Three essays on the social context of wealth accumulation and racial wealth inequality in the U.S.

What explains the long-term persistence of Black-White racial inequality in the United States? This dissertation seeks to answer this question by exploring how social context perpetuates racial wealth inequality. Black-White household wealth differentials consistently dwarf other outcome gaps, household wealth is a key factor in determining the social outcomes of individuals, and social context, broadly defined, consistently exerts a strong influence on social outcomes. This dissertation integrates these three insights into an examination of the Black-White wealth gap. The first chapter uses the Panel Study of Income Dynamics data set to explore how the total household wealth of the extended family network correlates with the household wealth accumulation of individuals, and tests for racial differences therein. Quantile regression analysis finds important racial differences in the accumulation of financial assets at the top of the wealth distribution. The second chapter combines household-level data from the Survey of Income and Program Participation with city-level data from various sources in order to explore how neighborhood social characteristics influence the ability of households to accumulate wealth. The social and economic characteristics of neighborhoods are found to exacerbate wealth inequality. The third chapter proposes a two-tiered model of wealth accumulation based on the empirical regularities established in the first two chapters. The proposed model of wealth accumulation combines threshold effects with network externalities to explain how the historical social isolation of Blacks in the United States has perpetuated the Black-White wealth gap and perpetuated Black-White differentials in social outcomes.

THREE ESSAYS ON THE SOCIAL CONTEXT OF WEALTH ACCUMULATION AND RACIAL WEALTH INEQUALITY IN THE U.S. by Johan Andres Uribe A dissertation submitted to the faculty of The University of Utah in partial fulfillment of the requirements for the degree of Doctor of Philosophy Department of Economics The University of Utah August 2016 Copyright © Johan Andres Uribe 2016 All Rights Reserved The University of Utah Graduate School STATEMENT OF DISSERTATION APPROVAL The dissertation of Johan Andres Uribe has been approved by the following supervisory committee members: Thomas Maloney , Chair 4/28/2016 Date Approved Ella Myers , Member 5/18/2016 Date Approved Gunseli Berik , Member 5/17/2016 Date Approved Eric Sjöberg , Member 4/28/2016 Date Approved Norman Waitzman , Member 5/17/2016 Date Approved and by Thomas Maloney , Chair/Dean of the Department/College/School of Economics and by David B. Kieda, Dean of The Graduate School. ABSTRACT What explains the long-term persistence of Black-White racial inequality in the United States? This dissertation seeks to answer this question by exploring how social context perpetuates racial wealth inequality. Black-White household wealth differentials consistently dwarf other outcome gaps, household wealth is a key factor in determining the social outcomes of individuals, and social context, broadly defined, consistently exerts a strong influence on social outcomes. This dissertation integrates these three insights into an examination of the Black-White wealth gap. The first chapter uses the Panel Study of Income Dynamics data set to explore how the total household wealth of the extended family network correlates with the household wealth accumulation of individuals, and tests for racial differences therein. Quantile regression analysis finds important racial differences in the accumulation of financial assets at the top of the wealth distribution. The second chapter combines household-level data from the Survey of Income and Program Participation with city-level data from various sources in order to explore how neighborhood social characteristics influence the ability of households to accumulate wealth. The social and economic characteristics of neighborhoods are found to exacerbate wealth inequality. The third chapter proposes a two-tiered model of wealth accumulation based on the empirical regularities established in the first two chapters. The proposed model of wealth accumulation combines threshold effects with network iv externalities to explain how the historical social isolation of Blacks in the United States has perpetuated the Black-White wealth gap and perpetuated Black-White differentials in social outcomes. TABLE OF CONTENTS ABSTRACT ....................................................................................................................... iii LIST OF TABLES ............................................................................................................ vii LIST OF FIGURES ........................................................................................................... ix Chapters 1. WHAT EVER HAPPENED TO THE RISING TIDE? RACIAL WEALTH INEQUALITY AND EXTENDED FAMILY NETWORKS........................................1 1.1 Abstract ..............................................................................................................1 1.2 Introduction ........................................................................................................2 1.3 Literature Review...............................................................................................5 1.4 Data and Network Structure...............................................................................8 1.5 Methodology ....................................................................................................23 1.6 Empirical Results .............................................................................................30 1.7 Discussion ........................................................................................................38 2. WHEREIN LIES THE FERTILE GROUND FOR RACIAL EQUALITY? AN ANALYSIS OF THE SPATIAL RELATIONSHIP AMONG INEQUALITY, SEGREGATION, AND RACIAL WEALTH DISPARITIES ....................................44 2.1 Introduction ......................................................................................................44 2.2 Data ..................................................................................................................46 2.3 Methods and Theory ........................................................................................56 2.4 Model and Specification ..................................................................................60 2.5 Limitations .......................................................................................................63 2.6 Results ..............................................................................................................63 2.7 Discussion ........................................................................................................66 3. WHAT'S WEALTH GOT TO DO WITH IT? RACE, RACIAL INEQUALITY, AND WEALTH IN THE UNITED STATES ........................................................................70 3.1 Introduction ......................................................................................................70 3.2 Why Race, Why Black? The Relevance of Race in the United States ............71 vi 3.3 Racial Inequality ..............................................................................................75 3.4 Two-Tiered Model of Racial Wealth Inequality ..............................................80 3.3 Discussion ........................................................................................................87 Appendices A. IMPUTATION MODEL...............................................................................................91 B. SPECIFICATION TESTS, REGRESSION, AND DESCRIPTIVE TABLES ............93 C. ADDITIONAL TABLES FROM CHAPTER 2 ...........................................................99 REFERENCES ................................................................................................................110 LIST OF TABLES Tables 1.1 Network Composition ................................................................................................. 11 1.2 Generational Makeup of Sample Households............................................................. 11 1.3 Network Characteristics in 2003 ................................................................................. 11 1.4 Network Wealth Tabulations for 1989, 1999, and 2013 ............................................. 12 1.5 Household Descriptive Statistics by Race in 2013 ..................................................... 14 1.6 Age at Household Formation ...................................................................................... 15 2.1 Household Descriptives - 2001 ................................................................................... 46 2.2 Education .................................................................................................................... 47 2.3 Marital Status .............................................................................................................. 47 2.4 Median Household Wealth by CBSA: 2001-2003...................................................... 49 2.5 Metropolitan Area Descriptives .................................................................................. 56 2.6 Effects of Metropolitan Characteristics on Household Wealth Accumulation ........... 62 B.1 Network Wealth Tabulations for Black Households by Year .................................... 94 B.2 Network Wealth Tabulations for Non-Black Households by Year ............................ 95 C.1 Pooled Regressions on Metros with HH Count 100 Using Segregation of Affluence........................................................................................................................ 100 viii C.2 Random Effects Regressions on Large Metros Using Income Segregation ............ 102 C.3 Pooled Regressions Excluding New York City ....................................................... 104 C.4 Pooled Regressions Excluding Los Angeles, CA .................................................... 106 C.5 Pooled Regressions Excluding Chicago, IL ............................................................. 108 LIST OF FIGURES Figures 1.1 Histogram of Household Wealth in 2013 ................................................................... 16 1.2 Scatterplots of Household Wealth with Extended Family Wealth and Age ............... 18 1.3 Box and Whisker Plot of Household Wealth .............................................................. 19 1.4 Median Household Wealth 1989-2013 ....................................................................... 21 1.5 Empirical Wealth Deciles by Race for 1989 and 2013 ............................................... 22 1.6 CRE Extended Family Wealth Full Specification ...................................................... 32 1.7 CRE Second-Generation Financial Wealth - Full Specification ................................ 36 1.8 Second-Generation Marginal Effects of Sibling Wealth and Offspring Wealth ........ 37 1.9 Marginal Effects of Extended Family Branches on Third-Generation Wealth .......... 39 2.1 Histogram of Change in Household Wealth Between 2001 and 2003 ....................... 50 2.2 Kernel of Change in Household Wealth Between 2001 and 2003 ............................. 51 2.3 Geographic Distribution of Household Wealth .......................................................... 51 B.1 Baseline CRE Model of Household Wealth .............................................................. 96 B.2 Baseline KFE Model of Household Wealth for Comparison .................................... 97 B.3 Marginal Effect on Household Nonfinancial Wealth for Third-Generation Sample…………………………………………………………………………………....98 CHAPTER 1 WHAT EVER HAPPENED TO THE RISING TIDE? RACIAL WEALTH INEQUALITY AND EXTENDED FAMILY NETWORKS 1 1.1 Abstract As wages stagnate and education and healthcare costs increase, household wealth arguably becomes more important than ever in the determination of social outcomes. However, the rising tide of asset values has left Black households behind. The mechanisms underlying the long-term persistence of racial wealth inequality are inadequately understood. Wealth clearly builds on itself over time, has large, tangible effects on the physical and social environment of communities, and represents the largest and most persistent difference between Whites and Blacks in the United States. This study investigates how differences in the structure and resources of extended family networks contribute to the perpetuation of racial wealth inequality. I organize a subsample of the Panel Study of Income Dynamics into extended family networks and use quantile regression to analyze the wealth linkages within extended family networks. 1 The PSID is based upon work supported by the National Science Foundation under Grant No. 9515005 and the National Institute on Aging under Grant Nos. R01AG6671 and Y1AG9188. Any opinions, conclusions, or recommendations expressed in this material are those of the author and do not necessarily reflect the views of the National Science Foundation, or the National Institute on Aging. 2 Upon decomposing household wealth into financial and nonfinancial components, I find that Black and non-Black households experience different accumulation dynamics. Second-generation non-Black households in the top two wealth deciles accumulate financial assets in tandem with their extended family networks, generating a virtuous cycle of wealth accumulation with extended family networks. However, Black households in the top wealth deciles do not benefit from a rising tide of family wealth, leaving them at a relative disadvantage. 