysis excludes regions without valid 1970 demographic information. In the remaining 254 regions, we focus our analysis on residents of the census tracts that are within 20 km of the CBD and with valid 1970 demographic information. We further constrain the sample to leave at least 10 km between each sample census tract and the region’s edge, so that we can accurately observe labor market opportunities in all commuting directions. Altogether, the result is 32,515 census tracts (28,476 of which are unique) whose residents are counted in the empirical work. However, we emphasize that we incorporate information from all potential commuting destinations outside of this sampled residential area as long as they are within a 1990 de…nition CTPP region. 3.2 Post-1990 Commuting and Employment Data We use the 2000 CTPP to construct commuting ‡ows, commute times and employment in each of 14 industries for year 2000. Unlike the 1990 version, the 2000 CTPP covers all commutes and employment in the U.S. down to the census tract level or below. We organize it to measure objects of interest within the 1990 de…nition region geographies described earlier. For 2010 commuting employment data, we process the LODES aggregation of the Longitudinal Employer Household Dynamics data. This data set has employment by industry and education plus commuting ‡ows for each census block in the U.S. However, it does not include commute times.4 3.3 Demographic Information We take census tract aggregates for 1970-2010 from the Decennial Census derived Neighborhood Change Database supplemented with some Summary Tape File 4 variables from 1980, as described in Baum-Snow & Hartley (2018), and some 2005-2009 and 2012-2016 tract aggregates from the American Community Survey (ACS). We use these data sets to measure aggregate outcomes and to control for pre-treatment trends. We note that because of the smaller ACS samples and the longer time windows the 2005-2009 and 2012-2016 tract aggregate data is noisier than than the 2000 census aggregate data. Following Aliprantis and Richter (2018), we construct a summary index of neighborhood quality for use throughout the analysis. This index is calculated as the …rst principal component of the nationwide cumulative distribution functions of fraction high school or more, fraction college or more, the negative of the poverty rate, the employment to population ratio, the negative of the unemployment rate, and the negative of the share of single headed households. The result is a percentile rank for each census tract nationwide. During the 2000-2007 period, tracts in our sample lost 0.9 percentiles in neighborhood quality on average, with a standard deviation of 13.1 percentiles. 4 Since Massachusetts is not included in the 2010 LODES place of work …le, we use its 2011 …le instead. 12 The average within-region standard deviation in the 2000-2007 change in neighborhood quality is 12.2 percentiles. These changes in quality are positively correlated with the change in fraction college, which exhibits a within-region standard deviation of 7 percentage points for its 2000-2007 change. This substantial variation in neighborhood change is the key treatment object of interest in this study. Table 1 Panel A presents summary statistics of relevant census demographic variables. 3.4 Commute Times Our empirical work requires information on commute times between each pair of census tracts in each region in 1990 and 2000. Since the CTPP only uses reports from the roughly one out of every six households who received the Decennial Census long form and ‡ows of fewer than 5 sampled workers are suppressed, many commutes and commute times are not observed in our data. Nevertheless, the CTPP is the most complete historical data on commute times between microgeographic units in a large number of U.S. cities. In particular, we observe this information for 7.4 million tract pairs in 1990 and 6.3 million tract pairs in 2000 in our sample of 254 regions. So as to limit the in‡uence of outliers, we focus on pairwise median commute times. The 1990 ‡ow-weighted median of median commute time in our sample area is 20 minutes with a standard deviation of 15.1 minutes and a distribution that is skewed to the right. The 2000 median commute time rose to 20.8 minutes with a standard deviation of 19.8 minutes. To …ll in the remaining commute times, we develop an empirical forecasting model based on distances between tract centroids and to the region’s CBD. We recognize that observed commutes may be subject to more road congestion than the less common commutes which we must forecast. As such, our prediction model may deliver overestimates of true commute times. However, the headline comparisons in the main empirical work below are for residential locations within CBD distance rings and thus should be subject to similar predicted commute time biases.5 After experimenting with a number of ‡exible forecasting models, we settled on the following simple forecasting equation. ln m ij = d ln Distanceij + +vm + r ln (Residence CBD Dis)i + w ln (Work CBD Dis)j um ij Here, the commute time from tract i to tract j in region m is constant elasticity in distance between i and j plus the CBD distances from home and work. The region …xed e¤ects allow average travel speeds to di¤er across regions (Couture et al., 2017). Including the two CBD distance terms adds 5 Allen, Arkolakis & Li (2017) uses travel times calculated using the Fast Marching Method algorithm instead. This also does not account for equilibrium e¤ects due to changes in congestion and would be di¢ cult to implement for 1990. 13 about 0.02 to the R-Squared. Adding additional terms to separate out radial from circumferential travel in a ‡exible way and/or introducing heterogeneity in the estimated elasticities adds less than 0.03 to the R-Squared. The estimated parameters are reported in Table A1. Our estimated elasticity of travel time with respect to distance is about 0.43. Starting or ending the trip 10% further from the CBD takes 0.7 percent less time, re‡ecting faster average travel speeds in the suburbs. The forecasting model …ts reasonable well with within R-squared values of 0.53 in 1990 and 0.50 in 2000. Figure A1 shows a graph of the region …xed e¤ects. Travel times in 2000 were highest in the Boston, Chicago, Jersey City, Los Angeles, Newark, New York, Paterson, San Francisco, Trenton, and Washington areas, for a given trip distance and origin and destination CBD distances.6 These parameters are used to predict bij for the location pairs between which we do not observe commute times, while incorporating that error terms in each region are drawn from independent normal distributions with di¤erent variance parameters. As may be expected, distributions of predicted commute times are typically longer than observed travel times, as short commute times are more likely to attract commuters in equilibrium. 3.5 Federal Reserve Bank of New York Consumer Credit Panel / Equifax (CCP) The CCP has information about block of residence, birth year, loan balances and creditworthiness for a random 5% nationwide sample of people with a social security number and a credit record. The sample runs from 1999 to the present. We use this data to construct credit histories of children born 1985-1989 and their parents starting in 2000. The observation counts from the CCP indicate that about 85% of the U.S. population in 2017 had a social secuity number and credit history. This share is stable across the age distribution.7 As in Chetty & Hendren (2018a, b), we identify "parents" as anybody coded to the same address as the child that is 16-45 years older than them in the …rst year we observe the child in the sample, typically at age 19-21. The tracts in our sample area contained about 133 million resdents as measured by the 2000 Census. Five percent of that is 6.55 million people. As of 2000Q1, there were 3.9 million people in our sample area in the CCP. This number is lower than 5 percent of the population primarily due to the fact that not everyone has a credit history, especially children under 18. This means that in order to determine where young adults in 2017 lived in 2000 when they were children, we must link them to an adult that is also in the 5 percent sample. We can only follow parents’residential locations back in time if they are also sampled, meaning that we have in essence a 5%2 = 0:25% nationwide random sample that we can use for analysis. We focus on children born from 19856 The New York region geography overlaps with those for Newark, Jersey City and Paterson. coverage is also good in earlier years. In 2000Q1, we observe 78% of 25-29 year olds and 85% of 30-34 year olds in the CCP, with this share above 81% for all older age groups. 7 The 14 1989, making them 11-15 years old in 2000 when we observe their parent’s residential location, and 28-32 in 2017 at the end of the sample period. This restriction results in a sample size of 10,859 parent-child pairs. There is a slight complication due to the fact that the CTPP regions can overlap. In these cases, we include any CCP individuals in the sample for for each of the overlapping regions, meaning they appear in the estimation sample multiple times. However, we assign them a weight of one divided by the number of CTPP regions that their tract is in. All of our estimation results use these weights. While educational attainment is not reported in the CCP, we use the sex-race distribution of the parent’s 2000 Census block and the sex-race-educational attainment distribution of their 2000 Census block group to compute weights which represent the probability that each parent is in one of four educational attainment groups: less than a high school degree, high school graduate, some college, and college degree or higher. When we report estimates by educational attainment group for the CCP, we weight by the product of these weights and the weights discussed in the previous paragraph that account for overlapping CTPP regions. When we report observation counts for the CCP results in Tables 8-10 we report the sum of the weights, re‡ecting the number of unique individuals represented in each speci…cation. Table 1 Panel B presents summary statistics of the CCP data. 3.6 PSID The geocoded PSID follows households over time, allowing us to look at outcomes for children through 2015 living in households hit by labor demand shocks in the 1990s. We focus on the 1,570 children in the PSID that were between the ages of 0 and 18 in 1990, were living with at least one of their parents and lived in our study region described above. Due to siblings and cluster sampling, only 684 census tracts are represented. Table 1 Panel C presents summary statistics for the PSID sample. 4 Tract Level Analysis Our main goal is to estimate causal e¤ects of increases in neighborhood quality on outcomes of incumbent resident children. While our main empirical analysis makes use of the full distribution of employment and population across all census tracts in each region, our fundamental source of identifying variation is in interactions between the 1990 industrial composition of census tract employment and subsequent national industry employment growth rates. In this section we show that tract-level Bartik (1991) type labor demand shocks successfully predict tract-level employment growth. However, separating out skilled and unskilled labor demand shocks at the neighborhood level is only possible for the 2000-2010 period. In Section 6, we explain how we spatially aggregate 15 these tract-speci…c shocks into market access shocks that measure exogenous variation in skillspeci…c labor demand conditions facing each residential neighborhood and justify the conditions required for their use in building our instruments. 4.1 Construction of Tract-Level Shocks We adapt the widely used Bartik (1991) local labor demand shocks to isolate exogenous variation in 1990-2000 and 2000-2010 tract-speci…c labor demand and employment growth.8 These shocks are constructed by predicting employment growth using 1990 tract industry composition as weights interacted with average national growth rates across industries among college graduates. This type of measure has been widely used to isolate exogenous variation in labor demand in empirical work on local labor markets going back to Blanchard & Katz (1992). Our implementation has some similarities to that in Diamond (2016), who also uses Bartik shocks for identi…cation while interpreting them as a component of skill group-speci…c productivity shocks in the context of a general equilibrium model of local labor markets. Unlike Diamond (2016), however, we employ these shocks to isolate variation in labor demand conditions across locations within metropolitan regions (as in Couture & Handbury, 2017), rather than between metropolitan regions. As a result, the assumptions required for identi…cation, discussed further below, are somewhat di¤erent. The key goal is to control for any variation in tract employment growth that may come from di¤erential trends within metro areas in amenities, housing productivity or unobserved initial demographic conditions. We construct the following two tract-speci…c Bartik shocks for the 2000-2005 period experienced by census tract j, where k indexes industry: BartikjS = X k BartikjU = X k 90 90 m0 (j)kS Empjk X 90 90 m0 (j)kS Empjk 05 [ln Em 0 (j)kS 00 ln Em 0 (j)kS ] (6) 05 [ln Em 0 (j)kU 00 ln Em 0 (j)kU ] (7) k 90 90 m0 (j)kU Empjk X 90 90 m0 (j)kU Empjk k In these equations, Emp90 jk is the number of workers in tract j and industry k in 1990 taken from 90 the CTPP data. m0 (j)kS and 90 m0 (j)kU are weights calculated from the census micro data using all states outside of j’s metropolitan area m for the fraction of metropolitan area m workers in industry k in 1990 that are "skilled" and "unskilled" respectively. We count skilled workers as those with a college education or more and unskilled workers as those with any lesser amount of education. 05 ln Em 0 (j)kS indicates the log of 2005 skilled employment in industry k in all states excluding those of 8 We use 2000-2005 shocks rather than 2000-2010 shocks in order to improve …rst stage strength, as we discuss below. 16 00 metropolitan area m and ln Em 0 (j)kS is the analogous object for 2000. We build analogous versions of these variables for 1990-2000 using the same 1990 employment shares interacted with employment growth rates over this earlier period. Table A2 lists the industry categories that we use to construct these variables with the most and least rapid employment growth rates during the 1990-2000 and 2000-2005 periods. We also construct uni…ed Bartik shocks in which the weights are set to one. It is straightforward to microfound the use of such Bartik shocks such that they represent the national component of productivity or output price growth as follows.9 Suppose …rms use skilled labor, unskilled labor and nationally traded capital to produce. This generates the following (reduced form) tract-industry speci…c aggregate labor demand equation for skill group S, where pk S U is the output price, wjk is the skilled wage and wjk is the unskilled wage. ln LSjk = f S (ln S jk ; ln U S U jk ; ln pk ; ln wjk ; ln wjk ) Additionally, decompose productivities ln Sjk and ln U jk to have tract-speci…c, industry-speci…c S S S S and idiosyncratic components: ln jk = aj + bk + ujk . Our goal is to achieve identi…cation from variation in productivity or output demand shocks across industries represented by di¤erential trends in ln Sjk , ln U jk and ln pk . Aggregating across industries at the tract level, we have d ln LSj = X S Sjk d ln LSjk k = X k + S S Sjk [f1S dbSk + f2S dbU k + f3 d ln pk ] X S S Sjk [endogjk ]; k S where Sjk is the share of base-year skilled employment in industry k, subscripts on f indicate partial S S U derivatives and endogjk = f4S d ln wjk + f5S d ln wjk . P S S S The idea of Bartik instruments is to use only variation in d ln LSj from k Sjk [f1 dbk + f2S dbU k + S S S S U S f3 d ln pk ] for identi…cation. To achieve power, we need that [f1 dbk +f2 dbk +f3 d ln pk ] is correlated 05 00 with [ln Em ln Em 0 (j)kS 0 (j)kS ], calculated only using locations in states outside of the metro area of tract j. We can think of identi…cation as coming either from exogenous components of di¤erences S in initial industry shares Sjk across tracts (Goldsmith-Pinkham et al., 2018) or from random shocks to industry growth (Borusyak, Hull & Jaravel, 2018). While random industry-speci…c growth rates would obviate the need for concern about exogeneity of tract-level Bartik instruments, our observation that shocks are correlated across industries leads us to organize our empirical strategy in order to minimize potential concerns that base year 9 Also see Adao, Kolesar, and Morales (2018) for a similar treatment. 17 90 0 industry shares Xm (j)kS Emp90 jk may be correlated with unobserved labor supply shifters driving 90 Emp90 jk m0 (j)kS k local employment growth. For example, areas with a heavy manufacturing presence may have declining amenities due to industrial pollution and plant closures that shift both labor supply and labor demand inwards. In our main empirical work laid out in Section 6, we sidestep this problem by only using Bartik shocks outside tracts of residence for identi…cation and have robustness checks that explicitly control for predicted employment growth near the origin tract. We also present an analysis of pre-treatment trends to further alleviate concerns that unobservables driving outcomes of interest may be correlated with the 1990 industrial composition of employment in commuting destination tracts. Our purpose in this section is only to indicate the sources of Bartik-type variation that are available for identi…cation in our setting. In our implementation of this Bartik research design, we follow best practices suggested by Goldsmith-Pinkham, Sorkin & Swift (2018). First, we maintain the same base year shares for Bartik shocks over both the 1990-2000 and 2000-2010 periods. Second, we show robustness to a number of di¤erent control sets. Third, we demonstrate that the instruments do not predict pretreatment trends, conditional on these control sets. Anticipating our use of spatial aggregations of tract-level Bartik shocks for identi…cation in the main empirical work, we defer the third test to our main implementation in Sections 6 and 7. 