1.2 Introduction Over the past 20 years, social scientists have built a body of empirical work investigating the role of household wealth in the determination of social outcomes. Researchers have documented how parent-child correlations in net worth, household ownership, and portfolio allocation play an important role in economic inequality (Chiteji & Hamilton, 2002; Chiteji & Stafford, 1999; Hilber & Liu, 2007; Karagiannaki, 2012). Researchers have explored causal mechanisms through which household wealth impacts health outcomes and influences both access to communities and the evolution of those communities (Meer, Miller, & Rosen, 2003; Oliver & Shapiro, 1997; Sugrue, 2005). Recently, researchers have delineated the pathways through which neighborhoods exert significant causal influence on educational and labor market outcomes (Chetty & Hendren, 2015; Chetty, Hendren, & Katz, 2015). In other words, researchers have made significant headway in understanding racial inequality, and the connection between wealth and social context plays a central role. However, a coherent explanation of the long-term evolution of racial wealth inequality has yet to take shape. Without addressing this gap in the literature, 3 it is problematic to draw clear empirical connections from the nexus of poverty to the panoply of adverse social outcomes, or to use the extant body of research to address public policy concerns related to racialized poverty. This study addresses this gap by elucidating the connections between extended family networks and household wealth accumulation. This contribution is important because it increases our understanding of the social forces involved in the long-term evolution of racial wealth inequality and opens up the possibility of addressing barriers to household wealth accumulation among Black communities. The central finding of this paper is that only second-generation non-Black households in the top two deciles of the household wealth distribution experience positive extended family effects on financial wealth accumulation. Relatively wealthy Black households do not benefit from the rising tide of family wealth. In other words, it is the privileges associated with belonging to the upper echelons of White society that correlate with self-reinforcing wealth accumulation dynamics. The novel use of the Panel Study of Income Dynamics, PSID, and the application of conditional-quantile regression analysis, represent the technical innovations that underlie this study. Taking advantage of the multigenerational structure of the PSID in conjunction with its rich wealth data, this study organizes a subsample of the PSID into extended family networks. This study defines extended family networks as the network of households whose heads are related to each other either through kinship, marriage, or nonfamily cohabitation.2 The quantile panel regression models the nonlinear marginal effects of extended family resources on household wealth accumulation across the conditional distribution of household wealth. 2 Informal social ties include people who live together in household units and share resources within the household, i.e., family friends who live together, but not renters. 4 The PSID is unique among datasets due to its longevity, panel structure, breadth of questions, and new-household-following rules. However, these advantages come with large drawbacks. The longevity of the PSID, running for nearly 50 years now, creates the possibility of substantial self-selection bias despite the low rate of attrition. Although PSID cohorts are representative of the overall population if the appropriate sample weights are used, the attrition bias issue is not fully obviated (Ziliak & Kniesner, 1998). It is still possible and likely for intergenerational correlations to suffer from attrition bias even if the sample remains representative, especially in a three-generation sample (Fitzgerald, Gottschalk, & Moffitt, 1998). For example, the individual households in my sample may be roughly representative of the United States in 2013 once the proper weights are applied, but the structure of the extended family networks is most likely not representative of the structure of extended family networks in the United States due to intergenerational attrition. Thus, the extended network effects cannot be perfectly generalized to the overall population of the United States, and self-selection may create bias in our estimates. There is no clear way to deal with the intergenerational attrition issues at the moment. Despite these serious limitations, the PSID represents a unique opportunity to answer important questions that would otherwise go unaddressed. Although the results should be interpreted with caution, the contributions of this study are important enough to warrant this investigation. Racial inequality is associated with educational inequality, differentials in health outcomes, income inequality, employment patterns, residential status, political power, and other outcomes (Oliver & Shapiro, 1997; Shea, Miles, & Hayward, 1996; Wilson, 2012). In other words, a coherent investigation of racial inequality must consider the social, 5 economic, and political dimensions of inequality. The primary advantage of focusing on household wealth accumulation is that it touches each of these different facets of racial inequality. Over the past 20 years, our understanding of what constitutes the determining factors of class or socioeconomic status has evolved, and it now incorporates wealth as a central factor (Conley, 2010; Oliver & Shapiro, 1997). Wealth exerts strong effects on social outcomes independent of the traditional factors associated with socioeconomic status. Family wealth, including home ownership, not only influences a child's potential for educational success, but also his or her health outcomes and future economic prospects (Karagiannaki, 2012; Shea, Miles, & Hayward, 1996). The paper proceeds as follows. A literature review will summarize the existing scholarship related to Black-White wealth inequality. The next section will discuss the data and how I organize the extended family networks. The fourth section describes the model. The fifth section presents the empirical results and discusses limitations. The paper concludes with a discussion of what the evidence tells us, how this research informs the ongoing public policy debates, and future avenues for investigation. 1.3 Literature Review Since the mid-90s, social scientists have increasingly looked to wealth as a central driver of racial inequality (Conley, 2010; Oliver & Shapiro, 1997). Part of the intrigue surrounding the racial wealth gap stems from the persistence of the wealth gap despite a significant narrowing of the income gap during the 1960s and 70s (Gittleman & Wolff, 2004; Mckernan, Ratcliffe, Steuerle, & Zhang, 2014). In this vein, scholars have documented the historical barriers, legal and otherwise, that specifically prevented Blacks 6 from accumulating wealth (Conley, 2010; Wilson, 2012). However, scholarship investigating the connections between the historical oppression of Blacks and the persistence of the Black-White wealth gap remains sparse, mostly limited to discussions of inheritances and perverse preferences (Barsky, Bound, Charles, & Lupton, 2002; Caballero, 1991; Hubbard, Skinner, & Zeldes, 1995). Relatively recent investigations have begun to empirically connect the social conditions of Black communities to the long-term persistence of wealth inequality, but this literature remains nascent (Chiteji & Hamilton, 2002; Conley, 2010; Oliver & Shapiro, 1997; Sampson, 2009). Research on sources of the Black-White gap in wealth is sparse due to several factors. Part of the problem with the study of wealth is the lack of data. Reliable national estimates of median wealth only started in 1983. With the exception of a national survey in 1963, we know little of the national trends in Black-White wealth inequality before then. A few historical studies make use of property and probate data from the 19th and early 20th centuries, but these leave out more than half a century of wealth trends (Higgs, 1982; Margo, 1984; Miller, 2011). Additionally, the prevailing theoretical models of household wealth accumulation reduce the process of wealth accumulation to income, risk aversion, rate of time preference, and transfers (Caballero, 1991; Modigliani, 1986). The theoretical literature consists of models derived from Modigliani's life cycle hypothesis with adjustments to incorporate uncertainty, bequests, transfers, intergenerational altruism and reciprocity, multiple overlapping generations, and social insurance (Arrondel & Masson, 2001; Caballero, 1991; Guiso, Jappelli, & Terlizzese, 1992; Hubbard et al., 1995; Modigliani, 1986). In all these models, a household's ability to accumulate wealth is connected to social context by only two mechanisms: preference formation during 7 childhood and transfers from parents. As such, these models' ability to inform the debate surrounding racial wealth inequality is limited to hypotheses about how savings rates are related to childhood environment and parental transfers. However, savings rates between Blacks and Whites are identical once income is controlled for, but the differences in transfers and portfolio allocation are large (Gittleman & Wolff, 2004). The differences in wealth transfers between Blacks and Whites are historically contingent; Blacks were systematically denied the ability to accumulate wealth for most of the 20th century, and hence have less to transfer (Margo, 1984; Oliver & Shapiro, 1997; Wilson, 2012). Differences in portfolio allocation, however, are an enigma. Blacks carry relatively more debt and are more heavily invested in real estate than comparable Whites, who prefer financial assets (Gittleman & Wolff, 2004). Although real estate gains were strong in the late 1990s and early 2000s, this investment ultimately did not work out well for Blacks (Rugh & Massey, 2010). There exists a strong parent-child correlation in portfolio allocation, but this explains only a portion of the enigma (Charles & Hurst, 2003; Chiteji & Stafford, 1999). A few studies have explored the idea that social relationships, broadly construed, impact wealth accumulation through home ownership, but they do not generally make the connection to portfolio allocation or wealth accumulation (Collins & Margo, 2011; Hilber & Liu, 2007; Sampson & Sharkey, 2008). Social scientists have begun to scratch at the nexus of social connections, familial transfers, portfolio allocation, and wealth accumulation. This study fills a gap in the literature by extending the analysis of wealth accumulation to incorporate social connections beyond the traditional parent-child relationship. 8 1.4 Data and Network Structure The Panel Study of Income Dynamics, PSID, began a survey of approximately 5,000 families in 1968, collecting individual- and household-level economic, demographic, and health information on a yearly basis. Three key features set the PSID apart from other longitudinal data sets. First, the scope of the PSID makes it a unique data set. Incorporating nearly 45 years' worth of detailed longitudinal data makes the PSID one of the world's longest running panel data sets, which enables researchers to investigate household dynamics across the entire life cycle, from early adulthood to retirement. Second, the relational structure of the PSID makes it valuable for researchers interested in family networks and intergenerational studies. The third key attribute of the PSID is the ability to measure and track household wealth. The PSID began to collect detailed wealth information for all member households starting in 1984 at 5-year intervals, and then at 2-year intervals starting in 1999. Thus, we have high-quality, detailed measures of wealth for most households for intermittent years between 1989 and 2013 (Bosworth & Smart, 2009).3 Household wealth is defined as the sum of the following: net value of primary home, net value of other real estate, net value of vehicles, net value of farm or businesses, net value of stocks and other financial instruments such as bonds, value of cash accounts, minus the value of any debts, which includes credit cards, student loans, and other debts.4, 5 The PSID defines a household as the set of persons living together in a residence who are either related by blood or marriage, plus nonrelatives who share income and resources with the rest of the household. A tenant renting a room in a house would not be 3 I restrict the years to 1989-2013 in order to focus on the most recent generations in the PSID. 4 All wealth components are self-reported values. 5 Monetary values used in this study are adjusted for inflation using 2000 as base year. 9 considered a member of the household, but a nonrelated family friend who permanently lives with and shares resources with the rest of the family would be considered a member of the household. Researchers originally set out to investigate the causes and correlates of long-term and intergenerational poverty, to track an ever-expanding web of related households. Each time that an adult left a household, he or she was tracked and the new household was added to the data set, to the extent that it was possible to track the adult and convince him or her to participate. The data set quickly expanded as children, siblings, divorced spouses,6 and nonrelated household members of the original 5,000 households left to establish their own respective households. Over the course of 45 years, the PSID has expanded to the point where we can identify and track 2,473 unique extended family networks that are still participating in the PSID. Each network incorporates up to three generations' worth of households, although I analyze only the two latest generations in this study.7 The networks are made up of the original 5,000 households, referred to as generation one, the households formed by their offspring, referred to as generation two, and the households formed by the offspring of the second generation, referred to as generation three. One extended family network could thus potentially include grandparents, parents, uncles/aunts, and adult children who have established their own households. On average, each network contains 2.43 second-generation households and 1.81 third-generation households, as shown in Table 1.1. This network composition is to be expected due to the age structure of the networks: third-generation households are still in the process of leaving their parents' households and 6 Divorced spouses of the sample are followed only if they were raising the children of sample members. 