4.2 Tract Level Empirics Since tract-level measures of employment by education are not available in 1990 and 2000, we impute employment by skill using the region-speci…c distribution of educational attainment by industry and the tract-speci…c industrial composition. These objects allow us to construct weights which represent the share of skilled and unskilled workers in each tract and year. We build these weights using education shares by industry for all PUMAs in the region of interest using data from the Census PUMS for the year in question. For 2010, we directly observe the fraction of jobs held in each tract by education in the LODES data. Table 2 shows tract-level regressions of employment growth rates for 2000-2010 (Panel A) or 1990-2000 (Panel B) of all, high skilled and low skilled employment on Bartik shocks BartikjS and BartikjU . We control for 2-2.5km wide CBD distance ring …xed e¤ects interacted with region …xed e¤ects, a quadratic in CBD distance and 10 and 20 year lags of tract demographic composition in all speci…cations.10 The results in columns 1 and 2 of Panel A show the e¤ects of the Bartik variables on total employment growth. These results show that high-skill Bartik shocks predict total 2000-2010 tract1 0 Accounting for the strong CBD distance-speci…c trends in employment growth requires our ‡exible controls for CBD distance. The lagged demographic controls are included to make this speci…cation comparable to those developed below in the main part of the empirical analysis. 18 level employment growth (column 1), and do so, even conditional on low-skill Bartik shocks. The results in columns 3 and 6 of Panel A show that the skill-speci…c shocks have impacts on skillspeci…c employment growth with the expected signs. The results in columns 4 and 7 show that this is true even conditional on the other shock, prefacing our ability to separate out skill-speci…c 2000-2010 labor demand shocks to residential neighborhoods. However, the results in columns 5 and 8 show that the identifying variation in these skill-speci…c shocks is quite di¤erent for the high and low skilled shocks. In particular, positive high-skill shocks and negative low-skill shocks provide most of the identifying variation, meaning we can only successfully isolate exogenous variation in tract-level skilled employment growth and unskilled employment declines for the 2000-2010 period. We note that analogous regressions using 2000-2010 shocks instead of 2000-2005 shocks have lower power due to the fact that there is little cross-industry shock variation for the 2005-2010 period. It is for this reason that we use the 2000-2005 period only to construct Bartik shocks. We use 2010 as the terminal year in which we measure employment by skill because this is the …rst year in which the LODES coverage is complete and in which we can observe the actual education composition of employment in each tract. Panel B of Table 2 shows the e¤ects of 1990-2000 Bartik shocks on contemporaneous employment growth. Here, the results in columns 4 and 7 are most informative about our power to separately predict high and low skilled employment growth with their respective shocks. The results in both columns show positive coe¢ cients on the high-skill shock and negative coe¢ cients on the lowskill shock. This means that we cannot separately predict both skilled and unskilled employment growth simultaneously with these two shocks. As such, our analysis for the 1990s necessarily focuses more on simply estimating the e¤ects of uni…ed labor demand shocks on neighborhood change and residents’ outcomes. Indeed, we show evidence from the PSID that neighborhood e¤ects are not separately identi…able from parents’wealth e¤ects in Section 7.2. Taken together, the results in Table 2 show that we can isolate exogenous variation in tractlevel skilled employment growth, while holding unskilled employment growth constant, in the 20002010 period. Since our primary goal is to isolate exogenous variation in skilled worker residential populations across tracts, we use tract-speci…c skilled Bartik shocks as our main source of identifying variation. The next section shows how we aggregate tract-level labor demand shocks to form the treatment variables hitting each residential tract used in our main empirical analysis. 5 Conceptual Framework In this section, we lay out the application of the urban economic geography model developed in Ahlfeldt et al. (2015) as adapted by Tsivanidis (2018) to generate a measure of "market access" that is readily measured with our data, is conceptually appealing, and has a convenient theoretical interpretation. In addition to facilitating clear thinking about various threats to identi…cation, the 19 model delivers treatment measures that incorporate sensible spatial aggregates of labor demand shocks that are relevant to residential locations. The model is used primarily as a vehicle to deliver theoretically grounded treatment variables that capture shifts in labor demand conditions that are oriented toward residential locations and can be constructed from identifying variation across work locations. We conceptualize each region as starting o¤ in long-run equilibrium in 1990. In this equilibrium, the labor market clears in each work location j and the housing market clears in each residential location i. We …rst develop the simpler case in which there is only one worker type and then show how the environment extends to accommodate both skilled and unskilled workers. 5.1 Preferences & Worker Productivity Starting from standard Cobb-Douglas preferences, we can write the indirect utility of individual ! living in tract i commuting to work in tract j and working in industry k as Vijk! = vi! Bi zijk! wjk Q1i e ; ij where Bi is a local consumer amenity, wjk is the price of a unit of skill in jk, Qi is the price of ij a unit of space in i, 1 e is the fraction of time spent commuting for those living in i and working in j and 1 is the housing expenditure share. One component of the local amenity Bi may be endogenous and depend on neighborhood demographic composition. zijk! is a worker-commute-industry speci…c productivity shock drawn from the Frechet distribution: " Fz (zijk! ) = e zijk! ; " > 1. vi! is a preference or amenity shock for living in i, also distributed Frechet: Fv (vi! ) = e vi! ; > 1: Workers …rst see the preference (or amenity) shock for each potential residential location and choose their places of residence to maximize expected utility, anticipating the distribution of wages net of commuting costs associated with each residential location. The productivity shocks are then revealed and individuals choose the highest utility work location-industry combination. If = "; this formulation reduces to the simpler version of the model in which there are only productivity shocks and no preference shocks, and workers choose commutes to maximize indirect utility in one step.11 1 1 Reformulation of the model to have individuals …rst choose a work location, based on productivity shocks indexed to j only, and then a commute, with amenity shocks speci…c to either i or ij, generates constant elasticity relationships that are isomorphic to those derived below. 20 5.2 Residential Population Given this setup, the fraction of residents in i who commute to j is ijji P k [wjk e =P P k j 0 [wj 0 k e ij " ] ij " ] P k [wjk e RM Ai ij " ] (8) P where RM Ai k RM Aik . RM Ai is our central Resident Market Access variable that is a summary of the strength of job opportunities in and around residential tract i. Below we show how we calculate RM Ai using information on employment and residential population in each tract. Equation (8) shows that the commuting probability to j is increasing in the wage and decreasing in the commuting cost. The expected income net of commuting cost from living in tract i is y i = (1 1 1 )(RM Ai ) " ; " (9) where ( ) is the gamma function and comes from taking the expectation over a Frechet distributed random variable. Note that this object is less than the average wage in tract i because the wage does not explicitly include commuting costs. Even though it is not observed directly, y i is a useful object to de…ne as it is constant elasticity in RM Ai , which we can measure with our data. Anticipating full income y i and housing cost Qi , preference shocks trace out the residential population supply to tract i. The probability that an individual’s utility is maximized by living in tract i is: 1 (10) Bi Qi 1 (RM Ai ) " i = Intuitively, this object is increasing in amenities and labor market opportunities but declining in the housing price.12 5.3 Commuting Gravity and Labor Supply In the data, we observe that commute lengths di¤er markedly across regions. New York has the longest commutes, with an average commute time of 35 minutes in 1990. Bryan-College Station, Texas has the shortest at just 13 minutes. The model can rationalize this discrepancy with di¤erent Frechet parameters " for each region. Regions in which people are willing to commute longer have more dispersion in their productivity draws (lower values of "). Indeed, a classic explanation for agglomeration economies is that larger cities like New York may exhibit more division of labor and heterogeneity in job types than smaller places like Bryan-College Station (Tian, 2018). Allowing for variation in " across regions will be important for our empirical implementation, as tract employment 1 2 The constant is 1= Xh Qi0 i0 1 i 1 Bi0 [RM Ai0 ] " . 21 growth impacts market access in areas accessible by longer commute times (meaning broader areas) in regions in which " is lower. Combining Equations (8) and (10) allows for the recovery of equilibrium commuting ‡ows ij = ijji i . These satisfy the following gravity relationship with commuting times. ln( ij ) = = ln + ln i + " X [wjk ] k ( ") j " # + ln h Bi Qi 1 1 (RM Ai ) " i " ij (11) ij That is, we can identify " with commuting ‡ow gravity regressions that include origin and destination …xed e¤ects. We implement the regressions described by Equation (11) by region in 1990 and 2000, weighting by ‡ow (the number of commuters). The resulting estimates of " are depicted in Figure A2. It shows values as low as about 0.02 in large cities including New York and Los Angeles, but also some smaller cities including Tulsa, OK and Mobile, AL. Large values up to about 0.12 are observed in some small cities, including Fargo, ND and Eau Claire, WI. A reasonable calibration of is 0.005 to 0.01, implying that a 10 minute commute takes 5-10 percent of the worker’s time endowment for working plus commuting. Comparing Panels A and B of Figure A2, the distribution of c" across regions is pretty stable between 1990 and 2000, though smaller cities are more likely to move around the distribution than are large cities. The correlation between the 1990 and 2000 region-speci…c estimates of c" is 0.61. X The labor supply to tract j can be derived by aggregating over commuting ‡ows Lj = ij . i Lj = " X k " wjk # Xh " e ij Bi Qi 1 1 RM Ai" i i (12) hP i " This object is increasing in the wage opportunities k wjk available in tract j and the residential population in locations that are more accessible to tract j. Following Tsivanidis (2018) and Donaldson and Hornbeck (2016), we de…ne the access from each work location j to residential locations of the labor pool as "Firm Market Access" F M Aj Xh e " ij i Bi Qi 1 RM Ai" 1 i : (13) Intuitively, this object is increasing in nearby residential amenities Bi but decreasing in nearby housing costs Qi and commute times to residential locations ij . From the perspective of …rms 1 in location j, the object e " ij RM Ai" re‡ects two forces. On the one hand, higher RM Ai in nearby residential locations is a positive labor supply shifter that tends to increase F M Aj . On the 22 other hand, higher RM Ai in such locations tends to re‡ect more competition for workers, thereby reducing F M Ai . If individuals’elasticity of substitution between neighborhoods in the residential demand system, , is higher than the labor supply elasticity to work tracts, ", then RMA’s positive labor supply shifter e¤ect outweighs its competition e¤ect, thereby increasing FMA.13 5.4 Measurement of Market Access We combine the expressions above for RM Ai and F M Aj into a system of recursive equations that can be solved given data on the number of jobs and workers in each tract, tract pair commute times and the parameter cluster " . Substitution of Equation (10) into Equation (13) yields an expression for FMA that depends only on RMA in each commuting origin, cluster " and the hP the parameter i " residential population of each commuting origin. Using Lj = w F M Aj from Equation (12) P P k jk " " ij and substituting into the de…nition of RM Ai = j e k [wjk ] delivers the expression for RM Ai below. The resulting system of equations is F M Aj = Xe i RM Ai = Xe j " ij i RM Ai " Lj . F M Aj (14) ij (15) This system of equations captures the interplay between employment and residential commuting linkages in a metropolitan region. The numerator of Equation (15) re‡ects that greater employment accessibility must raise available wages net of commuting cost to residential tract i. The denominator of Equation (15), F M Aj , captures the competition e¤ect - that wages become depressed if there are more competing potential workers living nearby. A nice feature of these "market access" objects is that their speci…cation only depends on indi¤erence of people across work and residential locations given some equilibrium wage vector wjk . We do not need to specify a structure of labor demand or …rm production in order for the empirically observable object RM Ai to capture labor demand conditions in tract i. Using data on employment Lj , residential population i and commute times ij plus estimates of " for each region from the gravity equation (Equation 11), we calculate F M Aj and RM Ai for each tract and year in our data. We also calculate reduced form analogs to market access "Resident P Market Potential" as RM Pi = j e " ij Lj . In the data, RM Pi and RM Ai have a correlation coe¢ cient of about 0.95. 1 3 Attempts to estimate in our data yield estimates of about 4, indicating that neighborhoods are highly substi" tutible in the residential demand system. 23 5.5 Model Closure and Equilibrium To close the model, we introduce the constant-elasticity housing supply function for each tract. His = (16) i Qi We allow each tract to have its own housing productivity but constrain the housing supply elasticity to be the same in all locations.14 Our preference speci…cation delivers the following Cobb-Douglas tract housing demand function. Hid = (1 ) Qi (1 1 1 )(RM Ai ) " " (17) i For simplicity, we assume that …rms do not use space in production. Equating housing supply (Equation 16) with housing demand (Equation 17) and substituting in for population, we derive the equilibrium relationship between housing price and RM Ai . ln Qi 1 + 1 + (1 = + ) + 1 + (1 ln[(1 ) ln Bi 1 1+ ) ]+ " " 1 ln + 1 + (1 ) 1 + 1 + (1 ) (1 ln(RM Ai ) ) (18) i This expression looks like a regression equation, with the error term composed of a linear combination of local amenity ln Bi and housing productivity ln i . Substituting back into the population supply condition (10), we have15 ln i ( + 1) + (1 ) ln (RM Ai ) " ( + 1) + (1 ) ( 1) ( 1) +( + ) ln Bi + 1 + (1 ) + 1 + (1 = K+ ) ln i (19) Finally, we calculate the average wage earned by workers in each residential tract. Note that average income net of commuting cost is a constant elasticity function of RMA, but average wage is not. However, if commuting costs are small then the two are close. 1 wi = [RM Ai ] " (1 1 XX ) " j k S ijkji e ij 1 > [RM Ai ] " (1 1 ) " In the context of the simple version of the model with one type, the empirical challenge is that RM Ai incorporates elements of error terms in Equations (18) and (19). In the following section, we 1 4 As 15 K RM Ai does not depend on housing supply elasticity, this assumption does not materially impact our analysis. ( 1) ( 1) = (1 + +1+ ) ln + +1+ ln[(1 ) (1 1" )] 24 lay out our strategy for using tract-speci…c Bartik shocks to isolate variation in ln RM Ai that is orthogonal to trends in local amenities and housing productivity that show up in these error terms. 