7 I omit the first-generation households, the original 5,000 households, because they have increasingly passed away, which leads to distinct bias in the selection of households in which members lived to an above-average age. 10 establishing their own. Extended family networks range from 1 household to 30 households with a median of 3 and a mean of 4 households per network. Black networks have a mean of 4.7 households per network, but non-Black households contain only 3.64 on average. This network structure makes it possible to explicitly model extended family structure as well as multigenerational dynamics. I restricted my analytical sample to the two most recent generations of PSID households, generations two and three. Generation two includes 5,585 unique households. These are mostly the adult children of the original 5,000 families that founded their own households and agreed to participate in the survey. Generation three includes 4,151 unique households. Tables 1.2, 1.3, and 1.4 provide raw, unweighted summaries of the network characteristics of my sample.8 Table 1.2 shows that the number of third-generation households increased almost every year, as we would expect since young families are constantly leaving their parents to start their own households. Table 1.3 shows the distribution of siblings, aunts/uncles, cousins, and offspring for our sample households in 2013. Looking at the first column of Table 1.3, we see that 959 second-generation households had zero siblings in their network. From Table 1.2, we see that 2,623 second-generation households participated in the survey in 2013. Thus, slightly less than one-third of all second-generation households had zero siblings in their network who participated in the survey in 2013.1 I lumped ambiguously related network members into the category of cousins. Because of the definition of households in the PSID, there arose instances of ambiguously related members, leading to the large "cousin" counts in Table 1.3 as well as the zero count in the cousins' cell. Somewhat 8 All tables were created using the stargazer package in R (Hlavac, 2014). 11 Table 1.1 Network Composition Extended Family Networks Generational Composition* Households per Network** Total Mean Total Mean Total Black 803 Second 2.439 5585 Black 4.741 3807 Non- Black 1670 Third 1.814 4151 non- Black 3.64 6084 Total 2473 Table 1.2 Generational Makeup of Sample Households Wealth observed in: 1989 1994 1999 2001 2003 2005 2007 2009 2011 2013 2nd Gen 3569 4077 2971 2979 2983 2898 2843 2794 2703 2623 3rd Gen 1010 1033 896 1058 1266 1460 1741 1976 2170 2339 Table 1.3 Network Characteristics in 2013 Siblings Siblings Cousins Offspring Uncles 2nd Gen 3rd Gen 3rd Gen 2nd Gen 3rd Gen 0 959 806 1534 1584 1 469 851 274 462 182 2 415 468 408 413 178 3 306 154 299 161 148 4 158 40 273 43 82 5 86 6 234 7 35 6 90 10 190 1 52 7+ 140 4 661 2 78 12 Table 1.4 Network Wealth Tabulations for 1989, 1999, and 2013 Non-Black Households Black Households 1989 Statistic N Mean Median N Mean Median Age of Head 2,767 34.3 33 1,812 33.4 32 HH Wealth 2,767 146,682 70,407 1,812 33,385 3,034 Network Wealth 2,767 252,804 87,072 1,812 70,820 4,197 Parent Wealth 32 231,969 141,918 26 106,001 51,937 Sibling Wealth 1,381 283,992 144,425 1,019 81,712 15,276 Uncle Wealth 8 220,104 317,175 6 43,200 7,997 Cousin Wealth 673 412,625 251,071 337 118,757 52,354 Offspring Wealth 1,031 19,819 0 447 4,531 0 1999 Statistic N Mean Median N Mean Median Age of Head 2,553 39.7 40 1,314 39.9 40 HH Wealth 2,553 175,707 46,512 1,314 43,853 9,302 Network Wealth 2,553 297,760 54,264 1,314 81,087 3,623 Parent Wealth 275 221,298 89,923 94 81,134 33,646 Sibling Wealth 1,431 284,940 75,763 709 87,202 14,677 Uncle Wealth 122 472,404 93,024 41 261,984 46,615 Cousin Wealth 622 347,183 124,719 274 81,116 46,615 Offspring Wealth 1,038 17,334 0 403 10,243 0 2013 Statistic N Mean Median N Mean Median Age of Head 3,035 44.5 45 1,927 44 47 HH Wealth 3,035 188,542 37,423 1,927 31,557 3,700 Network Wealth 3,035 664,310 176,059 1,927 114,225 22,200 Parent Wealth 892 436,324 127,122 417 59,820 20,165 Sibling Wealth 1,995 284,526 58,460 1,202 69,806 7,622 Uncle Wealth 513 446,782 134,310 242 75,517 0 Cousin Wealth 1,424 531,619 144,053 915 91,690 25,462 Offspring Wealth 788 92,799 20,831 306 29,696 1,480 Note: In 2013, 513 non-Black households and 242 Black households have uncles present in their extended family networks. Network wealth is the sum of the household wealth of all extended family households present in the data set in the given year, but exclusive of own-household wealth. Thus, the mean of network wealth for any year represents the total amount of wealth held by the extended family network of the average household, exclusive of their own wealth holdings. All monetary values are adjusted to constant year 2000 dollars. The full set of descriptive statistics, including standard deviations and min/max, for both Black and non-Black households, can be found in Appendix B. 13 more than two-thirds of all third-generation households had one or more siblings present in 2013. It is important to point out that year-to-year participation in the PSID varies considerably. Thus, Table 1.3 is a snapshot of family members who participated in the PSID in 2013. The important information from Table 1.3 is that the PSID contains an impressive amount of network structure, and thus lends itself to network analysis. Table 1.4 describes the location of wealth within Black extended family networks across three years: 1989, 1999, and 2013. The top right panel of Table 1.4 shows that in 1989, the median Black household held $3,034 worth of assets, belonged to a network with median total assets worth $4,197, and that the vast majority of that network wealth was located in the hands of the siblings. In 2013, the median Black household held $3,700 worth of assets, belonged to a much larger network, which held a total of $22,200 worth of assets at the median that was mostly held by parents, siblings, and cousins. Of the 1,927 Black households in 2013, 417 had parents who participated in the PSID that year, 1,202 of them had at least one sibling, and so on. Table 1.4 also describes the network location of wealth within non-Black extended family networks for the same 3 years. In 2013, the median non-Black household held $37,423 worth and assets and belonged to a network that held $176,000 worth of assets. These assets were distributed relatively evenly among parents, siblings, uncles, and cousins, with offspring predictably holding few assets. Table 1.5 provides a list and descriptive statistics for 2013 of the control variables used in this study. The top panel describes the Black households. The bottom pane describes the non-Black households. Education, gender, and race are not included among the descriptives. 14 Table 1.5 Household Descriptive Statistics by Race in 2013 Non-Black Households Statistic N Mean St. Dev. Min Median Max Age of Head 3,035 44.5 14.6 17 45 90 Household Wealth 3,035 188,542 684,595 -546,120 37,423 23,487,600 Household income 3,035 66,163 105,737 -34,040 48,100 2,453,840 Children 3,035 0.6 1.0 0.0 0.0 8 Divorced 18% Married 54% Black Households Statistic N Mean St. Dev. Min Median Max Age of Head 1,927 44 13.9 18 47 89 Household Wealth 1,927 31,557 103,973 -509,490 3,700 1,662,040 Household income 1,927 31,256 30,324 -518 22,671 377,252 Children 1,927 0.7 1.1 0.0 0.0 9 Divorced 26% Married 29% as control variables because they are collinear with the fixed effects in the majority of cases. Households are coded as female-headed only if there is no adult male living within the household at the time of the interview, or if the adult male is disabled. Race is incorporated into the model through the use of race-dummy interactions with both income and extended family wealth. Black and non-Black households in my sample are significantly different along nearly every dimension: wealth, income, schooling, marriage, and divorce. The difference in marriage and divorce rates is particularly striking. Only 29% of Black households in my sample were currently married in 2013, and 26% were single-divorced in 2013, leaving the other 45% as single-never married never divorced. Marriage and divorce both have large impacts on household wealth, thus leaving Black households at a significant disadvantage. 15 Additionally, the difference in magnitude between the mean household wealth of non- Black households and Black households is stunning: a difference of slightly more than 10- fold. This difference in magnitude dwarfs the oft-cited income gap. Table 1.6 shows the age at which households in my sample left the household of their parents to form their own. On average, third-generation households were 1 year older at the time of household formation as compared with the second-generation households. There does not seem to be any large difference in household formation patterns between Black and non-Black households. The histogram of household wealth, in Figure 1.1, shows the distribution of household wealth is left skewed with a fat tail. Among Black-headed households, the upper tails are thinner, with negligible amounts of households in the upper part of the White household distribution. This is important because in the quantile regression analysis, I am in essence estimating two different conditional distributions of household wealth: one for Black-headed households and one for non-Black-headed households. These conditional distributions are each estimated from the full sample, but they reflect the differences in the unconditional distribution of household wealth between Black and non-Black-headed households. I discuss the differences between the distributions. Table 1.6 Age at Household Foundation All 3rd Gen Second Gen Black 3rd Gen Non-Black 3rd Gen Black 2nd Gen Non-Black 2nd Gen Mean 24.98 25.66 24.46 25.65 25.98 24.76 25 Median 23 24 23 24 24 23 23 16 17 The scatterplot of household wealth and extended family wealth, Figure 1.2, demonstrates the simple bivariate relationship between the two key variables of this study. A bivariate smoothed spline was added to the figure to demonstrate the complex relationship between household wealth and extended family wealth. At low levels of wealth, the relationship between household wealth and extended family wealth is ambiguous, with alternating positive and negative slopes. The spline curve indicates a positive relationship between household wealth and extended family wealth at the upper section of the distribution. The two box and whisker plots, Figure 1.3, demonstrate the skewed nature of the data. The boxes represent the interquartile range of household wealth in each year. The interquartile range represents the spread between the 25th and 75th percentile of data. The length of the whiskers represents 1.5 times the interquartile range. In other words, any data points within the interquartile range are represented by the box, and the whiskers represent any points data points not too extremely removed from the interquartile range, within a distance of 1.5 times the interquartile range of the box. Any data points that are extremely far removed from the interquartile range are plotted as points outside of the whiskers. These points represent extraordinarily large values. As can be seen from the two box plots, my sample contains a sizable amount of extraordinarily large data points above median, every year. This trend emphasizes the highly skewed nature of the data, which I discussed earlier. The highly skewed nature of the data is one of the reasons I decided to use a quantile regression analysis, which is robust to skewed distributions. Figure 1.3 is also useful for demonstrating some of the changes over time inherent in my sample such as the increasing concentration of wealth at the top of the distribution. 18 Figure 1.2. Scatterplot of Household Wealth with Extended Family Wealth 19 Figure 1.3 Box and Whisker Plot of Household Wealth 20 Figure 1.4 displays the time path of median household wealth for Black and non- Black households in the United States. Although certainly interesting in its own right, this often-reproduced figure actually hides a large portion of the story. The resulting patterns are roughly in line with those produced by other data sets, although the PSID is known to produce lower estimates of wealth than the Survey of Consumer Finances (Bosworth & Smart, 2009). Several striking results emerge from Figure 1.5. First, Black and non-Black households in the bottom two deciles in 2013 hold identical amounts of wealth: approximately -$7,000 in the .1 decile and $0 at the .2 decile. Starting at the third decile, the wealth experiences starts to bifurcate. Black households do not accumulate wealth until the .6 or .7 decile, whereas non-Black households already start to accumulate substantial assets at the .3 decile. In 2013, a Black household in the .65 decile held as much wealth as a non-Black household in the .3 decile, approximately $18,000. Another trend is the relationship between the decile distribution of wealth for non-Blacks in 1989 and 2013. Moreover, the distribution of wealth has become much more unequal among non-Blacks over the past three decades, with the middle deciles losing wealth and the upper deciles gaining wealth. A comparable trend is not seen among Black households. Looking at Figure 1.5 again, we can see that the Black-White wealth ratio for the first decile stands at approximately one because Blacks and non-Blacks at the bottom of their respective wealth distributions hold almost the same amount of wealth, about -$7,000. The .2 and .3 decile wealth ratios stand at zero because Black households in both those deciles held zero wealth. An obvious trend emerges at the .4 decile: the degree of wealth inequality diminishes significantly the higher up we climb in the distribution 21 Figure 1.4. Median Household Wealth 1989-2013 $2,777 $2,898 $9,202 $7,283 $9,359 $8,817 $7,440 $4,335 $4,593 $5,565 $65,230 $37,947 $47,610 $55,421 $58,962 $67,376 $65,610 $52,978 $41,290 $42,180 0 10000 20000 30000 40000 50000 60000 70000 80000 1989 1991 1993 1995 1997 1999 2001 2003 2005 2007 2009 2011 2013 Black Households non-Black Households 22 Figure 1.5. Empirical Wealth Deciles by Race for 1989 and 2013 until it grows again at the .9 decile. This is a surprising trend because even though the absolute difference in wealth holdings is starkest in the upper deciles, the degree of wealth inequality is largest in the .2, .3 and .4 deciles. In essence, the absolutely poorest households, those in the bottom tenth, all hold nothing. Response rates of the PSID are generally very high compared to other longitudinal surveys, but the measure of wealth, made up of nine separate asset types, contains a significant amount of values coded as missing. Missing asset values make it impossible to accurately calculate the value of household wealth. Thus, many households have no available wealth information in one year, but may have wealth information in the subsequent year. This quirk of the data creates large year-to-year variability in the number of extended family households counted in the measure of total extended family wealth. $0 $0 $0 $3 $7 $18 $42 $84 -$8 $0 $0 $1 $6 $11 $28 $56 $101 $0 $3 $11 $29 $65 $102 $144 $215 $378 -$7 $1 $6 $17 $42 $79 $134 $225 $469 -10 90 190 290 390 490 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Thousands 1989 Wealth Quantiles for Blacks 2013 Wealth Quantile for Blacks 1989 Wealth Quantiles for non-Blacks 2013 Wealth Quantiles for non-Blacks 23 While our measure of total extended family wealth does capture real variation in the total amount of assets owned by the extended family, some of the year-to-year variation is due to partial nonresponse in addition to attrition. To the extent that the variation in the sum of extended family wealth due to non-response is randomly distributed throughout the sample, it is not necessarily a matter of concern. However, to minimize the potential bias and to maximize the use of available information, I utilize multiple imputation by chained equations, MICE, to impute the value of missing components of wealth. Appendix A discusses the MICE imputation method in detail. 1.5 Methodology 1.5.1 Identification Strategy The goal of this study is to estimate the treatment effect of increases in extended family wealth on own-wealth accumulation and racial differences therein. However, the identification of treatment effects requires a strategy to deal with endogeneity problems in the model. Unobserved heterogeneity can lead to endogeneity issues if not addressed properly. The most obvious sources of unobserved heterogeneity are an individual's ability, intelligence, inherent motivation, socialization, and social connections. The advantage of utilizing panel data is the ability to model the effects of time-invariant unobservable characteristics. If these particular unobserved traits and their effects do not change over time once someone becomes an adult, then we can control for these traits and interpret them as controlling for the long-term effects of childhood environment. The general category of childhood environmental effects would include the effect of particular family 24 resources, quality of education, informal education, the effects of role models, peer effects, access to family or neighborhood social networks, and countless other variables that go into the formation of an adult. To the extent that the effects of childhood environment accumulate until the individual becomes an adult, defined as when the individual creates his or her own household, and thereafter remain constant, then panel models should be able to control for the effects of childhood environment. The panel structure allows us to empirically separate the accumulated effect of childhood environment from the changing effects of the adult environment. Reverse causality can also introduce endogeneity into our model. If the dependent variable, household wealth, is causally related to any of our independent variables, then our estimates may be biased. It is within reason to imagine that household wealth influences the decision to get married, the number of children to have, and long-term earnings potential. As it stands, it is unlikely that we can interpret our independent variables as entirely exogenous. Dealing with endogeneity of this type requires the use of instrumental variables. In particular, merging a geocoded version of the PSID with a geographic data set such as the Small Area Income and Poverty Estimates would open the door to several instrumentation strategies but lies outside of the scope of this study. 1.5.2 Models The fixed effects (FE) estimator is a consistent estimator of the effect of extended family resources on own-wealth accumulation. The FE estimator directly models the relationship between unobserved unique household characteristics that may impact own-wealth accumulation while also allowing for the unobserved household characteristics to 25 be correlated with the X's, the observed household characteristics. FE estimation is attractive because it explicitly deals with the omitted variable problem, for time-invariant factors, which would otherwise cause biased estimates and endogeneity. However, the strengths of the FE model come at a price. A large number of degrees of freedom are eaten up by the inclusion of household dummy variables. For a data set such as this one, with a large number of households and relatively short panels, the degree of freedom constraint may seriously hinder the precision of the fixed effects estimates. Second, the use of household specific fixed effects precludes the use of other time-invariant variables that may be of significant interest, such as race, gender, or level of education. In this study, I follow the exposition of Koenker (2004) in the estimation of a quantile regression model for panel data with fixed effects. The advantage of the Koenker FE model, hereafter referred to as the KFE model, is twofold. First, the FE estimates are constrained to be equal across all quantiles, which reduces the number of estimated coefficients. Second, the KFE model introduces a penalty term, 𝜆1 , which shrinks the estimated fixed effects to a common constant. These two innovations allow for the use of a quantile fixed effects estimator while dealing with the increase in variability introduced by the large number of fixed effects estimates, thus attenuating potential bias (Koenker, 2004, p. 78). Multiple simulation studies have found that this and similar penalized fixed effects estimators for quantile regressions significantly reduce variability without introducing bias (Lamarche, 2010). Another way around the problems caused by fixed effects estimation is to use a random effects estimator to indirectly model the unobserved effects of the household on own-household wealth accumulation. The random effects estimator, instead of using 26 household specific dummy variables, introduces a household-specific component into the error term. This household-specific error component is assumed to be drawn from a population, independently and identically distributed with mean zero and variance 𝜎2 𝑣 . The random effects error term and its sample variance are calculated from the implicitly estimated fixed effects dummy variables (Ashley, 2012). Random effects estimators thus indirectly capture the effects of unobservable household-specific attributes. However, this model requires an additional assumption: all observed X's must be uncorrelated with the household-specific error term. This assumption is untenable in the current context. I would expect the unobserved time-invariant household specific attributes to be correlated with income, education, and household structure. The RE estimator offers more flexible specifications, but because we can now include race and education as independent variables, as well as superior efficiency, the estimates may be biased if this last assumption does not hold. One method for getting around the binding assumptions of the random effects estimator in the context of quantile regression estimation is to use the correlated random effects model introduced by Bache, Dahl, and Kristensen (2011). The intuition behind the correlated random effects, CRE, model "is that one can generate one or more ‘sufficient covariates' from the repeated observations which carry information that can correct for the bias" (Bache et al., 2011, p. 6). In other words, the bias introduced by the relationship between the RE error term and the subset of correlated X's can be negated by incorporating the relationship into the model. Normally, a Hausman test helps us decide between a fixed effects and a random effects model. The intuition behind the Hausman test is relatively simple. The random 27 effects estimator is preferable if estimating the fixed effect intercepts is not a concern, and the additional assumptions of the RE model hold. Since the fixed effects estimator is known to be consistent under a relatively loose set of assumptions, and the random effects estimator is known to be efficient but not necessarily consistent, if both models produce roughly the same point estimates, then the RE model is also consistent and therefore preferable. However, a standard Hausman test is not available for quantile regression models. Nonparametric versions of the Hausman test do exist, but their implementation is outside of the scope of this paper (Henderson, Carroll, & Li, 2008). However, the basic intuition of the Hausman test remains: if both models produce similar results, we have evidence that the CRE model is both efficient and consistent. I provide estimates from both the KFE model and the CRE model in section 1.6 and conclude that the CRE model is consistent and efficient. 1.5.3 Specification Equation 1.1 represents the key relationships investigated in this study. 𝑯𝑾𝒊,𝒕 is the total household wealth of household i in year t. 𝒛𝒊𝜶𝒊 is the unobserved household characteristics for household I, which can be interpreted as a fixed effect or random effect term. 𝑪𝒊,𝒕 is the set of control variables for household i in year t. The control variables include age of head, age squared, income and income interacted with race, dummy indicators of marital status, the number of children under 18 living in the household, a dummy indicator for households with imputed components of wealth, the size of the extended family network, and year fixed effects. Year dummies are included for every year in order to control for the economic circumstances affecting all families in a given year. 28 Income is defined as the total household income for household i in the year immediately preceding time period t.9 Income includes the total taxable income of the head, spouse, and all other family unit members, as well as all the transfer income and 𝑯𝑾𝒊,𝒕 = 𝒛𝒊𝜶𝒊 + 𝜸𝑪𝒊,𝒕 + 𝜷𝟏𝑭𝑾𝒊,𝒕 + 𝜷𝟐(𝑭𝑾𝒊,𝒕 × 𝒃𝒍𝒌𝒊) + 𝒖𝒊,𝒕 (𝑬𝒒. 𝟏. 𝟏) social security income of all members. 𝒃𝒍𝒌𝒊 is the dummy variable for households headed by a person who self-identifies as Black. The key variable of interest is 𝑭𝑾𝒊,𝒕, the sum of the household wealth of all households in the extended family network of family i in year t. 𝑭𝑾𝒊,𝒕 takes the sum of the household wealth of family i's parents' household, uncles' households, cousins' households, sisters' households, and so forth for all extended family households present in the data in year t. 𝑭𝑾𝒊,𝒕 represents the total sum of resources embedded in the extended family network, which includes not only the financial resources that come with wealth, but also the social resources associated with wealth. The relationships expressed in equation 1.1 allow me to answer the research question at the center of this paper: do differences between extended family network characteristics explain racial wealth inequality? In particular, the sign, size, and statistical precision of 𝜷𝟏 and 𝜷𝟐 allow me to answer this question while holding everything else constant. 𝜷𝟏 indicates how changes in the total sum of wealth of the extended family network influence the household wealth accumulation of non-Black households. 𝜷𝟐 represents the differential impact of extended family resources on the wealth accumulation of Black households relative to non-Black households. The linear combination of 𝜷𝟏 + 9 The PSID income questions are retrospective, asking the family for their income last year. 29 𝜷𝟐, the net effect of extended family wealth on Black household wealth accumulation, is presented in section 1.6. 1.5.4 Conditional Quantile Regression The quantile regression model provides two important advantages. 𝑸𝒑(𝑯𝑾𝒊,𝒕|𝑿) = 𝒛𝒊 𝒑𝜶𝒊 + 𝜸𝒑𝑪𝒊,𝒕 + 𝜷𝟏 𝒑𝑭𝑾𝒊,𝒕 + 𝜷𝟐 𝒑(𝑭𝑾𝒊,𝒕 × 𝒃𝒍𝒌𝒊) + 𝒖𝒊,𝒕 𝒑 𝒇𝒐𝒓 𝒑 =. 𝟏, . 𝟐, . 𝟑, … , . 𝟗 (𝑬𝒒. 𝟏. 𝟐) First, I can estimate the marginal effect of changes in the independent variables on the entire distribution of the response function (Hao & Naiman, 2007). For example, I can use the model to estimate the amount of wealth that a hypothetical household located at the first decile of the distribution would have, given the data and the model. I can then see how the wealth of this household would change if the household had one more child, got divorced, or had an uncle who become wealthier, while holding everything else constant. I can repeat this thought experiment for a typical household located approximately at the intersection of every decile from .1 to .9 of the household wealth distribution. To the extent that the model is an accurate representation of the relationship between the covariates and household wealth, we can think of our hypothetical household located at the first decile of the residual distribution as an accurate depiction of a household located at the first decile of the actual household wealth distribution. This hypothetical is important because in the study of inequality, researchers are interested in not just the conditional mean effects, but in the effects of changing economic conditions on the status of the poorest and the 30 wealthiest. Quantile regression gives us this information by estimating nine equations, one for each conditional decile. It is worth noting that quantile regressions use the entire sample in each of the nine regressions.10 The second advantage of using conditional quantile regressions is that it fixes the econometric problems caused by the skewed distribution of household wealth (Hao & Naiman, 2007). Traditional econometric methods based on minimizing the sum of squared errors are biased in the presence of large outliers and skewed distributions. Household wealth suffers from both these afflictions. Running a log-log model would attenuate the effect of outliers and non-normality but is not viable due to a large amount of zero and negative observations of household wealth. I could also drop all outliers, but aside from distorting the statistical inference, it would also obviate the very purpose of this study: to investigate how wealth inequality is impacted by the unequal distribution of extended family resources. As discussed earlier, I utilize two variants of the conditional quantile regression: the KFE model, which is analogous to the fixed effects model, and the CRE model, which is similar to a modified random effects model, discussed earlier. 