5.6 Extension to Multiple Skill Groups Incorporating multiple skill types in the model is straightforward. Given observations about residents and workers by skill in each location, we retain the same de…nitions of RM A and F M A as in Equations (15) and (14) with the addition of superscripts S for skilled or U for unskilled. Note that the introduction of two versions of each of these objects accommodates the possibility that U S unskilled and skilled wages wjk and wjk may both depend on the number of unskilled and skilled workers in tract j and industry k. Measurement of equilibrium RM ASi and RM AU i does not require taking a stand on how low and high skilled workers interact in production. Because the two types of workers compete for housing in each residential location, the housing market clearing condition delivers a housing cost that depends on both RM ASi and RM AU i . This feeds through to mean that both objects also predict the equilibrium unskilled and skilled S populations of each tract U i and i . In particular, the new equilibrium conditions are ln Qi = q+ ln U i = Pu + ln S i = Ps + 1 + 1 + (1 " " ) i + ! q (Bi ; i) ( 1) + 1 + (1 ( 1) + 1 + (1 ln RM AU + i ln RM ASi + ) i + ln BiU ) i + ln BiS ( 1) + 1 + (1 ( 1) + 1 + (1 ) ) ln ln i (20) i; (21) in which the component of total housing demand in tract i given by i equally shifts the shares of workers of each type that choose to live in i through its impact on housing cost.16 A few observations can be made from these expressions for equilibrium populations that are S of particular relevance for our empirical work. First, the college fraction S +i U is increasing in i i ln RM ASi holding ln RM AU constant. This justi…es our empirical strategy of using exogenous i shocks to ln RM ASi while holding ln RM AU i constant to deliver variation in residential locations. While analytical expressions for ln( S i S i U i S i S+ U i i across ) are simpler, we prefer to focus on as an outcome, as it is both commonly used in the gentri…cation literature (e.g. Brummet & Reed, 2019) and is empirically better behaved in tracts with few residents of one skill group. Since S S U S our empirical work is done in di¤erences and d S +i U = S +i U S +i U d ln( Ui ), the two measures i i i i i i i of neighborhood change are closely related. S Second, ln( Ui ) directly depends on unobserved tract consumer amenities BiU and BiS , which i S will end up in error terms in the empirical work. Moreover, RM AU i and RM Ai themselves are S+ U i i 16 i = ln h BiU U (RM AU i ) 1+ " + BiS S (RM AS i ) 1+ " 25 i functions of these amenities not only in tract i, but in all other tracts as well. This is because RM Ai depends on wages in all commuting destinations j, which in turn depend on populations in all commuting origins i0 (including i), which themselves depend on local amenities Bi0 . The existence of neighborhood e¤ects would mean that such local amenities are a function of the tract residential S composition S +i U . Therefore, estimation of treatment e¤ects of RM ASi on tract attributes (holdi i U S ing RM AU i constant) requires using variation in RM Ai that is independent of variation in RM Ai and shocks to local amenities. We use these observations in the next section, in which we discuss how we use tract-level Bartik shocks to isolate exogenous variation in RM Ai that is oriented toward skilled labor. Third, we think of the structural parameters and " as being heterogeneous across regions and as potentially heterogeneous across tracts (Baum-Snow & Han, 2018). Indeed, in Section 5.3 above we showed direct evidence that " is heterogeneous across regions. However, in the empirical work in the next two sections we do not have su¢ cient statistical power to estimate parameters separately by region. That is, our empirical setting constrains us to estimating average coe¢ cients which do not have straightforward structural interpretations.17 Panel A of Table 1 provides summary statistics about log di¤erences in skilled and unskilled RMA measures over our two study periods. In both study periods RM ASi grows more rapidly on average than does RM AU i , though the two objects have a similar amount of dispersion. Importantly, most of the dispersion in market access growth comes from comparisons between rather than within commuting regions. The average within region standard deviation of the growth rate of skilled RMA, weighting all tracts in our primary sample equally, is 0.09 for 1990-2000 and 0.02 for 2000-2010. In much of the empirical work, we standardize using the average within region standard deviation across our full sample of census tracts so as to be able to make comparisons of impacts of treatment across tracts within metropolitan regions.18 Comparing the mean growth rates of tract market access to tract-level employment growth shows why market access is a useful summary measure of accessible labor market opportunities. In the 1990s, the mean tract-level employment growth rate was 0.17 - much greater than the mean growth rate of market access because of employment booms in areas with little 1990 employment. With the 2007-2009 …nancial crisis, tracts with only a small amount of 2000 employment disproportionately lost jobs, leading to a mean 2000-2010 tract-level employment growth rate of -0.21. Using the market access aggregation of these tract-level changes captures variation in changes in job accessibility much more smoothly than the simpler tract-level employment growth measure. 1 7 Moreover, in most of our empirical work, we express RMA in terms of standard deviations, which allows us to make clearer connections with the neighborhood e¤ects literature. 1 8 Allen, Arkolakis & Li (2016) show that a unique equilibrium exists in this type of model as long as agglomerative forces are not too strong, and that there is a unique mapping from the joint spatial distribution of population and employment to tract amenities and productivities. In practice, solving the just-identi…ed system of 2J equations in 2J unknowns for RMA and FMA per region is relatively fast. The largest region, New York City, takes only a few minutes. 26 The left and middle panels of Figure 3 present heat maps of log skilled RMA in 2000 and 2010 for Los Angeles and Orange Counties, the bulk of the Los Angeles region in our data. Here one can see the smoothness of this measure across space. Adjacent tracts are most often in the same quartile (color shade) of each distribution. However, as seen in the right panel, the 2000-2010 change in log skilled RMA for this region is quite heterogeneous across space, with lots of variation across quartiles of its distribution between adjacent census tracts. Our empirical challenge, described in the following section, is to …nd instruments and controls that allow us to isolate plausibly exogenous variation in the change in log skilled RMA. 6 Empirical Implementation The main goal of our empirical work is to use variation in changes in employment opportunities across census tracts in each metropolitan region to recover long-run e¤ects on children. The neighborhood, parent and child outcome equations (1 - 5) laid out in Section 2 are the starting point for our empirical work. Due to various limitations to statistical power hinted at in our tract-level empirical analysis in Section 3, we are constrained from implementing these estimation equations exactly. We discuss these issues in this section along with various other identi…cation challenges to settle on equations that are similar to Equations (1) - (5) but can be feasibly taken to the data. The model is instructive both in showing how to construct a reasonable measure of labor market opportunities relevant to each location, RM Ai , and in highlighting various identi…cation challenges. We conceptualize each region as starting o¤ in long-run equilibrium with tract log populations by skill described by Equations (20) and (21). Each tract experiences shocks to local consumer amenities and wages in commuting destinations, which themselves are generated by national sector-speci…c productivity shocks. These wage shocks indirectly cause additional changes in local amenities BiU and BiS as tract demographic compositions change. The empirical challenge is …rst to isolate variation in the tract demographic change due to the arrival of new residents that come only because of productivity shocks to industries in nearby commuting destinations and then to estimate the knock-on e¤ects for incumbent resident parents and their children. Through the lens of the model, the key identi…cation challenge is that BiU and BiS may also change for other reasons that are correlated with but not caused by tract demographic change. 6.1 Instruments The model illustrates that ln RM ASi and ln RM AU i depend on local trends in consumer amenities and housing productivity, both of which are labor and population supply shifters. That is, we are concerned that growing employment and RM A occur in some areas because nearby residential neighborhoods are becoming nicer places to live, not vice-versa. These amenity changes may have 27 direct e¤ects on incumbent households and children that have nothing to do with neighborhood e¤ects. More generally, we are concerned that unobserved neighborhood attributes (including demographic characteristics) that are correlated with nearby employment growth may be driving child outcomes rather than demographic change that is caused by this employment growth. To deal with these issues, we require identifying variation that is uncorrelated with local amenity shocks and other unobserved drivers of tract demographic composition that in‡uence child outcomes. Our primary workhorse instruments are 2000-2005 or 1990-2000 di¤erenced counterfactual versions of RMA, which aggregate tract-level employment predicted with Bartik shocks. We construct the 2000 and 2005 levels of this counterfactual RMA as follows. In each tract, we assume that the 2000 and 2005 employment compositions maintain 1990 shares by industry and skill but get scaled up by the national growth rates for workers in each industry by skill. As with the tract-speci…c versions described in Equations (6) and (7), we apply industry-speci…c employment weights by skill from outside the region to construct growth rates of employment by skill. We omit tract i when constructing the instrument in order to reduce the possibility that 1990 industry composition might predict subsequent tract amenity changes. This logic delivers the following expression for counterfactual skilled RMA in 2005: S05 gAi RM = XX e k " 90 ij 05 90 90 Em0 (j)kS m0 (j)kS Empjk E 900 m (j)kS S05 Fg M Aj j6=i We maintain 1990 commute times and employment shares by industry and skill to build counterfactual RMA measures for all years. However, we update region-speci…c decay parameters " to be calculated with the 2000 data for instruments that apply to post-2000 growth in RMA. We construct two counterfactual versions of skilled and unskilled RMA for 2000, one that uses 1990 estimates of " to be di¤erenced with the 1990 data, and one that uses the 2000 estimates of " X to be di¤erenced with the 2005 counterfactual RMA. Since RM Ai = RM Aik , aggregation over k industries is fully consistent with the model. Since RMA is codetermined with FMA, we must also specify a counterfactual FMA, the commuting time discounted sum of population in commuting origins accessible from tract j, divided by RMA in these origins. To construct counterfactual FMA, we assume that the 1990 residential population in each tract changes proportionately, so that the spatial distribution of residences does not change in the counterfactual 2000 or 2005 environment relative to 1990. This rescaling of X X gAi without tract population achieves the required market clearing condition Fg M Aj = RM j i imposing any di¤erential labor supply shocks across tracts. The resulting year 2005 measure of 28 counterfactual FMA for skilled workers is thus: S05 Fg M Aj = Xe " 90 ij S90 i E 050 m (j)kS 90 Emp90 jk E 90 m0 (j)kS m0 (j)kS P P 90 90 j k m0 (j)kS Empj P P j k S05 gAi RM i6=j With this speci…cation of counterfactual FMA, counterfactual RMA for any year after 1990 only incorporates changes in skill-speci…c labor demand conditions.19 We use the di¤erence in the log of 2005 counterfactual skilled RMA and the log of 2000 counterS05 S00 gAi gAi ) as our main instrument for the 2000-2010 change in factual skilled RMA (ln RM ln RM log skilled RMA (ln RM AS10 ln RM AS00 ) because the di¤erence in the log of 2010 counterfactual i i S10 S00 gAi gAi ) does not skilled RMA and the log of 2000 counterfactual skilled RMA (ln RM ln RM have su¢ cient …rst stage power. This is consistent with our tract-level empirical results in Section 4.2. The left panel of Figure 4 shows a heat map of our main instrument for the 2000-2010 analysis S05 S00 gAi gAi , for Los Angeles and Orange counties. Tracts are shaded by below, ln RM ln RM quartile of this variable’s distribution within the Los Angeles study region. The right panel of Figure 4 shows the same object after being residualized from a large set of control variables that are used in our primary estimation speci…cations. CBD distance rings of 10 km and 20 km are also indicated in the right panel. Both panels in Figure 4 show considerable variation across quartiles of the distribution conditional on CBD distance. 6.2 Selection of Empirical Model Speci…cations We develop empirical speci…cations with the goal of estimating parameters in the tract equation (Equation 1) that describe treatment e¤ects of skilled RMA growth ( ln RM ASi ) on changes in tract attributes (particularly neighborhood quality and fraction college graduate), while holding unskilled RMA growth ( ln RM AU i ) constant. We use the tract-level demographic data to select our empirical speci…cations because this is our only data set that goes back several decades in time and fully covers the regions in our sample. This allows us to determine the set of control variables required to condition out statistically signi…cant relationships between the instruments and pretreatment trends in the dependent variables. After using the tract demographic data set to select our empirical speci…cation, we specify analogous estimation equations which facilitate the recovery of treatment e¤ects of skilled RMA growth ( ln RM ASi ) on parent and child outcomes. To get a sense of the main potential biases when attempting to estimate the e¤ects of labor 1 9 An alternative approach would be to construct a fully simulated instrument in which Q , S , U , RM A and i i i i F M Aj are all solved out given a Bartik-based counterfactual employment Lj in each tract. However, doing so would require knowledge of all of the structural parameters of the model. We have shown evidence that these parameters likely di¤er across regions and are di¢ cult to identify empirically as a result. 29 demand shocks on tract outcomes, we …rst examine the dynamics of neighborhood change descriptively. Table 3 reports OLS estimates of a1 from regressions of the form t nim = a0 + a1 t 1 nim + Xim + umr + im where nim is a tract-level outcome for tract i in region m. That is, it shows the extent to which changes in neighborhood quality, fraction college or income growth rates are serially correlated conditional on various control sets. Here, Xim is a vector of base controls which include a quadratic in CBD distance, log 1990 tract employment and 10 and 20 year lags of a house price index, rent index, log population, log family income, share African American, share white, share college graduate, and share without a high school degree, umr are "region-ring" …xed e¤ects that fully interact region …xed e¤ects with indicators for CBD distance rings of 0-2, 2-4, 4-6, 6-8, 8-10, 1012.5, 12.5-15, 15-17.5 and 17.5-20 km, and im is an error term.20 The descriptive results on the serial correlations of these decadal changes, reported in Table 3 columns 1, 2, 5, 6, 9 and 10, are striking. Conditional on region-ring …xed e¤ects and various combinations of additional controls, changes in our neighborhood quality index, fraction college and log average household income are all negatively serially correlated. Indeed, this negative serial correlation grows as more pre-determined controls are included in the regressions. The remaining two columns for each outcome in Table 3 (columns 3, 4, 7, 8, 11 and 12) show how our main endogenous treatment variable ln RM ASi for the 2000-2010 period relates to the lagged change and initial levels of the dependent variables. For example, the dependent variable in column 3 is the change in the share of residents with a college degree or higher from 1990 to 2000 and the dependent variable in column 4 is the share of residents with a college degree or higher in 2000. Here we see that growth in skilled RMA is positively related to prior neighborhood improvements and initial levels of neighborhood quality. The results for the 1990-2000 period are presented in Appendix Table A3. The 1990-2000 period shows similar mean reversion as the 2000-2010 period. However, the conditional correlations between skilled RMA growth 1990-2000 and the 1980-1990 changes in the dependent variables or 1990 levels of the dependent variables are negative rather than positive, as they are in the later study period. Given the tendency for the growth in neighborhood quality to revert to the mean, one important function of our IV empirical strategy is to help control for the associated di¤erential pre-treatment trends. Figure 5 shows two diagrams that lay out the fundamental identi…cation challenge and indicate the expected directions of OLS bias. On the vertical axis is tract fraction college and on the horizontal axis is time. The lines show two tracts with di¤erent evolutions of fraction college and 2 0 In each year, the tract-level house price (rent) index is formed from the residuals of a regression of log mean owner occupied home value (log mean gross rent) on housing unit structure characteristics (number of units in building, number of bedrooms in unit, age of building) of the tract and region …xed e¤ects. 30 how they would evolve di¤erently given two scenarios each for 2000-2010 RMA growth. The diagram on the left side shows the ideal 2000-2010 experiment, in which two tracts with identical levels and histories receive di¤erent rates of RMA growth. The red tract gets a positive RMA shock of one standard deviation and the blue tract gets no RMA shock. In an ideal experimental world, the blue tract is a valid counterfactual for the red tract that gets treated. Therefore, we can identi…y the treatment e¤ect of the one standard deviation RMA shock by comparing the 