1.6 Empirical Results I start by comparing the KFE and CRE models. The CRE models utilized in this study take into account the correlation among income, number of children, and marital status and the unobserved household specific attributes. Appendix B contains the regression tables from all covariates of our baseline specification, except the year effects. We can see that both the KFE and the CRE models produce nearly identical results for 10 This is different from separating the data into nine subsamples along the unconditional distribution of household wealth, and then running nine regressions, which would be a series of truncated regressions. 31 every variable. The baseline specification includes household wealth as the dependent variable, an age polynomial, income, married, divorced, children, extended family wealth, year dummies, and interaction terms for income and extended family wealth. Following the logic outlined in section 1.5, I interpret these results as an informal Hausman test, i.e., evidence in favor or using the correlated random effects model. The CRE model should thus provide more precise estimates and greater flexibility in model specification, which I utilize by including years of schooling and a dummy indicator of race in the full specification. All subsequent models use this full specification. Figure 1.6 presents the first result of my analysis: the marginal effects of extended family wealth on household wealth accumulation across the entire conditional distribution. The horizontal axis represents deciles. The vertical axis represents the marginal effect of a one-dollar increase in the sum of extended family wealth on own-wealth accumulation, holding everything else constant. The solid purple line represents the marginal effect of extended family wealth on own-household wealth accumulation for Black households across the wealth distribution, and the solid blue line represents the same effect of non- Black households. The 95% confidence intervals for each group are shaded in their respective color. A useful way to interpret these results is to consider a hypothetical household with attributes such that, conditional on the model, we would expect it to lie at exactly the ith decile of the household wealth distribution. To the extent that our model can take the covariates of our sample households and accurately predict in which decile of the household wealth distribution they belong, we can think of these results as the marginal effects of extended family wealth on own-household wealth for the households located at each decile of the wealth distribution. 32 Figure 1.6. CRE Extended Family Wealth Full Specification Two trends stand out in Figure 1.6. First, there are no statistically significant differences between the effects on Black and non-Black households. We see this because the confidence band for the Black household effect encompasses the non-Black household effect. Thus, we cannot reject the null that both effects are equal. Second, we see that hypothetical households, Black and non-Black, which we would expect to be located at the .8 or .9 deciles wealth distribution, given the model, receive statistically and economically significant benefits from the total accumulated wealth of their extended family network. Households at the ninth decile tend to, on average, accumulate an extra $92 worth of assets whenever one of their relatives accumulates an extra $1,000 of wealth. The effect is smaller for households in the .8 decile, and negligibly small for households located in any other decile. This effect is nontrivial, especially for households located at the .9 decile of the wealth distribution that on average hold several hundred thousand dollars' worth of assets 33 and belong to extended family networks with over a million dollars' worth of assets. This evidence indicates that belonging to the upper echelons of society correlate with a rising tide dynamic that lifts all members of the extended family network at the same time. However, the lack of a racial difference is surprising and warrants further attention, to which this study now turns. The graphs of the marginal effects of all covariates from this model can be found in Appendix B. Due to the inclusion of a race dummy as well as race interaction terms, the interpretation of these results is complicated. The complication arises from the fact that the race dummy and interaction terms create two separate conditional distributions of household wealth, one for Black households and one from non-Black households. For example, consider a household with the following attributes: age 40, income $100,000, married, two children, and an extended family network with 2 million dollars in net worth. To keep the calculations simple, I ignore the household specific intercepts and use a - $100,000 constant. My estimated quantile regression function for the ninth decile predicts that this household would hold approximately $435,000 worth of assets if it was not a Black household, but only $125,000 worth of assets if it was a Black household. If one uncle accumulated an extra $100,000 worth of assets from a good investment, the non-Black household would accumulate an extra $10,000 dollars, bringing its wealth up to $445,000. The Black household, on the other hand, would accumulate only an extra $7,000, bringing its wealth up to $132,00011. This difference in marginal effects, $10,000 versus $7,000, is relatively small, but the point is that when we consider the marginal effects of additional wealth accumulation using my model, we are really looking at how the differential 11 I round the coefficients and ignore statistical significant in these calculations in order to keep the example simple. 34 marginal effects impact two conditional distributions: households at the ninth decile of the non-Black conditional distribution, and households at the ninth decile of the Black conditional distribution. This point is important because the counterfactual of my regressions compares two otherwise identical households, but one is Black and the other is non-Black. Recall that in the unconditional ninth decile of the wealth distribution shown in Figure 1.5, non-Black households hold $469,000 worth of assets and Black households tend to hold $101,000 worth of assets. Although this example is contrived to produce illustrative results, it demonstrates the predictive usefulness of my model. Lastly, since Black households at the ninth decile tend to have about the same wealth as non-Black households at the sixth decile, it makes sense to compare these two groups rather than Black households at the ninth decile to non-Black households at the ninth decile. In order to further explore the underlying trends in the data, I first separate the sample by generation, running separate regressions on each sample by capturing possible cohort differences in wealth accumulation behavior. I then decompose wealth into financial wealth, nonfinancial wealth, and debt in order to focus on financial wealth accumulation. I focus on financial wealth because it is the biggest difference between the portfolios of Black and non-Black households. Following Chiteji and Hamilton (2002), I define nonfinancial wealth as the net value of housing, real estate, and vehicles. Financial wealth is defined as the net value of stocks, bonds, other financial instruments, cash accounts, and businesses. These are assets that directly generate or provide economic resources, whereas nonfinancial assets provide a stream of consumption (Chiteji & Hamilton, 2002; Oliver & Shapiro, 1997). Household wealth is decomposed in this manner to test the hypothesis that financial wealth accumulation follows different dynamics than nonfinancial wealth 35 accumulation. Figure 1.7 shows the marginal effects of changes in total extended family wealth on household-financial wealth accumulation across the financial wealth distribution for second-generation households. The associated graphs of all covariates can be found in Appendix B. In Figure 1.7, we see that Black and non-Black financial wealth accumulation dynamics diverge among second-generation households. The purple line is the marginal effect of one additional dollar of extended family wealth on own-household wealth for second-generation Black households. Non-Black households in the .8 and .9 deciles tend to experience statistically and economically significant gains in financial wealth whenever their extended family accumulates additional wealth. The marginal effect of an additional dollar of network wealth accumulation on financial wealth accumulation for second-generation Black households is negligible and insignificant across all deciles. Importantly, ninth decile black households behave similarly to 6th decile non-Black households: both groups reap little to no benefit from the addition wealth accumulation of their extended family networks. As I noted earlier, it makes sense to compare ninth decile blacks to sixth decile non-Blacks because they are in similar wealth positions. This similarity is indicative of a strong threshold effect: only high wealth households benefit from the additional wealth accumulation of their extended family networks and race may not be an important factor, on its own. On top of this threshold effect, however, there may also be unique racial obstacles preventing Black households from accumulating wealth, further exacerbating racial wealth inequality. Third-generation households tend to have much weaker ties to their extended family network and experience economically and statistically insignificant extended 36 Figure 1.7. CRE Second-Generation Financial Wealth - Full Specification family wealth effects on both financial and nonfinancial wealth accumulation. Third-generation regression figures are located in Appendix B. Since I am controlling for age and age squared, I interpret these generational differences as cohort effects. Those households whose heads were born in the 1980s and 1990s tend to experience much weaker wealth linkages with their extended family than second-generation households, born in the 1960s and 70s. Figure 1.8 respecifies my baseline model, breaking up second-generation extended family network wealth into familial categories: total sibling wealth and total offspring wealth. Offspring wealth has a large and significant effect on second-generation own-household wealth from the fourth decile to the ninth decile. Sibling wealth has a smaller, but still economically and statistically significant effect on the top two deciles. This result contributes to the empirical story of household wealth accumulation. 37 Figure 1.8 Second-Generation Marginal Effects of Sibling Wealth and Offspring Wealth 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Second generation: Marginal effect of sibling wealth -0.1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Second generation: Marginal effect of offspring wealth 38 In a similar vein, I look at how sibling wealth, cousin wealth, and parent wealth contribute to the wealth accumulation of third-generation households. Uncle wealth is omitted due to collinearity. Figure 1.9 shows that sibling wealth is the most important of the extended family wealth components for third-generation households. Whereas cousin wealth has a moderate effect at the ninth decile, and parent wealth has no statistically significant effect, sibling wealth has a moderately sized, steady, and significant effect on own-household accumulation from the fifth decile to the ninth decile. Initially, it seemed like there were no differences between the wealth accumulation dynamics of Black and non-Black households, but looking at financial wealth and generational differences, new patterns appear. Second-generation Black households do not accumulate additional financial wealth when their extended family becomes wealthier. Even the wealthiest Black households do not benefit, but wealthy non-Blacks benefit strongly. Third-generation households have economically small network effects when it comes to financial wealth, less than one fifth the size of the second-generation effects, with no appreciable racial difference. Finally, second-generation household wealth is connected to the wealth of their offspring, whereas third-generation households are connected to the wealth accumulation of their siblings. 