2000-2010 growth in outcome in the red to that in the blue tract. The 1990-2000 pre-trends for these two tracts are identical, meaning they would have the same mean-reverting trend absent a shock to RMA. The right panel in Figure 5 is closer to our empirical setting for the 2000-2010 period. In this diagram, we see the red tract with the more rapid 1990-2000 growth and higher 2000 level receiving the positive shock relative to the blue tract. We would like to use the blue tract as a counterfactual for the red tract absent a shock, but there is a selection problem. As a result, the simple di¤erence in di¤erence type estimator A-B underestimates the true treatment e¤ect in this setting. A triple di¤erence type estimator (A-A’) - (B-B’) would exacerbate this problem. Going from OLS to IV makes groups of tracts that receive di¤erent employment growth treatments more comparable, as does including demographic controls within IV. Since this is the nature of selection in our 2000-2010 study period, OLS understates the true causal e¤ects of RMA growth on neighborhood quality. The main role of the instruments is to quasi-randomize the distribution of unobserved neighborhood characteristics so that the treatment is unrelated to pre-treatment trends in neighborhood characteristics. Without such randomization, initial neighborhood attributes rather than changes in neighborhood composition or the labor market environment may drive child outcomes. Moreover, the instruments are useful because they allow us to isolate variation in employment opportunities for college graduates while holding employment opportunities for those without a college degree …xed. Since all of our instruments are de…ned using variation in 1990 employment composition near tract i, sorting on the industry composition of nearby employment is our main identi…cation concern. While excluding tract i from the calculation of the instruments helps with this problem, employment levels and compositions by industry are likely to be spatially correlated. Our …rst approach to address this issue is to explicitly control for tract i 1990 employment and its skill-speci…c 2000-2005 Bartik shocks in all regressions. In some speci…cations, we also control for Bartik shocks hitting 0-10 minute and 10-20 minute commuting time rings from tract i, forcing identi…cation to come from Bartik-type demand shocks hitting commuting destinations that are greater than 20 minute commutes away. Since typical commutes are longer in many cities, we can still achieve identi…cation even after excluding variation from locations within 20 minutes when constructing the treatment variables. However, reduced power and larger standard errors arise in this case for some outcomes. Second, we are concerned that the 1990 employment composition near each tract may respond to the local real estate market and/or demographic trends, which are related to unobserved demo- 31 graphic attributes or local amenities that may enter into child investments. As such, we control for 10 and 20 year lags of tract demographic composition and housing costs. We do not control for base year demographic characteristics from the Census because some of our analysis uses these objects as part of the outcome variable (the base year in a time di¤erence). Third, we must account for secular trends in neighborhood demographics as a function of CBD distance. With the central areas of many cities rebounding after 2000 due to improved local amenities (Baum-Snow & Hartley, 2018; Couture & Handbury, 2017), we wish to compare neighborhoods at similar CBD distances. To that end, we control for a quadratic in CBD distance in all speci…cations. In addition, in most speci…cations we control for CBD distance ring …xed e¤ects interacted with region …xed e¤ects, capturing potential di¤ering region-speci…c spatial demographic trends. The heat map in the right panel of Figure 4 highlights the sources of identifying variation we use for the Los Angeles region in the 2000-2005 period. As is evident in the map, there is considerable identifying variation available conditional on most CBD distances, with all four quartiles well-represented. We select our primary speci…cation for the main empirical work by evaluating which conditioning variables are required to render our instruments uncorrelated with observed pre-treatment trends in tract characteristics. To this end, Table 4 presents IV regressions of pre-treatment trends in fraction college and our neighborhood quality index on ln RM ASi for 2000-2010 instrumented with S gAi for the 2000-2005 period, including various sets of controls. We study the sensitivity ln RM U gAi as a control. We also study the sensitivity of of our results to including the instrument ln RM prediction of pre-trends to the inclusion of controls for skill-speci…c Bartik shocks aggregated across tracts within 0-10 and 10-20 minutes commuting time from each origin tract (Bartik ring controls). Including the two Bartik ring controls renders the relationship between the pre-treatment trends and the instruments to be economically and statistically insigni…cant in all cases studied except for U gAi control. However, given the descriptive evidence in fraction college when we include the ln RM Table 3, the logic of Figure 5 suggests that at least for the 2000-2010 study period, the existence of pre-trends will cause us to estimate lower bounds on true causal e¤ects. Consistent with this logic, in our main empirical work below we show that moving from OLS to IV and including additional controls in the IV both increase the estimated treatment e¤ects. Table A4 presents analogous results for the 1990-2000 study period, in which ln RM Ai for S gA . For this earlier period, we do not have su¢ cient 1990-2000 is instrumented with ln RM i U gA . Instead, we examine sensitivity of the identifying variation to be able to control for ln RM i inclusion of Bartik controls for 0-10 and 10-20 minute commute time rings. Here we see that including the Bartik ring controls eliminates di¤erential pre-treatment trends. 32 6.3 First Stage Results Table 5 presents the …rst stage results. Both instruments and endogenous variables are expressed in within-region standard deviation units for the average tract in our census sample. Anticipating the heterogeneity in treatment e¤ects as a function of CBD distance that we show below, we break out coe¢ cients on the instruments by 0-10 km and 10-20 km CBD distance bands. Table 5 Columns 1-4 show that 2000-2005 counterfactual skilled RMA growth rates signi…cantly predict actual 2000-2010 skilled RMA growth rates both near and far from CBDs in the full census U gAi and tract sample. Since the unskilled and skilled counterfactual RMA growth rates ( ln RM S gAi ) are positively correlated, the coe¢ cients on the skilled counterfactual RMA growth ln RM S U gAi ) rise when the unskilled counterfactual RMA growth rate ( ln RM gAi ) interrates ( ln RM acted with 0-10 km and 10-20 km CBD distance bands are included as controls. With all controls included (column 4), a one standard deviation increase in the skilled counterfactual RMA growth S gAi ) within 10 km of a CBD predicts a statistically signi…cant 0.09 standard deviation rate ( ln RM increase in the actual skilled RMA growth rate ( ln RM ASi ). In the suburbs, this coe¢ cient is more than twice as large at 0.23. In the right block of Table 5, we show analogous results for the CCP sample. Consistent with the analysis to follow, we impose some adjustments to handle the fact that the unit of analysis is now a person rather than a census tract. We inversely weight observations by the number of regions in which the tract appears and cluster standard errors on census tract. Without controls for unskilled RMA shocks in Columns 5 and 6, these …rst stage estimates exhibit similar patterns as those using the full census tract data set in Columns 1 and 2, though with larger standard errors. However, when such controls are included we lose …rst stage power, particularly in the inner CBD distance ring (columns 7-8). In the 1990-2000 study period, we are further constrained by the lack of separate identifying variation in skilled versus unskilled RMA shocks. Moreover, as the …rst stage coe¢ cients between the two CBD distance bands are much more similar and stable across speci…cations, we pool them to maximize statistical power. Without the Bartik ring controls, the …rst stage coe¢ cients are about 0.07 and with these controls they are about 0.04. These results are reported in Table A5. 6.4 Main Estimating Equations Here we re…ne the aspirational estimation equations (1)-(5) speci…ed in Section 2 to accommodate the constraints on data, power and identi…cation laid out in this section. In Section 2 we identi…ed both high and low skilled RMA shocks as treatment variables of interest. Constraints on power and available identifying variation in both the 2000-2010 and 1990-2000 study periods lead us to focus on recovering treatment e¤ects of t ln RM ASi on neighborhood, household and and individual outcomes. We consider each study period in turn. 33 U S gA and t ln RM gA do provide In the 2000-2010 study period, our two instruments t ln RM i i separate identifying variation. However, as seen in Tables 4 and 5, we lose precision and introduce U gA . To get around this precision problem, we pre-trends when including controls for t ln RM i estimate IV regressions of the form t ni! = n 0 + nS c t ln RM ASic + nS s zi!; = p 0 + pS c t ln RM ASic + pS s t = c 0 + cS c S t ln RM Aic cS s S t ln RM Ais yi! + t n ln RM ASis + Xi! ln RM ASis + Xi! + p c c Xi! S + eni! + epi! + eci! (22) (23) (24) gAi enters as the instrument for t ln RM AS . Subscripts c and s represent 0 in which t ln RM i to 10 km from the CBD and 10 to 20 km from the CBD, respectively. The tract equation (22) shows objects indexed by child !, but is also run for all neighborhoods using census data only. We U gAi as a control in census regressions, not in CCP regressions. only have power to include t ln RM U gAi constant, below we con…rm directly that b pS = b pS Because we cannot hold t ln RM s = 0, c obviating the need to control for the unskilled RMA growth shocks. In our initial treatment of this hierarchical empirical setup in Section 2, we indexed the second and third equations by the skill group of the parents. We do report coe¢ cients for di¤erent parent skill groups below but also show uni…ed results given our evidence that the shocks have no e¤ect on parental …nanical health (b pS = 0) for both skilled and unskilled parents. There are two main considerations that shape our choices of control variables Xi! . First, we choose our control set to minimize any pre-treatment trends in the data. This justi…es the inclusion of region-ring …xed e¤ects, 1990 log tract i employment, skilled and unskilled Bartik shocks for tract i (used as instruments in Table 2) and 10 and 20 year lags of a host of tract demographic variables and housing costs. We control for Bartik shocks aggregated to 0-10 minute and 10-20 minute commute time rings (henceforth, "Bartik ring controls") in robustness checks. Consistent with our discussion of selection biases in the context of Figure 5, all neighborhood e¤ects estimates grow larger with the inclusion of these controls. Second, we endeavor to control for variables that reduce the variances in error terms, thereby reducing standard errors of our coe¢ cient estimates. As such, we additionally control for parents’credit score and an indicator of whether they had any home loans in 2000. For the 1990-2000 study period, we are constrained to estimating a set of equations that include only one RMA variable. We use the following speci…cations. + v ni! (25) p + v pi! (26) c v ci! (27) t ni! = bn0 + bn t ln RM Ai + Xi! n t zi! = bp0 + bp t ln RM Ai + Xi! yi! bc0 ln RM Ai + Xi! = +b c t 34 + S gA . Note that We estimate these equations by IV, instrumenting for t ln RM Ai with t ln RM i rather than using long-run outcomes in the parent equation, we instead use changes in outcomes. This choice comes because the structure of the PSID data does not allow us to measure parent outcomes in a way that parallels the timing structure of the child outcomes. For this earlier period, our set of controls in Xi! are analogous to those in the later period, except that instead of parents’ credit score and home loan information for 2000 we include a rich set of controls for 1990 parent and household characteristics. 7 Results In this section, we …rst examine how shocks to RM ASi causally a¤ect neighborhood demographics in the 2000-2014 period using the census data, corresponding to estimates of nS in Equation (22). Next, we report estimates of pS and cS from Equations (23) and (24) using the CCP data. We combine these estimates to recover estimates of neighborhood e¤ects. Finally, we perform the analogous analysis for the 1990-2015 period using the census and PSID data instead. 7.1 7.1.1 2000-2017 Period Neighborhoods Table 6 reports estimates from Equation (22). The top block shows OLS results, with four speci…cations for each outcome, and the bottom block shows IV results. The treatment variable, the decadal change in skilled RMA growth (ln RM ASi ), is scaled by its average within-region standard deviation in our sample. The dependent variables are measured as 2000-2007 changes so as to focus on the period in which sampled children are still mostly living with their parents.21 Our estimates indicate that a one standard deviation increase in skilled RMA growth increases the neighborhood fraction college by 1.6-2.6 percentage points (23 to 38 percent of one withinregion standard deviation) for the 2000-2007 period, depending on the speci…cation. Once controls for aggregate skill-speci…c Bartik shocks for 0-10 minute and 10-20 minute commutes (Bartik ring controls) are included, the estimates are not a¤ected by conditioning on the growth in counterfactual U gAi ). The estimated e¤ects on the neighborhood quality index are a unskilled RMA ( 2010 ln RM bit smaller, at between a 1.7 and 3.4 percentile points increase (14 to 28 percent of a within-region standard deviation) per standard deviation increase in skilled RMA growth. The estimates are quantitatively similar for central areas and suburbs. For the 2000-2014 period (not reported), the estimated e¤ects on suburban fraction college are about one percentage point greater and those for 2 1 Our 2007 measures of college fraction and neighborhood quality are actually the 2005-2009 5 year ACS aggregates. This is the earliest range of years for which tract-level tabulations are available after the 2000 Census. 35 suburban neighborhood quality are about twice as large as those for the 2000-2007 period, though those for central areas are little changed. We have various pieces of evidence that the IV estimates in Table 6 are well identi…ed. The clearest potential threat to identi…cation is that neighborhoods with positive unobserved shocks to neighborhood amenities were nearby areas with a 1990 employment mix that was both skilled and experienced strong national growth in the 2000-2005 period. That is, unobserved neighborhood amenity shocks like municipal investments in improved schools and parks could be correlated with the instrument and drive growth in both college fraction and neighborhood amenities. We think this sort of threat to identi…cation is unlikely for two reasons. First, the evidence discussed in Section 6 indicates that the dynamics of neighborhood change generate endogeneity bias that would lead us to underestimate the true causal e¤ects if left uncorrected. Second, eliminating region-ring controls and/or lagged demographic controls from the speci…cations reported in Table 6 Columns 1, 3, 5 and 7 always reduces the estimated IV coe¢ cients (unreported). Third, the IV estimates are greater than the OLS estimates. That is, as we include more controls for selection and instrument, the estimated coe¢ cients of interest increase, consistent with moving from the observed correlations depicted in the right graph in Figure 5 to the more ideal experiment depicted in the left graph in Figure 5. More complete controls for selection of high skilled employment growth to areas that already have many skilled workers living nearby always increases our estimates. One important mechanism for neighborhood e¤ects that we examine below is school quality. Credible evaluation of this mechanism requires exogenous variation in neighborhood attributes within school districts. The evidence in Table 7 shows that we can isolate such exogenous variation. The results in columns 1 and 4 show that growth in skilled RMA increases the educational attainment of the neighborhood and the neighborhood quality index using variation only within public school districts. The results in columns 2 and 5 show that looking between districts, the coe¢ cients on interactions between RMA and school district quality are positive for fraction college and essentially zero for neighborhood quality, weakly indicating that skilled labor demand shocks to better school districts bring in even more college graduate households in these areas. Finally, the results in columns 3 and 6 show that using only within district variation, in better school districts the same labor demand shock has a smaller e¤ect on neighborhood demographic composition.22 However, shocks to skilled RMA still induce variation in neighborhood change within school districts in about the bottom two-thirds of the school quality distribution. 2 2 We measure school district quality using the sum of standardized elementary and math scores for fourth grade, as compiled in the Stanford Education Data Archive (Reardon et al., 2018). This measure has mean 0 and standard deviation 1 across students nationwide. In our sample, the mean is -0.23 and the standard deviation is 0.99 across tracts. 36 7.1.2 Parents and Children We …rst look at the exposure to neighborhood change experienced by the sampled families in the CCP. Table 8 presents estimates of the coe¢ cients on standardized skilled RMA growth in the IV regressions described by Equations (23) and (24). The set of controls is the same as in Column 1 of Table 6, with the addition of controls for the parent’s credit score and a mortgage loan indicator for 2000 (pre-treatment). The standard errors are clustered by census tract of resdence in 2000. Columns 1-3 report the estimated e¤ects on the 2000-2007 change in tract college fraction in the 2000 tract of residence by the parent’s imputed educational attainment. Columns 4-6 report the estimated e¤ects on 2007 college fraction in the 2017 tract of residence minus 2000 college fraction for the year 2000 tract of residence. Therefore, the di¤erences indicate how 2000-2017 migration that was induced by neighborhood change after year 2000 resulted in long-run di¤erences in exposure to college educated neighbors. Consistent with the census results, the evidence in columns 1-3 of Table 8 indicates that, on average, one standard deviation of skilled RMA growth caused a 0.9 to 1.3 percentage point increase in fraction college from 2000 to 2007, in the suburban tracts where CCP sample parents and children resided in 2000. These estimates are di¤erent only due to sampling variation.23 The estimates for the city tracts are positive but not statistically signi…cant. The results in columns 4-6 of Panel A show that skilled RMA shocks are estimated to cause parents to move in a way that o¤sets the causal e¤ect of skilled RMA growth on fraction college in their 2000 tracts of residence. However, in Panel B we see that suburban children, in the least educated neighborhoods in particular, react quite di¤erently to neighborhood shocks than their parents. Experiencing a one standard deviation better labor demand shock as a teenager growing up in the suburbs results in living in a tract that has about 2-3 percentage points higher college fraction by age 28-33, relative to not moving. The largest e¤ects are for children of parents that rent in the suburbs. That is, the results in Table 8 indicate that if measured in terms of neighborhood pro…le, children in the suburbs bene…t from nearby skilled labor demand shocks while their parents experience no signi…cant e¤ects. Moreover, because the e¤ects are larger for renters, our evidence is not consistent with parents using housing equity to contribute to this improvement for their children. The results for those living less than 10 km from a CBD in 2000 do not exhibit su¢ cient statistical power to draw …rm conclusions, though the point estimates for this region are near zero. Panels A and B of Figure 6 present dynamic versions of the estimates that are reported in columns 3-6 of Table 8 for 2017. The outcome is measured as the 2007 fraction college in the tract of residence in the year indicated on the horizontal axis minus the 2000 fraction college in the 2000 tract of residence. We do not observe credit records for children and infer that the youngest cohort 2 3 The di¤erence in sample size between Panel A and Panel B is due to attrition in the parent sample between 2000 and 2017. We restrict the sample in columns 1-3 to parents that are still in the CCP in 2017 so that the sample is the same as is used in columns 4-6 of Panel A. 37 experiences 2000-2007 residential locations that are the same as the parents’, thus the graphs for children start in 2008. The thick black line in Panel A shows that suburban families’average exposure to gentri…cation slowly declined over time due to migration out of more heavily treated tracts. The long-dashed line shows that roughly the same is true for families living in less educated suburban neighborhoods in 2000. This is the key treatment exposure to neighborhood change in which we are interested. As in Table 8, we see essentially zero average long-run improvement in neighborhood fraction college for the parents in Panel A of Figure 6. However, in Panel B, we do see such long-run improvements for the children. This positive e¤ect begins in 2015, after a period of no e¤ect. Figure 7 presents the central results of the paper. The thick black lines in this …gure show the estimated dynamic e¤ects of skilled RMA shocks to tracts of year 2000 residence on suburban children’s credit score, aggregate credit card limits, 30 day loan delinquency and mortgage holding in Panels A-D, respectively. These are estimated separately for each indicated year, where Equation (24) is the estimation equation. The speci…cations match those in column 1 of Table 6 Panel A, with additional controls the for parent’s credit score and whether the parent had a mortgage in 2000. We exclude the low skilled RMA shock and Bartik ring controls from our speci…cation as they only increase the estimated coe¢ cients; by excluding these controls we estimate lower bounds on the true coe¢ cients of interest. Moreover, since we directly show no evidence of wealth e¤ects of S U 2010 ln RM Ai for parents below, which was the motivation for conditioning on 2010 ln RM Ai , in practice the inclusion of this extra control variable is unnecessary to recover credible estimates of neighborhood e¤ects. The credit score results in Panel A show essentially no e¤ect in 2008, when sample children were 19-23 years old. However, the estimates rise monotonically over the following 5 years to a statistically signi…cant 23 points, or 21 percent of a standard deviation. Over the following 9 years, the estimates bounce around with little trend from a minimum of 18 points to a maximum of 28 points. The results for the sum of credit card limits (Panel B) and mortgage holding (Panel D) exhibit similar qualitative patterns, though their increases begin after 2011. By 2017, the estimated e¤ect on the sum of credit card limits for residents of the suburbs in 2000 rises to $1978, or 16 percent of a standard deviation. The estimated e¤ect on holding a mortgage rises to 9.9 percentage points in 2016, before dipping back down to 7.3 percentage points - 24 percent and 18 percent of a standard deviation, respectively. The results for loan delinquency in Panel C are qualitatively similar, though noisier and with a bigger decline from 2016 to 2017. The fact that we see no e¤ect in 2008 for any of these outcomes and that the responses rise monotonically over time is consistent with the results re‡ecting causal relationships. Were they not causal, the unobservables driving the estimated e¤ects would have to both be correlated with counterfactual skilled RMA shocks and growing only while sampled individuals are in their late 20s and early 30s. All four of these primary outcomes except the sum of credit card limits exhibit statistically signi…cant estimates in at least 38 one year. Having established our headline result that a standard deviation higher growth in skilled RMA increases children’s outcomes by about 20 percent of a standard deviation, we examine heterogeneity of the treatment e¤ects in three dimensions. First, city residents experience essentially zero estimated e¤ects of neighborhood change. Second, for credit score and loan delinquency, the estimated e¤ects in some years are slightly greater in absolute value for those growing up in less educated neighborhoods. These are seen in the short-dashed lines in Panels A and C of Figure 7. Finally, we compare the e¤ects for children of owners versus renters, with tenure proxied by whether the parent had a mortgage in 2000. Panel B of Table 9 reports the full set of results for children in 2017, with standard errors clustered by census tract of residence in 2000. Here we see somewhat larger e¤ects on credit score and credit card limits for children of owners and somewhat larger e¤ects on holding a mortgage for children of renters. However, none of these di¤erences are statistically signi…cant. None of the estimated e¤ects on children reported in Table 9 Panel B and Figure 7 can be explained by parental wealth e¤ects. The analogous regressions for the parents, described by Equation (23), are reported in Table 9 Panel A and Figure 8. Here we see point estimates that are consistently closer to zero and not statistically distinguisable from zero. Moreover, we see no statistically signi…cant or coherent di¤erence in the parent estimates for urban versus suburban areas. To examine the extent to which school quality or other factors correlated with school quality drive our estimated neighborhood e¤ects, we include school district …xed e¤ects rather than regionring …xed e¤ects in estimation equations like Equation (24). Recall from the tract-level results that even within school districts there is still identifying variation for the e¤ects of skilled RMA growth on the share of tract residents with a college degree. The analogous CCP-based results are reported in Table 10. They show insigni…cant and small (sometimes negative) coe¢ cients for all outcomes of interest when school district …xed e¤ects are included (columns 1, 4, 7 and 10). This indicates that variation in neighborhood fraction college within school district has no estimated e¤ect on long run outcomes for resident children. As a result, we conclude that our estimated neighborhood e¤ects must run through the schools in some way. The remaining results in Table 10 indicate two additional instructive facts about mechanisms. The results in Columns 2, 4, 6 and 8 have region-ring …xed e¤ects but interact the RMA treatment with school quality. The positive coe¢ cients on these interactions indicate that higher quality schools impart larger impacts on children than do lower quality schools. If college graduates are more likely to send their children to public schools if they are of higher quality, these results are consistent with the idea that schools mediate our estimated e¤ects in Table 9. However, these positive interactions disappear when controlling for school district …xed e¤ects, as seen in Columns 3, 6, 9 and 12. In better school districts, variation in neighborhood quality is not estimated to 39 e¤ect long-run outcomes of children. We caution, however, that these interacted speci…cations are under-powered. Overall, our analysis of the CCP provides compelling quantitative evidence on the existence of neighborhood e¤ects that operate through schools. Quantifying these neighborhood e¤ects indicates that a shock that causes neighborhood quality to improve by 25-38 percent of a standard deviation causes suburban children’s long-run outcomes to increase by 13-17 percent of a standard deviation but their parents’outcomes to not change. Dividing the estimate of the e¤ect on children by the estimate of the e¤ect on the neighborhood, implies a neighborhood e¤ect that can be summarized as a one standard deviation improvement in neighborhood quality leading to about a half of a standard deviation improvement in children’s long-run outcomes. A set of robustness checks presented in Appedix Tables A7-A9 reveal that adding controls for skill-speci…c Bartik shocks amongst commuting destinations 0-10 minutes and 10-20 minutes away (Bartik ring controls) to the speci…cations in Tables 8-10 show somewhat larger e¤ects of skilled RMA growth on long-run outcomes for the children and smaller e¤ects of skilled RMA growth on neighborhood educational attainment. These estimates imply larger neighborhood e¤ects than those reported in the previous paragraph as they have a larger numerator and smaller denominator. However, they also have slightly larger standard errors. For this reason, we view the results presented in Tables 8-10 as more conservative estimates. 7.2 1990-2000 Period As we discussed earlier, the 1990-2000 period presents some additional challenges for identifying neighborhood e¤ects. First, the skilled and unskilled RMA shocks are more highly correlated during this period, making econometric identi…cation more di¢ cult. Second, the identifying variation during this period produces parental wealth e¤ects which complicate the recovery of neighborhood e¤ects. Finally, the PSID’s smaller sample size generates lower statistical power. Despite these issues, the evidence shown here is fully consistent with the CCP results, shown above. So as to maintain statistical power, we focus on results from speci…cations where we do not interact growth in RMA with urban/suburban status. The census tract results in Appendix Table A6 indicate that shocks to RMA growth signi…cantly increase the tract-level fraction college and neighborhood quality, but only without controls for aggregated commute time ring Bartik shocks (Bartik ring controls). For this period, a one standard deviation increase in uni…ed RMA growth, with identifying variation from counterfactual skilled S gAi ), generates 1990-2000 increases in college fraction of 1 to 3 percentage RMA growth ( 2000 ln RM points and contemporaneous increases in neighborhood quality of 1.3 to 4.0 percentiles. Table 11 presents estimates of the e¤ects of RMA growth on family outcomes in the PSID. Panel A shows the e¤ects on outcomes relevant to families and children while they still reside with 40 their parents and Panel B shows long-run e¤ects on children. We present IV regressions with region …xed e¤ects and the full set of demographic controls and tract-level controls, with standard errors clustered by 1990 census tract of residence. The smaller sample size of the PSID means that there is not enough variation within region-rings for estimation. The results in Panel A show that RMA growth is estimated to e¤ect growth in family incomes in the 1990 tract of residence. Each standard deviation in RMA growth is estimated to lead to 32% higher family income growth during the 1990s.The results is driven by parents that are college graduates. As expected, shocks to RMA growth are estimated to have caused the fraction college in tract of 1990 residence to grow, with estimates in line with our Census results in Table A6. Scores on an applied problem test taken in the teenage years are 26% of a standard deviation higher for each additional standard deviation of RMA growth. Point estimates of the e¤ect of RMA growth on scores on a number of other tests (not shown) are also strongly positive, but they are not statistically signi…cant. We have two strong results for children’s long-run outcomes. Their 2015 family income is about 27 log points larger (30 percent of a standard deviation) for a one standard deviation greater RMA shock in youth and their 2015 employment rate is 13 percentage pionts higher, though only marginally signi…cant. Because we …nd measurable e¤ects on both children and parents, the PSID data do not allow us to distinguish between the e¤ects of improved labor market opportunities increasing parental investment versus neighborhood e¤ects that a¤ect children directly. However, the PSID does provide evidence that the sum of these objects is positive for the children. 8 Conclusions Using quasi-random variation in 1990-2000 and 2000-2005 skilled labor demand shocks for each urban census tract in the U.S. as a source of identifying variation, this paper investigates the e¤ects of neighborhood change on incumbent residents. Our estimates indicate that skilled labor demand shocks to areas within commuting distance raise neighborhood quality, fraction college and incomes in these neighborhoods. Our evidence from the PSID on the e¤ects of 1990-2000 growth in accessible employment opportunities shows that it results in increases in parental income, child test scores for teenagers and family income at age 25 to 43. Our evidence on 2000-2005 growth in employment opportunities from the CCP shows resulting improvements in a number of credit outcomes for those who were 11-15 at the beginning of exposure to treatment and living in the suburbs, but little or negative e¤ects for their parents. These estimated positive impacts for the children only manifest themselves once the children are at least age 19, and increase monotonically with age before levelling o¤. Taken together, the evidence in this paper indicates that a one standard deviation shock to neighborhood quality improves suburban children’s long-run outcomes by 40 to 70 percent of a 41 standard deviation, with these neighborhood e¤ects more pronounced for children growing up in less educated neighborhoods. Schools appear to represent the primary channel through which these neighborhood spillovers operate. Our analysis is one of the …rst attempts to look at the e¤ects of neighborhoods that change around people rather than children’s exposure to neighborhood change through the migration choices of their parents. Nevertheless, while di¤erent empirical settings make the results somewhat hard to compare, we …nd evidence that is broadly consistent with that from the literature using household moves. For example, Chetty and Hendren (2018a) and Laliberté (2018) …nd that one additional year of exposure to a neighborhood promotes convergence of about 4.4 percent between new arrivals up to age 23 and incumbent residents for various long-run labor market outcomes. Someone who is moved at age 13, the midpoint of our cohort, to a neighborhood that is one standard deviation higher quality would thus end up 44 percent of a standard deviation higher in the distribution of outcomes by age 23. This estimate is at the lower end of our range of estimates for the suburbs, but is consistent with an average of our city and suburb e¤ects. The compendium of evidence on neighborhood e¤ects suggests two fruitful avenues for future research. First, while it seems that schools matter as an important causal mechanism, it is not clear why schools matter. That schools matter more in better quality districts and mostly in the suburbs, where higher educated families are more likely to send their children to public schools, suggests that it is not simply a consequence of improvements in the tax base. But whether the e¤ects we …nd are truly from peer interactions, more active engagement from parents in schools, or more political support for expanding resources to public schools seems important. Second, the existing research on neighborhood e¤ects has not had the statistical power to establish the extent to which neighborhood e¤ects operate nonlinearly as functions of attributes of the treatment and/or the treated. If spillovers are linear, no aggregate gain is available from reallocating people across neighborhoods. In contrast, knowing details about the nature of any potential nonlinear relationships would allow for improving the e¢ cacy of inclusionary and low income housing policies. 