1.7 Discussion The wealth of households is tied to the wealth position of their extended family network through one of two ways: transfers or portfolio allocation. Racial differences in transfers may arise for a number of reasons. Blacks were historically precluded from accumulating wealth, and hence have less to pass down (Wilson, 2012). Additionally, there 39 Figure 1.9. Marginal Effects of Extended Family Branches on Third-Generation Wealth 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2 0.22 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 A. Marginal Effect of Cousin Wealth -0.02 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2 0.22 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 B. Marginal Effect of Parent Wealth 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2 0.22 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 C. Marginal Effect of Sibling Wealth 40 is a dispersion effect. Since Black households have more children, when parents or other family members transfer wealth to the youngest generation, it is split among more people, thus diminishing the transfer effect. This dispersion effect is compounded by the fact that Black households have lower incomes and higher rates of poverty. Thus, transfer they receive may not translate into wealth accumulation, but may instead be used for consumption (Chiteji & Hamilton, 2002). Household wealth is also connected to extended family wealth through portfolio allocation. It is well documented that children are affected by the portfolio decisions of their parents, but portfolio decisions may also be influenced by other family members (Chiteji & Stafford, 1999). Portfolio decisions might be altered through emulation or through information flows via the extended family network (Li, 2014). Racial differences in portfolio effects may exist for several reasons. First, Blacks historically have been shut out of mainstream financial markets and hence may distrust financial investments or know little about financial institutions. Second, there may be network disconnects that disproportionately affect Blacks. If there exists a broader social network that contains useful, unique information regarding business opportunities or financial information, then wealthier extended family networks would be the most likely to have access to this broader financial social network. Third, there may be threshold effects that prevent Black households from accumulating wealth. In particular, it seems that comparable Black and White households, in the ninth and sixth wealth deciles, respectively, both have trouble accumulating financial assets. Only the wealthiest non-Black households can be observed to be accumulating large amounts of financial assets in tandem with their extended family networks, which suggests that threshold effects and network effects may be working 41 together to exacerbate wealth inequality and racial wealth inequality. One interpretation of my results is that non-Black family networks can take advantage of this financial information, but Black extended family networks, even the wealthiest, might be shut out of these networks. This disconnect may be driven by discriminatory preferences, a threshold effect that Black households on average do not pass, or by the large degree of social separation between Blacks and non-Blacks in the United States. Thus, even wealthy Black networks would be unable to tap into this information. Since becoming wealthier does not confer useful information that can be passed on to other extended family members, wealthy Black households simply accumulate wealth on their own. In other words, Black households are shut out of the rising tide of rising asset values: accumulate wealth, gain social status, which leads to additional social connections, which provides useful information that can then be transferred to the extended family network, helping them accumulate wealth, and starting the cycle over again. My analysis cannot separate out these different effects. Both threshold effects, resource effects and network effects, may be in play. However, the evidence I have presented here tells a clear story. Wealthy, non-Black second-generation households benefit from the wealth accumulation of their extended family networks. As their family members accumulate wealth, they tend to accumulate additional financial wealth. Second-generation Black households, on the other hand, do not accumulate extra financial wealth when their extended family networks become wealthier. The fact that wealthy second-generation households mostly benefit from the additional wealth accumulation of their offspring, and that third-generation households mostly benefit from the wealth of their siblings, indicates that transfers are perhaps not the main impetus behind the positive 42 network wealth effects on financial wealth experienced by wealthy second-generation non- Black households. Although it is possible for offspring to transfer wealth to their parents, I find it unlikely that adult children are actively transferring wealth to their parents, especially among the upper half of the wealth distribution. Among third-generation households, parents do not appear to be the main source of positive network effects, which also indicates that transfers might not be the main mechanism for these network effects. A more likely scenario is that as offspring, for second-generation households, or siblings, for third-generation households, accumulate extra wealth, they gain access to information or learning that they can then pass on to their extended family networks. The difference in wealth accumulation dynamics between second- and third-generation households might be due to cohort and life cycle differences. Although I am not explicitly controlling for birth cohorts in this analysis, the generational structure of my sample serves as a type of cohort. Since third-generation households are still relatively young, they have yet to hit peak wealth accumulation years and would likely experience different wealth accumulation dynamics than second-generation households that are mostly at or near their peak wealth years. This analysis provides important insights into the racial inequality puzzle. Racial differences in wealth accumulation dynamics exist, particularly in the accumulation of financial wealth. Extended family networks are an important source of resources in the wealth accumulation process. Lastly, there is evidence that a threshold effect is holding back the wealth accumulation of Black households. These insights fill a gap in the literature, a gap that is directly relevant to the study of racial inequality, intergenerational mobility, and health differences. Importantly, this study charts a clear path forward. 43 Additional studies on network effects, portfolio differences, wealth accumulation nonlinearities, and intrafamilial transfers should yield important results. These results have important policy implications. The degree of social isolation of Blacks makes it harder for Black households to accumulate wealth. In the modern era, household wealth is of greater importance than ever. With stagnant wages, rising education costs, and rising health care costs, household wealth plays a larger role than ever in the determination of intergenerational mobility. Social desegregation, in schools, workplaces, and residential areas, can have a large impact on the way Americans accumulate wealth. Innovative housing policies, primary and secondary school bussing, and a renewed emphasis on antidiscrimination policies in the workplace can yield social benefits far beyond their immediate impacts on schooling or housing or wages. CHAPTER 2 WHEREIN LIES THE FERTILE GROUND FOR RACIAL EQUALITY? AN ANALYSIS OF THE SPATIAL RELATIONSHIP AMONG INEQUALITY, SEGREGATION, AND RACIAL WEALTH DISPARITIES 2.1 Introduction Two highly influential social movements, Black Lives Matter and Occupy Wall Street, have brought national attention and scholarly interest to racial inequality and economic inequality. Scholars have pointed out that rising inequality disproportionately impacts Blacks by lowering their chances of economic mobility, whereas others have pointed out that racial inequality increases overall economic inequality (Stiglitz, 2013; Wilson, 2012). However, the connection between the two has not been fully explored. For example, economic inequality is intimately related to income segregation, but racial segregation plays an important mediating role (Reardon & Bischoff, 2011a). The point is that income, income inequality, and income segregation make up a large portion of the social environment in which we live. This study answers the question of how the social environment of metropolitan areas influenced the wealth disparity between Blacks and Whites in the United States during the early 2000s. Using household-level data from the Survey of Income and Program Participation 2001 panel and metropolitan area measures 45 of inequality and segregation from multiple sources, I find that White households tend to accumulate additional wealth in metropolitan areas with higher rates of economic inequality, whereas Black households tend to benefit from residing in relatively egalitarian metropolitan areas. Although I control for measures of income segregation, income segregation does not appear to play a consistent role in explaining differentials in wealth accumulation. This research question is important because a rapidly emerging literature in economics has identified two important trends in the determination of social outcomes in the United States. First, household wealth matters. Researchers have detailed how household wealth plays an important role in the generation of health outcomes and largely determines access to communities (Meer, Miller, & Rosen, 2003; Oliver & Shapiro, 1997). Second, researchers have delineated the pathways through which neighborhoods and geographic context exert significant causal influence on educational and labor market outcomes (Chetty & Hendren, 2015; Chetty, Hendren, & Katz, 2015; Sampson, Morenoff, & Gannon-Rowley, 2002). The cutting edge understanding of the sizeable inequality in social outcomes in the United States must incorporate an analysis of household wealth and social context. This study proceeds as follows. The second section will discuss the data and present some preliminary data exploration. The methods and theory section uses evidence from the literature to build simple causal explanations connecting localized social context to racial wealth inequality and discusses how I broadly test for these dynamics. The results section presents the results from the econometric analysis of the data. The final section discusses the importance of cities and social context. 46 2.2 Data This study combines a household-level panel with city-level data in order to investigate how social context impacts racial wealth inequality. The household-level panel data come from the Survey of Income and Program Participation (SIPP) 2001 Panel, core and topical waves 3, 6, and 9 (U.S. Census Bureau, 2005). The SIPP is administered by the U.S. Census Bureau. The SIPP includes household wealth, the key dependent variable of this study, as well as socioeconomic and demographic control variables including age, race, educational level, and marital status of the household head, as well as total household income and the number of people in the household. The basic descriptive statistics of all household-level variables can be seen in Tables 2.1-2.3. An interesting aspect of Table 2.3 is the amount of households who are single and have never been married, which generally implies a low level of household wealth, despite the generally greater level of disposable income. Although these descriptive are unweighted, they are fairly different than the general population of the United States. The SIPP is uniquely well suited for my analysis for several reasons. First, it is a large nationally representative survey of the United States that contains high-quality, detailed household wealth information. Second, each panel follows every member of a Table 2.1 Household Descriptives - 2001 Statistic N Mean St. Dev. Min Median Max Total HH Net Worth 13,352 196,201 393,665 -900,552 60,775 18,057,560 Total HH Monthly Income 13,352 4,724.3 4,978.9 -49,777 3,477 104,954 Age of HH Head 13,352 48.5 16.7 16 46 85 Number People in HH 13,352 2.7 1.6 1 2 17 47 Table 2.2 Education 2001 % < High School 2,079 15.5 High School 3,436 25.7 Some College 2,478 18.5 College 3,921 29.4 Postgraduate 1,438 10.8 Table 2.3 Marital Status 2001 % Married 6,941 52 Divorced 3,320 24.9 Never married 3,091 23.2 household for 3 to 4 years at 4-month intervals, asking questions related to the socio-economic status of every household member for every month of the preceding 4 months, although wealth information is collected on only a yearly basis. Third, the SIPP contains geographic information at the state and metropolitan statistical area, MSA, level for the years 1996-2003, and only at the state level for 1991-1995 and 2005-2013. The SIPP 2001 panel contains approximately 13,300 sample households, most observed across 3 years, yielding 35,000 household-year observations. The SIPP was designed to be representative of the United States as a whole. However, the downside of using the SIPP is that it was not designed to be representative of individual states or metropolitan statistical areas. Thus, sample sizes for cities and states vary considerably. Household wealth is defined as the total net worth of all members of the household: the sum of assets, including real estate, vehicles, savings, retirement accounts, checking 48 accounts, financial assets, and other assets, minus liabilities, which include debt, credit card debt, student debt, unpaid bills, legal debt, and miscellaneous debt. Table 2.4 gives the unweighted estimates of household wealth per metropolitan statistical area (MSA) derived from the SIPP sample for 12 of the 44 largest MSAs in my sample, including the interquartile range and the number of households residing in each MSA. Median household wealth by race across metropolitan areas contains a large amount of year-to-year variation due to the small sample size of households per metropolitan area. The approximately 13,000 households in the sample are distributed unevenly throughout 44 metropolitan areas. The large variance in the estimates can be clearly seen from the large interquartile range reported in Table 2.4. The most obvious pattern in Table 2.4 is the large difference in White-Black household wealth among large northeastern cities such as Boston, Chicago, and New York, relative to the rest of the United States. However, the racial difference in these cities is not driven so much by lower levels of wealth among Blacks, as much as by the extraordinarily high levels of wealth among White households. The histogram and kernel in Figures 2.1 and 2.2 present the distribution of change in household wealth between 2001 and 2003, i.e., wealth in 2003 minus wealth in 2001 for each sample household. The median Black household gained $144 in wealth, whereas the median White household gained $9,472 during this time period. As can be seen above, Black households were far less likely to experience a large increase in wealth. To get a better sense of the distribution of household wealth across the United States, Figure 2.3 highlights the large amount of variation in median household wealth across states in 2001. 