42 References Adao, Rodrigo, Michal Kolesar and Eduardo Morales (2018) "Shift-Share Designs: Theory and Inference," NBER Working Paper No. 24944. Ahlfeldt, Gabriel, Stephen Redding, Daniel Sturm, Nikolaus Wolf (2015) "The Economics of Density: Evidence from the Berlin Wall," Econometrica 83(6): 2127-2189. Aliprantis, Dionissi and F. G.-C. Richter (2018) "Evidence of Neighborhood E¤ects from Moving to Opportunity: LATEs of Neighborhood Quality," manuscript. Altonji, J.G. & R.K. Mans…eld (2018) "Estimating Group E¤ects Using Averages of Observables to Control for Sorting on Unobservables: School and Neighborhood E¤ects," American Economic Review, 108(10): 2902-2946. Allen, Treb, C. Arkolakis & Li (2016) "Optimal City Structure," manuscript. Andersson, F., J. Haltiwanger, M. Kutzbach, G. Palloni, H. Pollakowski, and D. H. Weinberg (2016). "Childhood housing and adult earnings: A between-siblings analysis of housing vouchers and public housing." NBER Working Paper No. 22721. Åslund, O., Edin, P.-A., Fredriksson, P. and H. Grönqvist (2011), “Peers, neighborhoods, and immigrant student achievement: Evidence from a placement policy,” American Economic Journal: Applied Economics 3, 67-95. Bartik, Timothy (1991) Who Bene…ts from State and Local Economic Development Policies? Kalamazoo, MI: W.E. Upjohn Institute for Employment Research. Baum-Snow, Nathaniel & Lu Han (2018) "The Microgeography of Housing Supply," manuscript. Baum-Snow, Nathaniel & Daniel Hartley (2018) "Accounting for Central Neighborhood Change," manuscript. Bayer, P., Ross, S.L. and G. Topa (2008), “Place of work and place of residence: Informal hiring networks and labor market outcomes,” Journal of Political Economy 116, 1150-1196. Beaman, L. (2012), “Social networks and the dynamics of labor market outcomes:Evidence from refugees resettled in the U.S.,” Review of Economic Studies 79, 128-161. Blanchard, Olivier Jean and Lawrence F. Katz. (1992) "Regional Evolutions," Brookings Papers on Economic Activity, 23(1): 1-76. Borusyak, K. and X. Jaravel. (2018) "Consistency and Inference in Bartik Research Designs," manuscript. Brummet, Quentin and Davin Reed (2019) "Gentri…cation and the Location and Well-Being of Original Neighborhood Residents," manuscript. Charles, Kerwin Ko…, Erik Hurst and M. J. Notowidigdo. (2018) "Housing Boom and Busts, Labor Market Opportunities, and College Attendance," American Economic Review, 108(10): 29472994. Chetty, Raj, John N. Friedman, Nathaniel Hendren, Maggie R. Jones and Sonya R. Porter. (2018) "The Opportunity Atlas: Mapping the Childhood Roots of Social Mobility," manuscript. 43 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 129: 1553-1623. Chetty, Raj, Nathaniel Hendren, and Lawrence Katz (2016) "The E¤ects of Exposure to Better Neighborhoods on Children: New Evidence from the Moving to Opportunity Experiment," American Economic Review 106(4): 855-902. Chetty, Raj and Nathaniel Hendren (2018a) "The Impacts of Neighborhoods on Intergenerational Mobility I: Childhood Exposure E¤ects," Quarterly Journal of Economics, forthcoming. Chetty, Raj and Nathaniel Hendren (2018b) "The Impacts of Neighborhoods on Intergenerational Mobility II: County Level Estimates," Quarterly Journal of Economics, forthcoming. Couture, Victor and Jessie Handbury (2016) "Urban Revival in America, 2000 to 2010," manuscript. Couture, Victor, G. Duranton & M. Turner (2017) "Speed," Review of Economics and Statistics, forthcoming. Chyn, Eric (2016) "Moved to Opportunity: The Long-Run E¤ect of Public Housing Demolition on Labor Market Outcomes of Children," manuscript. Currie, Janet, Lucas Davis, Michael Greenstone, and Reed Walker (2015) "Environmental Health Risks and Housing Values: Evidence from 1,600 Toxic Plant Openings and Closings," American Economic Review, 105:2. Damm, A.P. (2009) “Ethnic enclaves and immigrant labor market outcomes: Quasi-experimental evidence,” Journal of Labor Economics 27, 281-314. Damm, A.P. and C. Dustmann (2014) “Does growing up in a high crime neighborhood a¤ect youth criminal behavior?” American Economic Review 104, 1806-1832. Davis, Steve & Till von Wachter (2012) "Recessions and the Costs of Job Loss" Brookings Papers on Economic Activity. Donaldson, Dave and Richard Hornbeck (2016) "Railroads and American economic growth: A “market access” approach." The Quarterly Journal of Economics 131(2): 799-858. Edin, P.-A., Fredriksson, P. and O. Aslund (2003) “Ethnic enclaves and the economic success of immigrants: Evidence from a natural experiment,” Quarterly Journal of Economics 118, 329-357. Reardon, S. F., Ho, A. D., Shear, B. R., Fahle, E. M., Kalogrides, D., & DiSalvo, R. (2018). Stanford Education Data Archive (Version 2.1). Retrieved from http://purl.stanford.edu/db586ns4974. Fogli, Alessandra and Veronica Guerrieri (2018) “The End of the American Dream? Inequality and Segregation in US Cities," manuscript. Goldsmith-Pinkham, Paul, Isaac Sorkin & Henry Swift (2018) "Bartik Instruments: What, When, Why and How?" manuscript. Gould, E.D., Lavy, V. and D. Paserman (2011) “Sixty years after the magic carpet ride: The long-run e¤ect of the early childhood environment on social and economic outcomes,” Review of 44 Ecoconomic Studies 78, 938-973. Gould, E.D., Simhon, A. and B.A. Weinberg (2015) "Does Parental Quality Matter? Evidence on the Transmission of Human Capital Using Variation in Parental In‡uence from Death, Divorce, and Family Size," manuscript. Graham, B.S. (2018) "Identifying and Estimating Neighborhood E¤ects," Journal of Economic Literature 56(2): 450-500. Hellerstein, J.K., M.P. McInerney and D. Neumark (2011) “Neighbors and co-workers: The importance of residential labor market networks,” Journal of Labor Economics 29, 659-695. Hellerstein, J.K., Kutzbach, M.J. and D. Neumark (2014) “Do labor market networks have an important spatial dimension?” Journal of Urban Economics 79, 39-58. Hoynes, H.W., D.W. Schanzenbach and D. Almond (2018) "Long Run Impacts of Childhood Access to the Safety Net," American Economic Review, forthcoming. Laliberté, J-W.P. (2018) "Long-term Contextual E¤ects in Education: Schools and Neighborhoods," manuscript. Meltzer, R. and P. Ghorbani (2017) "Does gentri…cation increase employment opportunities in low-income neighborhoods?," Regional Science and Urban Economics 66, 52-73. Monte, Ferdinando, Steven Redding & Esteban Rossi-Hansberg (2016) "Commuting, Migration and Local Employment Elasticities," American Economic Review Oreopoulous, P. (2003) “The long-run consequences of living in a poor neighborhood,”Quarterly Journal of Economics 118, 1533-1575. Sampson, R. J. (2008) Moving to inequality: neighborhood e¤ects and experiments meet social structure. American Journal of Sociology 114(1), 189-231. Severen, C. (2019) "Commuting, Labor, and Housing Market E¤ects of Mass Transportation: Welfare and Identi…cation," manuscript Tian, L. (2018) “Division of Labor and Productivity Advantage of Cities: Theory and Evidence from Brazil," manuscript. Tsivanidis, N. (2018) "The Aggregate And Distributional E¤ects Of Urban Transit Infrastructure: Evidence From Bogotá’s TransMilenio," manuscript. 45 Table 1: Summary Statistics Residents of Sample Census Tracts in 254 Regions Panel A: Census Tract Sample (32,515 Tracts) Mean Growth Rates Fraction College Neighborhood Quality (Pctiles) Employment in Tract Unified RMA Skilled RMA Unskilled RMA Shocks Tract Bartik Shock, Skilled Tract Bartik Shock, Unskilled Simulated RMA, Skilled Simulated RMA, Unskilled 0.035 -0.9 -0.21 0.05 0.09 0.03 0.14 0.03 0.07 0.01 St Dev. In Rgn Stdev. 2000-2007/10 0.073 0.069 13.1 12.2 0.88 0.86 0.05 0.015 0.07 0.018 0.07 0.015 2000-2005 0.03 0.02 0.04 0.04 0.003 0.001 0.006 0.001 Mean 0.026 -2.7 0.17 0.03 0.13 -0.02 0.28 0.07 0.15 0.03 St Dev. 1990-2000 0.051 12.5 0.85 0.16 0.17 0.17 1990-2000 0.05 0.04 0.009 0.010 In Rgn Stdev. 0.048 12.0 0.76 0.08 0.09 0.08 0.04 0.04 0.002 0.002 Panel B: CCP Samples (Children Born 1985-1989, 9,083 Tracts in 2000) Migration Since 2000 Fraction College (2000 or 2007) Equifax Risk Score ™ Sum of Credit Card Limits Any Loan 30 Days Past Due Mortgage Indicator Parents, 2000 Mean St Dev. 0.00 0.00 0.26 0.18 655 108 11,224 16,571 0.23 0.42 0.45 0.50 Obs 10,859 10,858 10,352 10,333 10,333 10,333 Parents, 2017 Mean 0.69 0.30 703 19,862 0.18 0.43 St Dev. 0.46 0.18 108 26,227 0.38 0.49 Obs 10,433 10,397 9,408 9,394 9,394 9,394 Panel C: PSID Sample (1,519 Children born 1972-1990, 684 Tracts in 1990) ln (2015 Family Income) Employed in 2015 Max Years of Education Age Adjusted Applied Problem Score Migration Probability, 1990-2001 Dln (Parent Family Income), 1990-2001 Avg Annual Dln Rent, 1990 Tract Dln (Parent Rent), 1990-2001 Mean 10.3 0.79 13.8 0.02 0.67 0.32 0.00 0.27 Standard Dev. 1.1 0.41 2.4 0.39 0.47 0.87 0.02 1.00 Children, 2017 Mean 0.85 0.30 653 8,408 0.25 0.22 St Dev. 0.36 0.18 109 10,962 0.43 0.41 Obs 10,859 10,794 10,277 10,279 10,279 10,279 Table 2: Effects of Tract Level Bartik Shocks on Tract Level Employment Growth Rate by Education Total (1) (2) (3) College (4) (5) (6) <HS to Some College (7) (8) Panel A: 2000-2010 Skilled Shock S Bartikj Unskilled Shock 1.600*** 1.463*** 0.726*** 1.618*** (0.230) (0.278) 0.129 (0.251) (0.303) -0.838*** U Bartikj Skilled Shock (0.147) 1.385*** (0.160) 0.962*** (0.273) 0.550*** (0.119) (0.144) 120.5 66.21 S (85.34) 1.604*** (77.01) 1.379*** S (0.304) -0.816*** (0.273) 0.545*** U (0.164) -2.173 (0.147) 0.765 U (1.762) (1.589) Bartikj if neg., 0 o/w Skilled Shock Bartikj if pos., 0 o/w Unskilled Shock Bartikj if neg., 0 o/w Unskilled Shock Bartikj if pos., 0 o/w Observations (Within) R-Squared 32,459 0.030 32,459 0.030 32,362 0.041 32,362 0.042 32,362 0.042 32,449 0.032 32,449 0.032 32,449 0.032 Panel B: 1990-2000 Skilled Shock S Bartikj Unskilled Shock 0.977*** 1.933*** 1.259*** 3.392*** (0.108) (0.168) -1.254*** (0.113) (0.174) -2.796*** U Bartikj Skilled Shock (0.168) 1.323*** (0.174) 0.269** (0.169) -0.745*** (0.109) (0.169) -9.212*** -10.34*** S (2.044) 3.478*** (1.981) 1.400*** S (0.175) -1.941*** (0.169) 0.118 U (0.388) -2.910*** (0.376) -0.866*** U (0.188) (0.183) Bartikj if neg., 0 o/w Skilled Shock Bartikj if pos., 0 o/w Unskilled Shock Bartikj if neg., 0 o/w Unskilled Shock Bartikj if pos., 0 o/w Observations (Within) R-Squared 32,016 0.194 32,016 0.196 32,016 0.169 32,016 0.176 32,016 0.177 32,016 0.197 32,016 0.199 32,016 0.200 Notes: Each column in each panel reports the results from a separate regression of change in log tract employment for the group indicated at top on the indicated tract level labor demand shocks, region-ring fixed effects and our base set of controls. Bartik shocks in Panel A are for 2000-2005 whereas those in Panel B are for 1990-2000. The full sample in each panel includes all tracts within 20 km of a CBD and with positive employment in the base and terminal years. Base controls are a quadratic in CBD distance, log 1990 tract employment and 10 and 20 year lags of the house price index, rent index, log population, log family income, share African American, share White, share college graduate, and share with less than high school. There are 2,060 inluded region-ring fixed effects. Heteroskedasticity robust standard errors are reported. Standard errors constructed using Adao, Kolesar & Morales' (2018) approach are about twice as large. Table 3: Descriptive OLS Regressions About Neighborhood Change and Relationships With RMA, 2000-2010 1990-2000 Change in Dependent Variable D ln Skilled RMA, 2000-2010 Obs (Within) R-Squared Region-Ring Fixed Effects Base Controls Fraction College Graduate Neighborhood Quality Index 90-00 Chg 2000 Level 90-00 Chg 2000 Level 2000-2007 Change 2000-2007 Change (1) (2) (3) (4) (5) (6) (7) (8) -0.124*** -0.168*** -0.255*** -0.288*** (0.00641) (0.00667) (0.00562) (0.00572) 0.00205*** 0.00209*** 0.0772 0.390*** (0.000324) (0.000326) (0.0604) (0.0667) log Average HH Income 90-00 Chg 2000 Level 2000-2007 Change (9) (10) (11) (12) -0.301*** -0.379*** (0.00610) (0.00625) 0.00229** 0.00233** (0.000911) (0.000921) 32,413 0.012 32,413 0.053 32,430 0.115 32,460 0.870 32,510 0.063 32,510 0.114 32,515 0.079 32,515 0.745 32,286 0.075 32,286 0.155 32,342 0.135 32,395 0.783 Yes No Yes Yes Yes Yes Yes Yes Yes No Yes Yes Yes Yes Yes Yes Yes No Yes Yes Yes Yes Yes Yes Notes: Each column is a separate regression of the indicated variable at top on its change over the prior decade or the subsequent change in log skilled RMA and the controls listed at bottom. Base controls are origin tract skill-specific Bartik shocks, a quadratic in CBD distance, log 1990 tract employment and 10 and 20 year lags of a house price index, rent index, log population, log family income, share African American, share White, share college graduate, and share with less than high school. Results using a home price index as an alternative outcome show the same qualitative pattern as that seen for the three outcomes shown. Table 4: Examination of Differential Pre-2000 Trends, IV Regressions D ln Skilled RMA Obs First Stage F Unskilled Simul. RMA Controls Bartik Ring Controls Region-Ring Fixed Effects Base Controls Fraction College Graduate (1) (2) (3) (4) 0.0253*** -0.00265 0.0143*** 0.0111** (0.00510) (0.00752) (0.00348) (0.00482) Neighborhood Quality Index (5) (6) (7) (8) 3.146*** -0.809 1.982*** 0.749 (0.918) (1.406) (0.644) (0.887) 32,294 143.6 32,294 57.05 32,294 277.0 32,294 142.1 32,379 143.2 32,379 56.98 32,379 278.3 32,379 143.2 No No Yes Yes No Yes Yes Yes Yes No Yes Yes Yes Yes Yes Yes No No Yes Yes No Yes Yes Yes Yes No Yes Yes Yes Yes Yes Yes Notes: Each column reports coefficients from a separate IV regression of the 1990-2000 change in the variable listed at top on the change in log skilled RMA in the following decade, instrumented with the change in counterfactual log skilled RMA. See the notes to Table 3 for the base controls. The instrument is built using employment in all tracts excluding the origin tract. Table 5: First Stage Results - Resident Market Access (1) Change in Cntrfctl. log Skill. RMA, 0.0573*** < 10 km from CBD (0.00686) Change in Cntrfctl. log Skill. RMA, 0.0643*** > 10 km from CBD (0.00740) Observations (Within) R-Squared / R-Squared Cntrfctl. log Unskill. RMA Controls Bartik Ring Controls Region-Ring FE Base Controls Census Sample (2) (3) 0.0440*** 0.122*** (0.00719) (0.0164) 0.0429*** 0.300*** (0.00821) (0.0181) (4) 0.0893*** (0.0166) 0.233*** (0.0189) (5) 0.100*** (0.0319) 0.142*** (0.0372) CCP Child Sample (6) (7) 0.0923*** 0.0256 (0.0326) (0.0701) 0.129*** 0.177** (0.0439) (0.0694) (8) -0.0286 (0.0700) 0.0705 (0.0704) 32,515 0.031 32,515 0.040 32,515 0.038 32,515 0.044 10,251 0.956 10,251 0.957 10,251 0.956 10,251 0.957 No No Yes Yes No Yes Yes Yes Yes No Yes Yes Yes Yes Yes Yes No No Yes Yes No Yes Yes Yes Yes No Yes Yes Yes Yes Yes Yes Notes: All RMA measures are expressed in standard deviation units. Base controls are listed in the notes to Table 3. Testing whether the coefficients are jointly equal to zero in columns (5) - (8) yields F-statistics of 12.1, 6.9, 3.3 and 0.6, respectively. Table 6: Effects of Skilled RMA on Census Tract Outcomes, 2000-2007 (1) OLS, 1st Diff. D Skilled RMA, <10 km from CBD D Skilled RMA, >10 km from CBD IV, 1st Diff. D Skilled RMA, <10 km from CBD D Skilled RMA, >10 km from CBD Observations First Stage F Unskilled Simul. RMA Controls Bartik Ring Controls Region-Ring FE Base Controls Fraction College Graduate (2) (3) (4) (5) Neighborhood Quality Index (6) (7) (8) 0.000166 0.000123 0.000112 0.000100 (0.000471) (0.000471) (0.000472) (0.000472) 0.00178*** 0.00157*** 0.00175*** 0.00158*** (0.000597) (0.000603) (0.000597) (0.000603) 0.0256 (0.0795) -0.0295 (0.101) 0.0267 (0.0795) 0.00749 (0.102) 0.0206 (0.0795) -0.0329 (0.101) 0.0255 (0.0795) 0.00837 (0.102) 0.0249*** (0.00774) 0.0261*** (0.00687) 0.0206** (0.00931) 0.0194* (0.0101) 0.0167** (0.00845) 0.0161*** (0.00380) 0.0210** (0.00965) 0.0181*** (0.00508) 2.405* (1.253) 2.624** (1.101) 2.047 (1.535) 2.145 (1.646) 1.739 (1.408) 1.729*** (0.627) 3.387** (1.634) 2.878*** (0.851) 32,283 63.31 32,283 33.09 32,283 50.04 32,283 38.44 32,375 63.46 32,375 33.20 32,375 49.76 32,375 38.11 No No Yes Yes No Yes Yes Yes Yes No Yes Yes Yes Yes Yes Yes No No Yes Yes No Yes Yes Yes Yes No Yes Yes Yes Yes Yes Yes Notes: Entries show OLS or IV coefficients and standard errors. RMA measures are expressed in standard deviation units. Analogous results for the CCP sample are about half the magnitudes reported above, reflecting the larger first stage coefficients reported in Table 5. Table 7: Effects of Standardized Skilled RMA on 2000-2007 Tract Outcomes - School District Analysis (1) IV, 1st Diff. DRMA, <10 km from CBD DRMA, >10 km from CBD DRMA, <10 km from CBD X School Quality Pctile DRMA, >10 km from CBD X School Quality Pctile Observations First Stage F Region-Ring FE Base Controls School District FE Region-Ring FE College Graduate Share (2) (3) (4) Neighborhood Quality Index (5) (6) 0.0159*** (0.00257) 0.0160*** (0.00260) 0.0257*** (0.00778) 0.0202*** (0.00509) 0.000740*** (0.000253) 0.000926*** (0.000190) 0.0181*** (0.00288) 0.0182*** (0.00291) -0.0153*** (0.00409) -0.0153*** (0.00420) 0.736* (0.411) 0.830** (0.416) 2.561** (1.264) 2.387*** (0.822) -0.0464 (0.0410) 0.0198 (0.0307) 1.088** (0.455) 1.222*** (0.460) -1.670** (0.649) -1.503** (0.666) 32,015 283.7 31,373 30.91 31,131 51.57 32,107 293.7 31,436 30.85 31,194 51.33 No Yes Yes No Yes Yes No Yes No Yes Yes No No Yes Yes No Yes Yes No Yes No Yes Yes No Notes: Each column reports one IV regression that includes the set of controls listed at bottom. College-oriented simulated DlnRMA interacted with the two indicated CBD distance bands enters as instruments, as in Table 5. CCP sample results are similar, though with larger standard errors. Analogous regressions including additional controls for low and high skilled shocks in 0-10 and 10-20 minute commute time rings show similar results. Table 8: Impacts of RMA on Neighborhood Exposure and Migration Region-Ring Fixed Effects and Full Set of Controls Educational Attainment Weight 2000-2007 Change in Fraction College in Tract of 2000 Residence None < HS College + Fraction College in 2017 Tract of Residence Fraction College in 2000 Tract of Residence None < HS College + Panel A: Parents All, 0-10 km from CBD All, 10-20 km from CBD Observations 0.00722 (0.00697) 0.00872** (0.00420) 0.00459 (0.00784) 0.00832* (0.00489) 0.000163 (0.00651) 0.0124** (0.00507) -0.0102 (0.0163) -0.000407 (0.00939) -0.0206 (0.0213) 0.00843 (0.0121) -0.0222 (0.0186) -0.00470 (0.0119) 9,848 1,769 3,163 9,826 1,765 3,156 Panel B: Children (Born 1985-1989) All, 0-10 km from CBD All, 10-20 km from CBD Observations Renters, 0-10 km from CBD Owners, 0-10 km from CBD Renters, 10-20 km from CBD Owners, 10-20 km from CBD 0.00628 (0.00663) 0.0101** (0.00428) 0.00429 (0.00706) 0.0123** (0.00539) 0.000926 (0.00652) 0.0134*** (0.00513) -0.00174 (0.0187) 0.0210* (0.0116) -0.00690 (0.0208) 0.0348** (0.0142) 0.00185 (0.0216) 0.0205 (0.0144) 10,246 1,853 3,278 10,188 1,845 3,258 0.00871 (0.00692) 0.00595 (0.00684) 0.0113*** (0.00436) 0.00815** (0.00410) 0.00670 (0.00734) 0.00388 (0.00726) 0.0138** (0.00554) 0.0106** (0.00528) 0.00301 (0.00681) 0.000256 (0.00663) 0.0143*** (0.00524) 0.0112** (0.00494) 0.00132 (0.0189) -0.00188 (0.0188) 0.0230* (0.0120) 0.0184 (0.0116) 0.00155 (0.0218) -0.00891 (0.0218) 0.0412*** (0.0148) 0.0273* (0.0142) -0.00530 (0.0228) 0.00405 (0.0221) 0.0174 (0.0154) 0.0255* (0.0148) Observations 10,246 1,853 3,278 10,188 1,845 3,258 Notes: Each block of results is from a separate IV regression of the outcome at top on region-ring fixed effects, the tract's 2000-2005 Bartik shock, 1990 tract employment, 10 and 20 year lags of demographic controls and RMA interacted with the variables listed at left. Each RMA variable is instrumented with the growth rate in counterfactual skilled RMA interacted with the objects listed at left. Representative first stage results are in Table 5. Standard errors are clustered by census tract of residence in 2000. Table 9: Consumer Credit Panel Results - Impacts of Standardized RMA Region-Ring Fixed Effects and Full Set of Controls Outcome Educational Attainment Weight None Equifax Risk Score ™ < HS College + Sum of Credit Card Limits None < HS College + Any Loan 30 Days Past Due None < HS College + None Mortgage Dummy < HS College + Panel A: Parent Outcomes, 2017 All, 0-10 km from CBD All, 10-20 km from CBD Observations 3.297 (9.082) -5.010 (5.079) 18.77 (13.23) -9.943 (7.398) -10.17 (9.387) -8.433 (5.538) 2,967 (2,656) 456.5 (2,625) 3,977 (3,401) -335.5 (2,922) 2,228 (2,843) -2,615 (2,372) -0.00776 (0.0411) -0.0181 (0.0201) -0.0788 (0.0627) 0.00764 (0.0304) 0.0208 (0.0395) -0.0137 (0.0228) 0.0583 (0.0584) 0.0288 (0.0316) 0.0984 (0.0853) 0.0504 (0.0433) 0.0272 (0.0554) 0.00535 (0.0364) 8,995 1,594 2,924 8,984 1,589 2,924 8,984 1,589 2,924 8,984 1,589 2,924 Panel B: Child Outcomes, 2017 All, 0-10 km from CBD All, 10-20 km from CBD Renters, 0-10 km from CBD Owners, 0-10 km from CBD Renters, 10-20 km from CBD Owners, 10-20 km from CBD Observations -0.203 (11.97) 18.28** (9.184) 5.014 (13.99) 18.91 (12.01) 6.393 (12.00) 14.86 (9.320) -45.50 (1,147) 1,978* (1,028) 904.3 (1,383) 2,634** (1,306) -1,040 (1,395) 1,767* (1,063) -0.0794 (0.0577) -0.0205 (0.0343) -0.126 (0.0813) 0.0130 (0.0463) -0.0447 (0.0512) 0.00637 (0.0361) 0.0633 (0.0532) 0.0728* (0.0392) 0.0906 (0.0638) 0.0825* (0.0496) 0.0464 (0.0546) 0.0639 (0.0397) -1.150 (11.96) -0.160 (12.06) 17.77* (9.456) 19.76** (9.200) 3.069 (14.04) 5.399 (14.27) 17.71 (12.40) 21.59* (12.14) 4.044 (12.16) 7.147 (12.12) 13.78 (9.593) 17.88* (9.234) -184.3 (1,148) 4.265 (1,159) 1,926* (1,061) 2,129** (1,030) 963.5 (1,373) 848.7 (1,405) 2,662** (1,331) 2,607** (1,296) -1,562 (1,476) -773.7 (1,451) 1,571 (1,123) 2,416** (1,083) -0.0529 (0.0552) -0.0884 (0.0562) -0.0103 (0.0352) -0.0508 (0.0353) -0.0954 (0.0755) -0.143* (0.0781) 0.0320 (0.0485) -0.0226 (0.0476) -0.0192 (0.0518) -0.0579 (0.0517) 0.0158 (0.0359) -0.0248 (0.0356) 0.0792 (0.0551) 0.0585 (0.0560) 0.0794** (0.0394) 0.0530 (0.0385) 0.0878 (0.0626) 0.0942 (0.0645) 0.0816 (0.0507) 0.0821* (0.0491) 0.0685 (0.0570) 0.0359 (0.0568) 0.0726* (0.0413) 0.0354 (0.0404) 9,736 1,737 3,151 9,736 1,737 3,146 9,736 1,737 3,146 9,736 1,737 3,146 Notes: Reported coefficients are graphed as the final points in Figures 7 and 8. See the notes to Table 8 for a description of the specifications. Specifications with additional Bartik ring controls generate coefficients that are 30-50 percent larger and standard errors that are 10-30 percent larger. Standard errors are clustered by census tract of residence in 2000. Table 10: Consumer Credit Panel Results for Children - School District Fixed Effects Equifax Risk Score ™ DRMA, <10 km from CBD DRMA, >10 km from CBD DRMA, <10 km from CBD X School Quality Pctile DRMA, >10 km from CBD X School Quality Pctile Observations School District FE Region-Ring FE Sum of Credit Card Limits Any Loan 30 Days Past Due Mortgage Indicator -2.708 (7.594) -3.430 (7.698) -3.034 (12.63) 18.11* (9.493) 0.247 (0.599) 0.995** (0.444) -0.118 (8.965) -0.818 (9.111) -12.45 (12.43) -11.17 (12.68) -748.3 (690.5) -741.2 (702.2) -581.7 (1,211) 1,883* (1,057) 143.0** (67.96) 135.7** (53.86) -658.6 (765.5) -672.9 (779.9) -301.9 (1,092) -301.7 (1,111) 0.0243 (0.0340) 0.0233 (0.0344) -0.0621 (0.0582) -0.0173 (0.0353) 0.00115 (0.00261) -0.00294* (0.00178) 0.00713 (0.0406) 0.00555 (0.0411) 0.0755 (0.0531) 0.0711 (0.0539) -0.00102 (0.0286) -0.00626 (0.0290) 0.0568 (0.0548) 0.0678* (0.0399) 0.00128 (0.00271) 0.00133 (0.00185) 0.00304 (0.0317) -0.00292 (0.0322) -0.0126 (0.0427) -0.0135 (0.0434) 9,736 Yes No 9,374 No Yes 9,374 Yes No 9,736 Yes No 9,372 No Yes 9,372 Yes No 9,736 Yes No 9,372 No Yes 9,372 Yes No 9,736 Yes No 9,372 No Yes 9,372 Yes No Notes: Results are analogous to those reported in Table 9 except for the included fixed effects and interaction terms. Standard errors are clustered by census tract of residence in 2000. Table 11: PSID Results, All Tracts 0-20 km from CBDs Panel A: Outcomes When Children and Parents Live Together Dln(Family Income), 1990-2001 1990-2000 DFraction College, 1990 Tract Children's Applied Problem Score First Stage F 0.315* (0.185) 11.87 0.0285** (0.0132) 12.45 0.255** (0.127) 6.831 D ln (RMA) X Parent >= Coll D ln (RMA) X Parent < Coll First Stage F 0.455** (0.198) 0.102 (0.221) 5.415 0.0289** (0.0129) 0.0280* (0.0144) 5.614 0.256** (0.123) 0.251 (0.156) 3.099 D ln (RMA) X own D ln (RMA) X rent First Stage F 0.280 (0.204) 0.202 (0.303) 3.078 0.0286** (0.0122) 0.0287* (0.0172) 2.977 0.266** (0.128) 0.310 (0.397) 0.203 Observations 1,371 1,416 696 D ln (RMA) Panel B: Long-run Child Outcomes Max Years of Education Employed in 2015 First Stage F -0.175 (0.471) 13.92 0.133* (0.0767) 12.56 0.274** (0.113) 12.33 D ln (RMA) X Parent >= Coll D ln (RMA) X Parent < Coll First Stage F 0.419 (0.639) -0.818 (0.662) 6.689 0.159** (0.0773) 0.0876 (0.0877) 5.871 0.308*** (0.111) 0.169 (0.139) 5.264 D ln (RMA) X own D ln (RMA) X rent First Stage F -0.188 (0.509) -0.228 (0.816) 2.196 0.116 (0.0830) -0.0158 (0.187) 1.967 0.278** (0.116) 0.303 (0.207) 2.966 Observations 1,374 1,057 945 D ln (RMA) ln (2015 Family Notes: Estimates are from IV regressions that include controls for region fixed effects, a quadratic in CBD distance, 1990 household income, child age in 1990, child sex, number of family members in 1990, household head's race in 1990, mother’s age in 1990, single parent indicator as of 1990, parent divorce indicator as of 1990, and living with father and mother indicator for 1990. Tract-level controls include log 1990 employment, 1990-2000 skill-specific Bartik shocks, and 1980 and 1970 levels of the following variables: house price index, rent index, log population, log family income, share African American, share White, share college graduate, and share with less than high school. Standard errors are clustered by census tract of residence in 1990. Figure 1: Exposure to Neighborhood Change - Census Tract Data 10 15 Densities of 1990-2000 Changes in Tract College Fraction Less than High School 0 5 College Graduates -.2 -.1 0 .1 .2 .3 10 Densities of 2000-2010 Changes in Tract College Fraction 8 Less than High School 6 High School or Less 0 2 4 College College Graduates Graduates -.2 -.1 0 .1 .2 .3 Notes: Figures show densities of childrens' exposure to change in neighborhood fraction college in the 1990s and 2000s by parents' education. These are calculated by taking the number of children ages 0-19 in each census tract and assigning parents' education based on the fraction of those 25 and older in the tract in the indicated education groups. Figure 2: Neighborhood Attributes for Movers and Stayers, Consumer Credit Panel Panel A: 2000 Tract Fraction College Destination Tract of Movers Destination Tract of Movers Original Tract of Movers Original Tract of Movers Stayers Stayers Parents Children Panel B: 2000-2007 Changes in Tract Fraction College Stayers Stayers Original Tract of Movers Original Tract of Movers Destination Tract of Movers Destination Tract of Movers Parents Children Notes: Panel A shows distributions of year 2000 college fraction of the neighborhoods of parents and their children by whether they moved by year 2017. Panel B shows distributions of 2000-2007 change in college fraction for stayers and the original tracts of movers. Red lines show distributions of college fraction measured in 2007 of movers' 2017 tracts of residence minus 2000 college fraction in year 2000 tract of residence. Figure 3: Skilled Resident Market Access for Los Angeles and Orange Counties (.0794876,.1194439] (.0655649,.0794876] (.0459721,.0655649] [-.092948,.0459721] (.0804842,.119606] (.0682347,.0804842] (.0474298,.0682347] [-.0916181,.0474298] log College RMA, 2000 log College RMA, 2010 Notes: Heat maps indicate quartiles of log College RMA and its 2000-2010 growth. The CBD is indicated in bright yellow. (.0087886,.0294669] (.0039237,.0087886] (-.0000428,.0039237] [-.0235488,-.0000428] 2000-2010 Change in log College RMA Figure 4: Change in Simulated RMA for College Workers, Los Angeles and Orange Counties (.0000867,.0006156] (-.0001036,.0000867] (-.000367,-.0001036] [-.0010563,-.000367] (.0000738,.0006158] (-.0000979,.0000738] (-.0003351,-.0000979] [-.0010487,-.0003351] Notes: The left panel depicts our main instrumental variable for the 2000-2010 analysis and the right panel shows the same object after being residualized. Residuals are taken from regressions of the instrument on region-ring fixed effects, origin tract Bartik shocks, origin tract 1990 employment, a quadratic in CBD distance and two decades of lagged demographic characteristics. CBD distance rings of 10 km and 20 km are also shown. Figure 5: Schematic Diagram of the Identification Challenge for the 2000-2010 Period Tract Fraction College Treat. Effect Tract A Actual: 1 SD RMA Growth Counterfactual: No RMA Growth Tract Fraction College A Tract A Actual: 1 SD RMA Growth Counterfactual: No RMA Growth A B A' = B' Tract B Counterfactual: 1 SD RMA Growth Actual: No RMA Growth A' B B' 1990 2000 2010 Ideal Experiment: Two identical tracts w/ diff. post-2000 experiences Treatment Effect = A - B Relative Pre-Trend = A' - B' = 0 Tract B Counterfactual: 1 SD RMA Growth Actual: No RMA Growth 1990 2000 2010 Empirical Setting: two different tracts Estimated Diff in Diff = A - B < Treatment Effect Relative Pre-Trend = A' - B' > 0 Figure 6: Dynamic Treatment Effects on Neighborhood Attributes Accounting for Migration 0.06 0.06 Sub, 97.5% Sub, 97.5% 0.03 Sub, <HS 0.03 Sub, <HS 0 City -0.03 -0.06 2000 Sub City -0.03 Sub, 2.5% 2004 2008 Sub 0 2012 Sub, 2.5% 2016 -0.06 2008 2010 2012 2014 2016 Panel A: Parents' 2000-2007 Change in Fraction College in Tract of Residence Panel B: Children's 2000-2007 Change in Fraction College in Tract of Residence Notes: Each panel graphs coefficients from separate IV regressions of the indicated outcome applying to the indicated area and education group in the indicated year on base controls and region fixed effects. Graphs exclude plots for education groups with similar results as those for everyone pooled. Indicated confidence intervals use robust standard errors. Change in fraction college is measured for the tract of residence in the indicated year measured as of 2005-2009 minus tract of residence in 2000 measured as of 2000. Figure 7: Dynamic Treatment Effects on CCP Children 60 Sub, 97.5% 40 4000 Sub, 97.5% Sub, <HS Sub, <HS 2000 Sub Sub 20 Sub, Coll+ 0 0 -20 2008 Sub, 2.5% 2010 City City 2012 2014 2016 -2000 2008 Panel A: Credit Score Sub, 2.5% 2010 2012 2016 Panel B: Sum of Credit Card Limits 0.1 0.2 Sub, 97.5% Sub, 97.5% City 0 0.1 Sub City Sub -0.1 Sub, <HS 0 Sub, <HS Sub, 2.5% -0.2 2008 2014 2010 Sub, 2.5% 2012 2014 2016 -0.1 2008 2010 2012 2014 Panel C: Probability 30 Days Past Due Panel D: Probability of Having a Mortgage Notes: See the notes to Figure 6 for an explanation of the plots. Table 9 Panel B presents detailed results for 2017. 2016 Figure 8: Dynamic Treatment Effects on CCP Parents 60 4000 City 40 2000 Sub, 97.5% 20 Sub Sub 0 0 City City -20 2001 Sub, 2.5% 2005 2009 2013 2017 -2000 2001 Panel A: Credit Score 0.1 Sub, 97.5% Sub, 2.5% 2005 2009 2013 2017 Panel B: Sum of Credit Card Limits 0.2 City 0 Sub, 97.5% 0.1 Sub Sub -0.1 City 0 Sub, 2.5% Sub, 2.5% -0.2 2001 2005 2009 2013 2017 -0.1 2001 2005 2009 2013 Panel C: Probability 30 Days Past Due Panel D: Probability of Having a Mortgage Notes: See the notes to Figure 6 for an explanation of the plots. Table 9 Panel A presents detailed results for 2017. 2017 Table A1: Coefficients from Commute Time Regressions log Distance between i and j log Distance from CBD to Residence i log Distance from CBD to Work j R-Squared 1990 0.427 (0.000) -0.068 (0.000) -0.068 (0.000) 2000 0.434 (0.000) -0.063 (0.000) -0.072 (0.000) 0.532 0.504 Notes: Regressions are of log one-way commute time on the variables listed at left and region fixed effects. Both regressions have contemporaneous commuting flow weights. Table A2: Ranking of Industry Growth Rates for College Bartik Shocks Industry Employment Growth Rate Panel A: 1990-2000 Top …. … Bottom Business and repair services Other professional and related services Communications and other public utilities Personal services Health services 0.71 0.41 0.39 0.37 0.33 Construction Agriculture, forestry, and fisheries Wholesale trade Armed Forces Mining 0.10 0.06 0.02 -0.22 -0.41 Panel B: 2000-2005 Top …. … Bottom Construction Retail Trade Finance, Insurance and Real Estate Personal Services Wholesale Trade 0.23 0.20 0.20 0.19 0.18 Business and Repair Services Communications and Public Utilities Manufacturing, Nondurable Manufacturing, Durable Armed Forces 0.07 0.07 0.03 0.01 -0.02 Notes: Since Bartik shocks exclude the origin metro area, they are slightly different across metro areas. Here we report industry-specific employment growth rates used for shocks in Bismark, ND. Table A3: Descriptive OLS Regressions About Neighborhood Change and Relationships With RMA for the 1990-2000 Period Fraction College Graduate 1990-2000 Change 1980-1990 Chg 1990 Level (1) (2) (3) (4) 1980-1990 Growth in Dependent Variable D ln Unified RMA, 1990-2000 Obs (Within) R-Squared Region-Ring Fixed Effects Base Controls -0.000340 (0.00491) -0.0483*** (0.00617) Neighborhood Quality Index 1990-2000 Change 1980-1990 Chg 1990 Level (5) (6) (7) (8) -0.213*** (0.00532) -0.00145*** (0.000267) -0.00145*** (0.000267) -0.248*** (0.00526) log Average HH Income 1990-2000 Change 1980-1990 Chg 1990 Level (9) (10) (11) (12) -0.00204*** (0.000733) -0.308*** (0.0739) -0.221*** (0.0770) -0.264*** (0.00555) -0.00419*** (0.00114) -0.00419*** (0.00114) 32,457 0.000 32,457 0.079 32,478 0.471 32,478 0.845 32,515 0.050 32,515 0.116 32,515 0.050 32,515 0.732 32,342 0.000 32,342 0.110 32,394 0.984 32,394 0.748 Yes No Yes Yes Yes Yes Yes Yes Yes No Yes Yes Yes Yes Yes Yes Yes No Yes Yes Yes Yes Yes Yes Notes: Each column is a separate regression of the variable listed at top on its change over the prior decade or change in unified RMA and the controls listed at bottom. See the notes to Table 3, which presents analogous regressions for the 2000-2010 period, for more details. Table A4: Examination of Differential Pre-1990 Trends D ln Unified RMA, 1990-2000 Obs First Stage F Region-Ring Fixed Effects Base Controls Bartik Ring Controls Fraction College Graduate Neighborhood Quality Index (1) (2) (3) (4) 0.0299*** 0.0125 3.955*** 4.653** (0.00439) (0.00830) (0.861) (1.816) 32,342 65.31 32,342 26.54 32,379 65.28 32,379 26.51 Yes Yes No Yes Yes Yes Yes Yes No Yes Yes Yes Notes: Each column reports coefficients from a separate IV regression of the 1980-1990 change in the variable listed at top on the change