49 Table 2.4 Median Household Wealth by CBSA: 2001-2003 2001 Metro Area non-Black IQR N Black IQR N Atlanta-Sandy Springs-Roswell, GA 89,575 231,310 267 19,070 100,826 101 Boston-Cambridge-Newton, MA-NH 153,000 340,794 455 100 11,654 21 Chicago-Naperville-Elgin, IL-IN-WI 129,320 244,921 637 4,976 56,108 152 Dallas-Fort Worth-Arlington, TX 56,272 156,105 400 5,554 47,180 79 Detroit-Warren-Dearborn, MI 105,822 230,612 356 14,530 55,008 115 Houston-The Woodlands-Sugar Land, TX 49,900 166,854 337 3,448 50,658 82 Los Angeles-Long Beach-Anaheim, CA 62,222 176,282 386 8,375 59,231 87 Miami-Fort Lauderdale-West Palm Beach, FL 94,850 295,845 1,485 885 74,812 362 New York-Newark-Jersey City, NY-NJ- PA 110,725 259,025 469 7,700 37,200 113 Philadelphia-Camden-Wilmington, PA-NJ-DE-MD 204,186 474,247 520 14,902 156,202 56 San Francisco-Oakland-Hayward, CA 57,908 265,770 313 18,824 69,623 14 2002 Metro Area non-Black IQR N Black IQR N Atlanta-Sandy Springs- Roswell, GA 95,000 198,036 247 27,892 77,687 93 Boston-Cambridge-Newton, MA-NH 168,516 347,592 404 101 38,838 19 Chicago-Naperville-Elgin, IL-IN-WI 127,013 274,313 566 6,590 75,446 131 Dallas-Fort Worth-Arlington, TX 52,424 139,287 353 6,719 42,658 66 Detroit-Warren-Dearborn, MI 126,400 235,344 309 6,150 60,393 90 Houston-The Woodlands-Sugar Land, TX 52,456 184,196 283 14,460 61,709 64 Los Angeles-Long Beach-Anaheim, CA 74,684 180,338 337 25,675 70,325 75 Miami-Fort Lauderdale-West Palm Beach, FL 130,182 355,896 1,284 1,750 73,180 30 0 New York-Newark-Jersey City, NY-NJ- PA 124,086 233,839 410 13,195 42,872 99 Philadelphia-Camden-Wilmington, PA-NJ-DE-MD 212,440 565,667 457 8,887 173,303 50 San Francisco-Oakland-Hayward, CA 102,076 270,099 272 39,500 149,275 15 50 Table 2.4 Continued 2003 Metro Area non-Black IQR N Black IQR N Atlanta-Sandy Springs-Roswell, GA 89,869 225,131 244 20,498 119,098 93 Boston-Cambridge-Newton, MA-NH 198,055 419,140 387 2,350 5,492 15 Chicago-Naperville-Elgin, IL-IN-WI 178,750 335,092 517 9,216 79,825 110 Dallas-Fort Worth-Arlington, TX 68,591 154,454 334 16,000 48,830 61 Detroit-Warren-Dearborn, MI 134,586 245,327 272 7,000 63,700 73 Houston-The Woodlands-Sugar Land, TX 63,950 219,622 267 2,600 44,298 54 Los Angeles-Long Beach-Anaheim, CA 108,206 230,203 324 23,750 73,438 70 Miami-Fort Lauderdale-West Palm Beach, FL 136,876 394,566 1,198 1,838 97,848 276 New York-Newark-Jersey City, NY-NJ- PA 150,189 268,458 396 30,000 73,942 89 Philadelphia-Camden-Wilmington, PA-NJ-DE-MD 322,009 605,413 424 35,750 267,202 45 San Francisco-Oakland-Hayward, CA 111,500 303,866 251 24,038 140,473 15 Figure 2.1. Histogram of Change in Household Wealth Between 2001 and 2003 51 Figure 2.2. Kernel of Change in Household Wealth Between 2001 and 2003 Figure 2.3. Geographic Distribution of Household Wealth in 2001, Weighted 52 The geographic data contained in this analysis are drawn primarily from two sources. The data on income segregation are taken from the US2010 Project, http://www.s4.brown.edu/us2010/, which provides public access to the income segregation data constructed by Reardon and Bischoff (2011). The rest of the geographic data utilized in this study is drawn from The Equality of Opportunity Project, http://www.equality-of-opportunity. org/, run by Raj Chetty and Nathaniel Hendren. One important caveat to keep in mind is that the geographic variables are not exactly matched to the geographic location of the sample households. The SIPP 2001 panel is coded using the 1993 MSA/CSA definitions, whereas the data from the US2010 Project are coded based on the 2000 census Core Based Statistical Areas definition, which is slightly different. Furthermore, the data taken from The Equality of Opportunity Project are coded at the commuting zone, CZ, level. A commuting zone is a different geographic unit than MSA. Most commuting zones and corresponding MSA are centered around the same city, but with different boundaries. However, there is significant overlap, with some metropolitan areas mapping very closely to commuting zones. An MSA is based around an urban center with a population of at least 50,000 people and any surrounding counties that are tied to the urban center via commuting patterns. However, metropolitan statistical areas are somewhat arbitrary because they are bound by a minimum population threshold, political boundaries of county lines, and an emphasis on commuting patterns. Commuting zones (CZs) are an alternative geographic unit based on labor market patterns with no minimum population thresholds. CZs are a geographic unit that combines where people live and where they tend to work. One difference between metropolitan statistical areas and commuting zones is the inclusion of small, rural labor 53 markets in the definition of commuting zones, whereas metropolitan statistical areas completely ignore these areas or subsume them into a larger urban unit. Another difference is that MSAs sometimes treat two neighboring urban areas as distinct units because they each have large populations, are separated by a county line, and have distinct commuting patterns. However, the two MSAs may be economically intertwined due to shared labor markets, in which case both urban areas would be placed into one CZ. When dealing with large urban centers, the geographic delineations of MSAs and CZs are sometimes identical. For example, the Portland, OR MSA and CZ are the same. However, the Seattle-Tacoma-Bellevue MSA is much smaller than the Seattle CZ because the Seattle CZ includes nearby urban centers in the labor market of Seattle. An in-depth discussion of the differences between MSAs and CZs lies outside the scope of this paper, but suffice to say that on average, the MSAs and CZs of the major urban areas of the United States contain significant overlap, which provides meaningful geographic information. Combining MSA and CZ data into a unified analysis obviously poses certain problems, but on average, the potential differences are outweighed by the similarities, and hence this analysis is meaningful. The measures of income segregation taken from the US2010 Project, income segregation, poverty segregation, and affluence segregation, are calculated from the rank order information theory index. The measure of poverty segregation divides the population into poor and nonpoor along the bottom decile, and calculates the ratio of census tract variation in poor to nonpoor residents relative to the overall metropolitan area variation in poor to nonpoor residents. For example, if the poorest 10% of the population of a metropolitan area all live together in census tracts, which contain only other poor families, 54 then poverty segregation is equal to 1, indicating perfect poverty segregation. If, instead, each census tract contains 10% poor households, then the metropolitan area is perfectly unsegregated, and poverty segregation is equal to zero. The definition of affluence segregation is analogous, but for the top decile. The overall measure of income segregation takes a weighted average of income segregation at every centile of the income distribution, with the weights maximized at the median and approaching zero at the tails (Reardon & Bischoff, 2011a). These measures of income segregation are useful because they are independent of the level of income and income inequality within a metropolitan area. The segregation index relies exclusively on measuring the spatial variation between ranks of the distribution (Reardon & Bischoff, 2011a). This independence of income segregation from income inequality or income levels is important because I specifically want to control for all three of these metropolitan attributes. The average income, income inequality, and the spatial segregation of household income are theoretically separate factors that change the social makeup for a metropolitan area in different ways. The other geographic variables can be divided into three categories. First are variables that control for the economic structure of a metro area: high school dropout rate, fraction working in manufacturing, and earning income tax credit (EITC) exposure. The high school dropout rate was calculated from the National Center for Education Statistics CCD 2000-2001 and is adjusted for the income per-capita of the local area.12 Faction working in manufacturing measures the proportion of people 16 or over working in 12 The income adjusted high school dropout rate is calculated as the residual of an OLS regression of income per capita on the high school dropout rate and thus represents a measure of the quality of schooling independent of local income. 55 manufacturing, as reported in the 2000 census. EITC exposure refers to the extra amount that certain states choose to contribute to the Earned Income Tax Credit on top of the amount the federal government provides. I include EITC exposure because it increases the incentive to work for those at the poverty level. Although the main effect of EITC state top-ups is on income, getting people off means-tested welfare programs also impacts their choice to save their rates of wealth accumulation (Chetty, Friedman, & Saez, 2013). Next are variables that control for the demographic aspects of a metro area: the fraction of the population made up by Black households and the fraction foreign born, both of which are measured using the 2000 Census. Third is the set of variables that account for factors that influence the social structure of the metro area: income segregation, household income per capita, income inequality, social capital, racial segregation, and the share of income accruing to the top 1% of households. Income segregation is defined by the rank-order information theory index, discussed above. Household income per capita is defined as the sum total of income for the commuting zone in the 2000 census, divided by the number of people ages 16-64. Social capital is a standardized index of various measures of civic engagement, such as filling out census forms and participation in business, political, religious, and professional associations at the county level (Rupasingha, Goetz, & Freshwater, 2006). Since social capital is a standardized index, it ranges from negative values to positive values, representing how much more social capital a geographic area contains relative to other areas of the United States. For this study, I limited the sample to 44 metro areas that contained at least 100 SIPP households in 2001. Table 2.5 presents the metropolitan area descriptive statistics. 56 Table 2.5 Metropolitan Area Descriptives, Unweighted Statistic N Mean St. Dev. Min Median Max high school drop rate 28 0.008 0.017 -0.020 0.006 0.053 household income per cap 44 41.064 4.894 28.267 40.769 54.014 fraction in manuf. 44 0.126 0.050 0.039 0.120 0.272 EITC exposure 44 1.334 3.483 0.000 0.000 17.286 fraction foreign born 44 0.113 0.085 0.025 0.089 0.397 fraction Black 44 0.138 0.088 0.010 0.129 0.371 population 44 3,068,044 2,952,164 262,282 2,347,654 16,393,360 poverty segregation 44 0.091 0.018 0.044 0.095 0.115 income segregation 44 0.145 0.023 0.097 0.145 0.192 affluence segregation 44 0.139 0.042 0.059 0.131 0.289 racial segregation 44 0.264 0.095 0.103 0.259 0.474 Gini 44 0.487 0.066 0.358 0.477 0.684 income share of top 1% 44 15.822 3.642 10.785 15.280 29.143 social capital 44 -0.439 0.900 -2.317 -0.443 1.348 2.3 Methods and Theory The key innovation of this study lies in the use of geographic data to explore how the social landscape of one's city impacts household wealth accumulation and racial inequality. In particular, this study seeks to investigate how income levels, income inequality, and income segregation impact the social character of a city, which in turn impacts the Black-White differential in household wealth accumulation. In a simple accounting framework, the accumulation of wealth depends on bequests, income, savings rate, and portfolio efficiency. However, this accounting framework tells us very little about why racial wealth inequality is so large in the United States. Black households do, on average, have lower income, and receive smaller bequests 57 from family members (Wolff & Gittleman, 2011). However, even compared to similar White households, Black households have dramatically lower levels of wealth (Gittleman & Wolff, 2004; Oliver & Shapiro, 1997). Similarities in income and education can mask considerable differences in lifetime earnings due to greater income volatility among Blacks as well as very different starting points. Recent empirical evidence suggests that unemployed Black men must spend more time and effort in order to find a new job, compared to similar White men, due to discrimination (Fryer, Pager, & Spenkuch, 2013; Pager, Western, & Bonikowski, 2009). This discriminatory job market dynamic decreases the household wealth of Blacks relative to Whites due to the drawdown of wealth during spells of unemployment or loss of credit rating. The extent to which the disparity in the job search process is influenced by localized social factors has not been extensively studied, but I hypothesize that job search disparities represent one mechanism through which social context influences Black-White income disparities and disparities in wealth controlling for income. Although I am not aware of any empirical studies investigating the connection between localized social factors and discriminatory attitudes, sociological field studies involving interviews with employers have found that employers explicitly take into account the geographic distribution of poverty and race in their formulation of racial/class stereotypes (Kirschenman & Nickerman, 2001). This qualitative evidence lends credence to my hypothesis that localized social factors influence racial and class-based attitudes. Additionally, this discriminatory job-search dynamic would not necessarily be reflected in the household-level income data except in very long panels, which accurately measure differences in lifetime earnings. Such panels are rare, and the small sample sizes of panels such as the PSID could mask this dynamic, which might be observable only as a 58 geographic correlate of wealth disparities. Highly income-segregated places tend to also be highly racially segregated (Reardon & Bischoff, 2011a). In particular, income segregation is greater among Blacks than Whites. Thus, in highly segregated cities, the self-fulfilling cycle of racial and class stereotypes feeding into greater poverty, which feeds into racialized stereotypes about poverty, can create a long-term dynamic where employment discrimination would be greater in these cities. Similar hypotheses of the creation or perpetuation of racial and class stereotypes have been proposed in the historical and sociological literature, but quantitative evidence remains scarce (Kruse, 2007; Massey & Denton, 1989; Sugrue, 2005; Wacquant, 2010; Wilson, 2012). This inequality-stereotype- discrimination dynamic links trends in income inequality and income segregation to localized discriminatory behavior and thus to racial disparities in wealth accumulation. This story maps closely to a large body of literature that shows that racial stereotypes and racial disparities are not a national phenomenon in the United States, but are instead localized phenomenon with large variation across regions (Crowder, South, & Chavez, 2006; Pais, South, & Crowder, 2012). I can indirectly test for evidence in favor on this stereotype-inequality-discrimination hypothesis by investigating if higher rates of average income, income segregation, and income inequality are negatively associated with Black household wealth accumulation. Another dynamic that might provide a causal link between local social context and racial wealth disparities is network effects. This particular dynamic has been discussed widely in the literature, although usually only in general discussions on the effects of rising inequality (Chetty, Hendren, Kline, & Saez, 2014; Corak, 2013; Sampson, 2009). 