in log unified RMA, instrumented with the change in counterfactual log skilled RMA. See the notes to Table 3 for the base controls. The instrument is built using employment in all tracts excluding the origin tract. Regressions do not control for the change in counterfactual log unskilled RMA, as these are not separately identified from the skilled shocks for the 1990-2000 period. Table A5: First Stage Results - 1990-2000 Period Census Sample (1) (2) Change in Cntrfctl. log Skill. RMA 0.0698*** 0.0397*** (0.00440) (0.00507) Tracts in PSID Sample (3) (4) 0.0701** 0.0604 (0.0294) (0.0404) Observations (Within) R-Squared 32,515 0.065 32,515 0.118 1,519 0.906 1,519 0.914 Region-Ring FE Base Controls Bartik Ring Controls Yes Yes No Yes Yes Yes Yes Yes No Yes Yes Yes Notes: All RMA measures are expressed in standard deviation units. Table A6: Effects of RMA on Census Tract Outcomes, 1990-2000 OLS, 1st Diff. D Unified RMA IV, 1st Diff. D Unified RMA Observations First Stage F Region-Ring FE Base Controls Bartik Ring Controls Fraction College Graduate (1) (2) 0.000400 0.000246 (0.000382) (0.000394) Neighborhood Quality (3) (4) 0.288*** 0.313*** (0.0704) (0.0727) 0.0299*** (0.00459) 0.0101 (0.00878) 4.003*** (0.813) 1.269 (1.624) 32,294 253.3 32,294 62.31 32,379 251.0 32,379 61.28 Yes Yes No Yes Yes Yes Yes Yes No Yes Yes Yes Notes: Entries show OLS or IV coefficients and standard errors. All RMA measures are expressed in standard deviation units. Table A7: Impacts of RMA on Neighborhood Exposure and Migration Region-Ring Fixed Effects, Full Set of Controls and Bartik Ring Controls Educational Attainment Weight 2000-2007 Change in Fraction College in Tract of 2000 Residence None < HS College + Fraction College in 2017 Tract of Residence Fraction College in 2000 Tract of Residence None < HS College + Panel A: Parents All, 0-10 km from CBD All, 10-20 km from CBD Observations 0.00336 (0.00721) 0.00439 (0.00463) -0.00272 (0.00789) 1.11e-05 (0.00537) -0.00102 (0.00730) 0.0109* (0.00611) -0.0231 (0.0183) -0.0124 (0.0110) -0.0364 (0.0257) -0.00546 (0.0137) -0.0316 (0.0218) -0.0170 (0.0153) 9,848 1,769 3,163 9,826 1,765 3,156 Panel B: Children (Born 1985-1989) All, 0-10 km from CBD All, 10-20 km from CBD Renters, 0-10 km from CBD Owners, 0-10 km from CBD Renters, 10-20 km from CBD Owners, 10-20 km from CBD Observations 0.00157 (0.00688) 0.00472 (0.00468) -0.00452 (0.00709) 0.00181 (0.00564) -0.000871 (0.00734) 0.0109* (0.00621) -0.00676 (0.0209) 0.0165 (0.0140) -0.0214 (0.0236) 0.0199 (0.0160) 0.00133 (0.0244) 0.0209 (0.0180) 0.00444 (0.00726) 0.00184 (0.00709) 0.00640 (0.00495) 0.00345 (0.00455) -0.00210 (0.00726) -0.00445 (0.00717) 0.00351 (0.00588) 0.000671 (0.00557) 0.00225 (0.00789) -0.000831 (0.00745) 0.0128* (0.00663) 0.00942 (0.00600) -0.00297 (0.0213) -0.00617 (0.0209) 0.0192 (0.0147) 0.0145 (0.0137) -0.0114 (0.0242) -0.0219 (0.0242) 0.0278* (0.0167) 0.0137 (0.0159) -0.00832 (0.0267) 0.00121 (0.0252) 0.0150 (0.0198) 0.0233 (0.0180) 10,246 1,853 3,278 10,188 1,845 3,258 Notes: These estimates show robustness of Table 8 estimates to the inclusion Bartik ring controls. Standard errors are clustered by census tract of residence in 2000. Table A8: Consumer Credit Panel Results - Impacts of Standardized RMA Region-Ring Fixed Effects, Full Set of Controls and Bartik Ring Controls Outcome Educational Attainment Weight None Equifax Risk Score ™ < HS College + Sum of Credit Card Limits None < HS College + Any Loan 30 Days Past Due None < HS College + None Mortgage Dummy < HS College + Panel A: Parent Outcomes, 2017 All, 0-10 km from CBD All, 10-20 km from CBD Observations 12.59 (10.94) 4.048 (6.548) 28.20* (16.10) 0.239 (8.483) -1.020 (11.53) 2.924 (6.933) 3,208 (2,999) 477.1 (3,221) 3,962 (3,540) -756.6 (3,187) 2,717 (3,231) -1,861 (2,960) -0.0146 (0.0441) -0.0279 (0.0243) -0.0676 (0.0643) 0.0151 (0.0357) 0.0102 (0.0443) -0.0315 (0.0279) 0.0956 (0.0681) 0.0656 (0.0411) 0.117 (0.0904) 0.0807 (0.0503) 0.0702 (0.0673) 0.0522 (0.0459) 8,995 1,594 2,924 8,984 1,589 2,924 8,984 1,589 2,924 8,984 1,589 2,924 Panel B: Child Outcomes, 2017 All, 0-10 km from CBD All, 10-20 km from CBD Renters, 0-10 km from CBD Owners, 0-10 km from CBD Renters, 10-20 km from CBD Owners, 10-20 km from CBD Observations 2.907 (13.39) 24.15** (11.83) 12.67 (15.82) 29.72* (15.97) 10.84 (13.86) 22.20* (12.01) -154.8 (1,304) 1,890 (1,228) 651.2 (1,549) 2,272 (1,544) -1,167 (1,605) 1,712 (1,307) -0.112 (0.0683) -0.0488 (0.0436) -0.181* (0.0971) -0.0393 (0.0586) -0.0794 (0.0616) -0.0229 (0.0457) 0.0839 (0.0603) 0.108** (0.0530) 0.106 (0.0709) 0.112* (0.0632) 0.0635 (0.0610) 0.102** (0.0518) 2.213 (13.50) 2.747 (13.38) 23.66* (12.36) 25.17** (11.73) 10.59 (15.71) 12.70 (15.91) 28.32* (16.51) 31.97** (16.11) 8.170 (14.31) 10.57 (13.74) 20.40 (12.88) 23.77** (11.73) -321.8 (1,325) -153.0 (1,308) 1,800 (1,282) 1,984 (1,214) 749.9 (1,542) 627.7 (1,561) 2,337 (1,587) 2,270 (1,533) -1,930 (1,775) -1,136 (1,685) 1,239 (1,435) 2,095 (1,297) -0.0769 (0.0639) -0.112* (0.0641) -0.0297 (0.0454) -0.0698 (0.0443) -0.140 (0.0881) -0.189** (0.0909) -0.0104 (0.0615) -0.0662 (0.0604) -0.0437 (0.0613) -0.0812 (0.0594) -0.000800 (0.0472) -0.0403 (0.0442) 0.107* (0.0645) 0.0841 (0.0639) 0.120** (0.0537) 0.0920* (0.0510) 0.101 (0.0696) 0.107 (0.0707) 0.109* (0.0646) 0.109* (0.0620) 0.0987 (0.0675) 0.0628 (0.0643) 0.124** (0.0559) 0.0831 (0.0506) 9,736 1,737 3,151 9,736 1,737 3,146 9,736 1,737 3,146 9,736 1,737 3,146 Notes: These estimates show robustness of Table 9 estimates to the inclusion Bartik ring controls. Standard errors are clustered by census tract of residence in 2000. Table A9: Consumer Credit Panel Results for Children - School District Fixed Effects - Full Controls and Bartik Ring Controls Equifax Risk Score ™ DRMA, <10 km from CBD DRMA, >10 km from CBD DRMA, <10 km from CBD X School Quality Pctile DRMA, >10 km from CBD X School Quality Pctile Observations School District FE Region-Ring FE Sum of Credit Card Limits -3.369 (10.09) -4.125 (10.20) -0.0177 (13.76) 24.16** (11.88) 0.293 (0.601) 0.991** (0.446) 1.058 (12.66) 0.352 (12.84) -12.92 (14.37) -11.63 (14.68) -1,291 (963.3) -1,285 (975.7) -627.2 (1,344) 1,870 (1,232) 146.3** (68.06) 140.2*** (53.40) -1,256 (1,133) -1,274 (1,149) 151.8 (1,232) 154.8 (1,254) 9,736 Yes No 9,374 No Yes 9,374 Yes No 9,736 Yes No 9,372 No Yes 9,372 Yes No Any Loan 30 Days Past Due -0.000608 -0.102 (0.0405) (0.0685) -0.00172 -0.0565 (0.0410) (0.0436) 0.00125 (0.00268) -0.00296* (0.00180) 9,736 Yes No 9,372 No Yes Mortgage Indicator -0.0406 (0.0558) -0.0426 (0.0565) 0.108 (0.0669) 0.105 (0.0680) -0.0184 (0.0381) -0.0237 (0.0385) 0.0723 (0.0599) 0.100* (0.0515) 0.00148 (0.00273) 0.00129 (0.00188) -0.0222 (0.0458) -0.0285 (0.0464) 0.00582 (0.0484) 0.00512 (0.0492) 9,372 Yes No 9,736 Yes No 9,372 No Yes 9,372 Yes No Notes: These estimates show robustness of Table 10 estimates to the inclusion Bartik ring controls. Standard errors are clustered by census tract of residence in 2000. 6 Figure A1: Densities of Fraction College 4 < HS, 1990 < HS, 2010 2 College +, 1990 0 College +, 2010 0 .2 .4 .6 Notes: Densities are for children by parents' education and are calculated using census tract tabulations. See the notes to Figure 1 for more details. .8 Figure A2: Fixed Effects Estimates from Commute Time Model .75 1990 Commute Time Model Fixed Effects Newark Jersey City -.5 -.25 0 .25 .5 New York Washington Chicago Paterson Trenton San Francisco Boston Los Angeles Philadelphia Miami Atlanta Seattle Pittsburgh Dallas Baltimore Detroit New Orleans Norfolk Tampa Titusville St LouisHouston Cleveland Charlotte Fort Lauderdale Bristol Denver Cincinnati Orlando Greensboro Charleston Richmond Phoenix New Britain Providence Baldwin Salinas Norwalk Laredo Memphis Columbus Biloxi Mobile Portland Milwaukee New Brunswick Raleigh Albany Gary West Louisville Palm Sacramento BeachMinneapolis SanKansas Antonio San Buffalo Savannah Birmingham Elkhart CityDiego Hartford Baton Rouge Wilmington Pensacola Tucson Austin Norwich Sarasota Castleton Dayton Bibb City El Paso Harrisburg Berry Hill Atlantic City Part Brownsville Meriden Long Worcester Branch Fayetteville Lakeland Las Vegas Augusta New Haven Roanoke Huntington-Ashland,WV-OH Bridgeport Albuquerque Springfield Omaha Tallahassee Portland Chattanooga Allentown New New Haven Bedford Oklahoma Salt LakeCity Ciel Rochester Columbia Fort Smith Fort Wayne Madison Little Rock Stamford Toledo Wilmington Danville Poughkeepsie Huntsville Syracuse Akron Steubenville Gainesville Odessa Macon Spokane Beaumont Rockford Springfield Parkersburg Des Youngstown Moines Tulsa Waterbury Knoxville Fresno Davenport Jackson Montgomery Corpus Colorado Christi Springs Brockton Shreveport Wichita Springfield Charlottesville Daytona Beach Lynchburg Charleston Canton Stockton Saginaw Santa Barbara Lexington-Fayette Evansville Lowell Greenville Yak Boise City Bakersfield Lafayette Lawrence Hickory-Morganton,NC Lubbock Reno Erie Reading LansingGrand Rapids Asheville Lincoln Binghamton Appleton Flint Sioux Falls Peoria Texarkana Pittsfield Wheeling Lancaster Spartanburg,SC Fayetteville McAllen Fort Collins Florence Lake Kalamazoo Charles Albany Anderson Duluth Gadsden Champaign Tuscaloosa Topeka Modesto Lafayette Richland Decatur,AL Amarillo Hagerstown Cedar Rapids Billings SantaFe,NM Pine Bluff Monroe Manchester Eugene Abilene Tyler Cumberland Muskegon Jackson,TN Waco Dubuque Decatur Janesville Battle Johnstown Creek Sherman Green Bay Sioux City Fitchburg Scranton St Joseph Rome,GA Athens Waterloo Bloomington Victoria Lima Muncie Utica Bryan Pueblo Terre Haute Lewiston Jackson Visalia Burlington San Angelo Eau Claire Columbia Elmira Mansfield Altoona Wichita Bangor Danville,IL Fargo Lawton Goldsboro,NC Sharon, PAFalls Provo Bloomington Rochester RapidCity,SD Great Falls MesaCounty,CO Wausau Las Cruces Missoula,MT 11 13 log 1990 Population 15 17 .75 2000 Commute Time Model Fixed Effects Newark .5 Jersey City -.5 -.25 0 .25 New York Paterson Washington Trenton San Francisco Boston Chicago Los Angeles Miami Atlanta Philadelphia Dallas Seattle Baltimore Titusville Tampa Denver Houston Detroit Fort Lauderdale Pittsburgh Baldwin Las Vegas St Louis Norfolk Phoenix Portland West Palm Beach New Orleans Orlando San Diego Cleveland Providence El Paso Sacramento Biloxi Norwalk Memphis Austin Charlotte New Brunswick Charleston San Antonio Bristol Tucson Cincinnati Milwaukee Richmond Atlantic City Pensacola Salinas Sarasota Berry Hill Savannah Minneapolis Gary Wilmington Salt Lake Ciel Tallahassee Raleigh New Britain Castleton Louisville Brownsville Albany Buffalo Elkhart Greensboro Lakeland Bridgeport Augusta Kansas City Meriden Springfield Worcester Columbus Macon Harrisburg LaredoRoanoke Baton Little Rouge New Rock Long Haven Branch Colorado Springs Mobile Birmingham Huntington-Ashland,WV-OH Lowell Daytona Part Beach Knoxville Columbia New New Haven Bedford Fort Smith Albuquerque Fresno Stockton Hartford Burlington Gainesville Lawrence Portland Greenville Madison Norwich Spokane Brockton Stamford Lexington-Fayette Fort Fayetteville Wayne Allentown Charleston Omaha Richland Jackson Waterbury Grand Rapids City Beaumont Reno Rochester Spartanburg,SC Dayton Danville Oklahoma Reading Evansville Saginaw Corpus Christi Bibb CityMcAllen Chattanooga TylerFitchburg Charlottesville Fayetteville Fort Collins Yak Toledo Springfield Lansing Modesto Bakersfield Parkersburg Montgomery Des Moines Flint Syracuse Lincoln Santa Barbara Tulsa Boise City Lancaster Rockford Binghamton Manchester Akron Visalia Poughkeepsie Youngstown Erie Shreveport Wichita Rome,GA Davenport Hickory-Morganton,NC Lafayette Wilmington Sioux Falls Albany Waco Wheeling Appleton Utica Lynchburg Kalamazoo Asheville Huntsville Muskegon Eugene Lafayette Muncie Cedar Rapids Hagerstown Pine Bluff Elmira Billings Canton Athens Bryan Odessa Springfield Duluth Pueblo Texarkana Sherman Lewiston Pittsfield Decatur,AL San Angelo Eau Claire SantaFe,NM Steubenville Green Bay Janesville Lubbock Abilene Topeka Victoria Jackson,TN MesaCounty,CO Jackson Peoria Gadsden Battle Creek Monroe Sioux City Scranton Johnstown Tuscaloosa Amarillo Florence Fargo Provo Waterloo Terre Haute Bloomington Anderson Cumberland Wichita Lake Charles StSharon, Wausau Joseph Dubuque Bangor Champaign Mansfield PA Falls Lima Columbia Altoona Bloomington Lawton Rochester Danville,IL Great Falls RapidCity,SD Decatur Goldsboro,NC Las Cruces Missoula,MT 11 13 log 2000 Population Notes: Fixed effects are normalized to be mean 0 across regions. 15 17 Figure A3: Estimates of ke by Metro Area .12 1990 Kappa X Epsilon 0 .04 .08 Eau Claire Janesville Fargo Appleton Steubenville UticaLancaster Pittsfield Waterbury Bangor Scranton New Bedford Meriden Sherman Binghamton Fayetteville Sharon, PA Hickory-Morganton,NC Anderson Norwich Mansfield Erie Great Falls Elkhart Elmira Altoona Stamford Waterloo New Britain Brockton MadisonSpringfield NewBristol Haven Wausau Hartford Lewiston Danville Fayetteville Green Parkersburg LawrenceAtlantic City Wichita FallsBay Florence Rockford Canton Cedar Rapids New Haven Gary Beaumont Allentown Milwaukee Champaign Yak SiouxWilmington City Harrisburg Texarkana Biloxi Peoria Bridgeport Minneapolis Richland Poughkeepsie McAllen Titusville MesaCounty,CO Buffalo Rome,GA Las Cruces Youngstown Saginaw Salinas Lafayette Norwalk Worcester Roanoke Davenport Des Moines Fort Smith Cumberland Rochester Burlington Daytona BeachLong Branch Bloomington Duluth Terre Haute Goldsboro,NC New Brunswick Abilene Provo Providence Corpus Christi Akron Columbia Chattanooga Kalamazoo Wichita Albany Castleton Bryan Billings Huntington-Ashland,WV-OH Part St Louis Odessa Fitchburg San Angelo Missoula,MT Visalia Waco Johnstown Brownsville Sioux Falls Montgomery Fort Collins Dayton Macon Lincoln Wheeling Reno Lowell Reading Monroe Grand Rapids Cincinnati Pittsburgh Lansing Lawton Bloomington Lakeland Santa Barbara Greensboro Gadsden Asheville Modesto Dubuque Boise City Sarasota Bibb City Lynchburg Salt Lake Ciel Cleveland Boston Toledo RapidCity,SD Norfolk Hagerstown Syracuse Stockton Fort Wayne Pine Bluff Pensacola Louisville San Diego Tyler Manchester Omaha Columbus Danville,IL Charleston Gainesville Kansas City Jackson,TN Spokane Little Rock Springfield Eugene Portland Seattle Portland Jackson Baltimore Evansville Colorado Springs Raleigh Trenton El Fresno Paso Spartanburg,SC Athens Charlotte Richmond Lima Shreveport Orlando Denver Tampa Tuscaloosa Muncie Wilmington Oklahoma City New Orleans Springfield West Palm Beach Sacramento Atlanta Philadelphia Victoria Flint Charlottesville Augusta SanPaterson Francisco Fort Rochester Columbia AmarilloHuntsville Berry Hill Lauderdale Dallas Knoxville Greenville Laredo Baldwin Baton Rouge Birmingham Memphis DetroitWashington Decatur Austin Chicago Battle Creek Albuquerque StAlbany Joseph Muskegon Lexington-Fayette Las Vegas Topeka Tallahassee Savannah Lafayette Houston Los Angeles Charleston Tucson Lake Charles Miami Phoenix SantaFe,NM Bakersfield San Antonio Pueblo Lubbock Jackson Decatur,AL New York Newark MobileTulsa Jersey City 11 13 log 1990 Population 15 17 .12 2000 Kappa X Epsilon Dubuque Champaign 0 .04 .08 Sioux City Sharon, PA Erie Lima Janesville Wausau Florence RapidCity,SD BloomingtonFayetteville Appleton Lawton Steubenville Norwich Mansfield Bangor JohnstownAsheville Altoona Lancaster Billings Fargo EauLake Claire Missoula,MT Columbia MesaCounty,CO Charles Yak Pittsfield Danville,IL Canton Great Falls Anderson Goldsboro,NC Hickory-Morganton,NC Jackson Wilmington Wichita Fitchburg FallsPortland Hagerstown Waterbury Muskegon Elkhart Harrisburg Richland Fort Collins Lynchburg Des Moines Madison Duluth Elmira Fayetteville Scranton Decatur,AL Manchester Sherman Sioux Falls Parkersburg Brockton Biloxi Provo Davenport Peoria Texarkana Fort Smith Grand Rapids Battle Creek Raleigh Cedar Rapids Boise City Allentown Abilene Lewiston Meriden Wheeling Las Cruces Gary Stamford Atlantic City Worcester Saginaw Rockford Minneapolis Utica St Waterloo Joseph Reading Santa Barbara Akron Danville San Angelo Green Bay Dayton Evansville Milwaukee Salinas Toledo Charleston Bristol Huntington-Ashland,WV-OH Part Lansing Springfield Kalamazoo Cumberland Poughkeepsie Binghamton Lafayette Charlottesville Chattanooga Reno Muncie Lowell Youngstown Providence SantaFe,NM Greensboro Hartford Burlington Eugene Odessa New Britain Macon Rochester Lexington-Fayette Huntsville Titusville McAllen Berry Hill New Syracuse Albany Bedford Pine Bluff Castleton New Brunswick Roanoke Corpus Christi Springfield Monroe Knoxville Bloomington Springfield Portland Norwalk New HavenColumbus Tyler Lincoln Daytona Beach Montgomery Pensacola Colorado Springs Waco Pittsburgh Spartanburg,SC Decatur Wichita Cincinnati Gadsden Charleston Little Bridgeport Rock Augusta Visalia LongSalt Branch Lawrence Norfolk Louisville Kansas AmarilloShreveport Charlotte Wilmington Topeka Greenville Spokane Boston Lake Ciel Albuquerque St City Louis Seattle NewColumbia Haven Modesto Buffalo Laredo Tucson Tallahassee Fort Wayne Beaumont Mobile Austin Richmond Lafayette El Rouge Paso Omaha Tulsa Oklahoma City San Francisco Brownsville Baton Sarasota Orlando Cleveland Terre Haute BibbFlint City Denver Savannah Sacramento Detroit Bakersfield Birmingham Baldwin Trenton Tuscaloosa Washington Atlanta Paterson Rochester Dallas Rome,GA Lakeland Fresno Bryan Philadelphia Stockton Victoria PuebloGainesville Jackson,TN Lubbock Jackson Tampa San Phoenix Diego Baltimore Houston Chicago San Antonio New Orleans Miami Fort Lauderdale West Palm Beach Memphis Los Angeles New York Albany Athens Newark Jersey City Las Vegas 11 13 log 2000 Population 15 17 Notes: Plotted points are coefficients on one-way commute time from regressions of log commute flow on origin and destination fixed effects and the one-way commute time in minutes, weighted by commute flow.