59 Generally, rising inequality is envisioned as creating larger social separation between income classes, which creates separated social networks between income groups. Thus, lower income groups would not benefit from the social capital, information flows, peer effects, and other benefits that accrue to upper-income networks. This dynamic is supported by strong empirical evidence that rising income inequality does tend to literally make the wealthy live further away from the poor, and the poor live increasingly segregated on their own (Reardon & Bischoff, 2011a). Additionally, rising inequality disproportionately increases income segregation among Blacks. Thus, as income inequality increases, middle-income Blacks are segregated from middle- and upper-income Whites, and low-income Blacks are segregated from everybody. I would expect this dynamic to impart negative network effects on poor and middle-income Blacks who become increasingly separated from the rest of the society. I hypothesize that this increasing social separation could lead to lower rates of portfolio efficiency: even middle-income Blacks would have worse performing portfolios if they have less access to information, business contacts, and social resources. Lastly, local social context can influence wealth inequality directly by impacting the housing market, which represents a disproportionately large portion of Black wealth portfolios (Chiteji & Stafford, 1999). If rising income inequality and income segregation create falling tax revenues in poor neighborhoods and rising tax revenues in high-income neighborhoods, I would expect the shift in public goods, social services, and neighborhood quality to increase property values in high-income neighborhoods and to decrease property values in low-income neighborhoods, everything else equal (Reardon & Bischoff, 2011a). To the extent that increasing income segregation disproportionately affects low- and 60 middle-income Blacks as discussed earlier, this dynamic would cause a disproportionate negative effect on Black wealth, even among middle- income Blacks, relative to Whites. I obviously cannot identify which of the three hypotheses is at play with the same limited set of coefficients. However, generally testing the hypothesis that local social context influences racial wealth inequality is an important step forward for the scholarly literature on this subject. To date, no studies have rigorously investigated this dynamic. 2.4 Model and Specification The combination of household and geographic data in my sample limits the number of available econometric models at my disposal. First, the geographic data are static, mostly for the year 2000. Thus, it is not possible to use a fixed effects model to control for unobserved heterogeneity because the fixed effects would be collinear with the geographic data. Instead, I used an alternative: a pooled regression with household- and year-clustered standard errors. Although this model does not control for unobserved heterogeneity, it does account for the correlation of standard errors within households and within years while allowing me to investigate the effects of living in various metro areas. As a robustness check, I also run a series of random effects models using the same specifications. These can be seen in Appendix C. The random effects estimator indirectly models the unobserved effects of the household on own-household wealth accumulation. The random effects estimator, instead of using household-specific dummy variables, introduces a household-specific component into the error term. This household-specific error component is assumed to be drawn from a population, independently and identically distributed with mean zero and variance 𝜎2 𝑣. The random effects error term and its sample 61 variance are calculated from the implicitly estimated fixed effects dummy variables (Ashley, 2012). Random effects estimators thus indirectly capture the effects of unobservable household-specific attributes. However, this model requires an additional assumption: all observed X's must be uncorrelated with the household-specific error term. I would expect the unobserved time-invariant household specific attributes to be correlated with income, education, and household structure. Whereas the random effects estimator offers more flexible specifications, because we can now include race and education as independent variables, as well as superior efficiency, the estimates may be biased if this last assumption does not hold. I find that the random effects regressions produce the same results as the pooled-clustered regressions. For expositional purposes, I present the pooled regression results as the main results, but I do not necessarily see them as preferable to the random effects results. Measures of metropolitan area income level, income inequality, and income segregation are the three coefficients of interest I use to test my proposed hypotheses. Measures of income level and income segregation are straightforward; I use income per capita for income level, and the Reardon and Bischoff measures of income segregation for income segregation. For income inequality, I use two alternative measures: the gini coefficient and the income share of the top 1% in different specifications. Since there is no existing literature on which geographic variables have an impact on household wealth accumulation, I use an informal stepwise method to pick and test five alternative econometric specifications. The purpose of the five alternative specifications is to test different measures of inequality as well as to test which set of geographic controls is appropriate. The first specification presented in column 1 of Table 2.6 presents the results 62 Table 2.6 Effects of Metropolitan Characteristics on Household Wealth Accumulation Dependent Variable: Scaled total net worth of Household (1) (2) (3) (4) (5) HH inc per cap -0.001 0.004* 0.004 0.006*** 0.006** (0.009) (0.002) (0.003) (0.002) (0.002) income segreg -0.025 -0.037* -0.035* -0.033* (0.016) (0.021) (0.020) (0.020) racial segreg -0.680 (0.713) gini -0.357 0.161*** -0.079 (0.446) (0.054) (0.085) Inc, share 1% 0.010** 0.002 0.004** (0.005) (0.002) (0.002) social capital 0.082 0.016 -0.001 -0.008 (0.061) (0.010) (0.008) (0.008) B*HH income 0.001 -0.004* -0.003 -0.005*** -0.005** (0.007) (0.002) (0.002) (0.002) (0.002) B*income segre 0.027 0.027 0.025 0.024 (0.023) (0.024) (0.022) (0.022) B*racial segreg. 0.129 (0.304) B*gini 0.784 -0.138*** 0.044 (0.548) (0.053) (0.095) B*social capital -0.019 -0.012 0.001 0.005 (0.015) (0.011) (0.008) (0.008) B* inc share 1% -0.018** -0.002 -0.004** (0.008) (0.002) (0.001) Constant -0.216 -0.563*** -0.573*** -0.675*** -0.590*** (0.506) (0.112) (0.137) (0.115) (0.148) Observations 23,384 35,877 35,877 35,877 35,877 R2 0.017 0.021 0.021 0.021 0.021 Adjusted R2 0.016 0.021 0.020 0.020 0.020 Residual Std. Error 1.364 (df = 23349) 1.111 (df = 35852) 1.111 (df = 35856) 1.111 (df = 35858) 1.111 (df = 35858) *p<0.1; **p<0.05; ***p<0.01 Note: Standard errors are heteroscedastic robust and clustered by household and year. 63 from the full model with a large set of covariates. I remove one or two covariates from each subsequent column, moving from the full specification toward a relatively parsimonious specification in column 5. 2.5 Limitations This study cannot, at the moment, circumvent two important empirical limitations. First, bias stemming from reverse causality or unobserved heterogeneity has not been controlled for. Second, a particular type of reverse causality, the self-selection of certain types of households into metropolitan areas, can also bias the results. That being said, this study is doing something entirely novel that has large potential impacts on an ongoing national policy debate: is inequality a problem, what are the consequences, and how can we address them? The results are highly suggestive and important, even if they contain potential bias. 2.6 Results The results from the analysis are sensitive to the model specification, which is not unexpected given the short panel with which I am working and the close correlations between metro income, income inequality, and income segregation. This collinearity can be dealt with in future studies with better data. In particular, using a finer grained geographic unit would allow for more precision. Although these three factors are conceptually different and measured independently from one another, the degree of collinearity between them may impact the estimation and precision. The results presented in Table 2.6 exclude all household-level variables and 64 miscellaneous geographic covariates not directly related to the central hypotheses. The full results can be seen in Appendix C. Looking at the model specification in column 4 of Table 2.6, higher-income metros tend to favor White wealth accumulation, and metros with higher rates of income inequality benefit White households as well. Conversely, Black households do worse in higher-income and higher-inequality metropolitan areas. Additionally, income segregation is associated with less wealth accumulation for everyone, on average, holding all else equal. Since the dependent variable, total household wealth, is measured in standardized units and household income per capita is measured units of $1,000, column 4 tells us that a White family moving to a city with a household income per capita $1,000 higher would, on average, experience a wealth boost of .006 standard deviations of household wealth. These results roughly comport with my proposed hypotheses. I interpret these results as weak, partial evidence of my proposed hypotheses. The model specification in column 3 uses the top 1% income shares rather than the gini as the measure of income inequality. This specification finds evidence that higher rates of income inequality benefit Whites but have no effect on Black households. These results are very similar to the results from column 4, despite the alternative measure of income inequality. I interpret this as evidence that income inequality has a robust differential effect on the wealth accumulation of households. Once again, income segregation has a negative effect on all households. The specification in column 2 contains two more geographic control variables than the specification in column 3: fraction foreign born and fraction in manufacturing, although they are not shown in Table 2.6. The inclusion of these two covariates shrinks the effect of 65 income inequality, rendering it statistically insignificant at the 95% level. However, income segregation remains important. As mentioned earlier, these results are somewhat sensitive to the specification. Since the city-level covariates are the main subject of this analysis and also the most binding limitation in terms of data, I test the sensitivity of the results to alternative restrictions on the sample of cities included in my sample. This set of alternative sample-restriction tests is important in order to check that the data from any one city that happens to be included in the sample are not driving the results. For these tests, I selected three metro areas and excluded all household members who reside in these cities, one by one with replacement. I first excluded the New York City metro area, then the Los Angeles metro area, and lastly, the Chicago metro area. These three cities were chosen because they are the three largest cities in the United States by population, but also because they each have a unique history, economy, and set of social institutions. Thus, it is useful to check if one of these three cities has a unique and strong effect on the wealth accumulation of its residents. The three auxiliary regressions follow the same specification and functional form as my main regression, reported above. These three auxiliary regressions can be found in Appendix C. I find no evidence that these three cities influence the central results. Excluding Chicago or Los Angeles leads to almost no changes in the results, whereas excluding New York City does reduce the size of the coefficients on income segregation and income inequality, but only by approximately 24%, not enough to change the sign of the coefficient or significantly alter my conclusions. The exclusion of households from New York City does marginally decrease the p-values of the coefficients on income segregation and income inequality. These sample-sensitivity tests indicate that my results 66 are not heavily influenced by any of these three unique cities. 2.7 Discussion Income segregation and income inequality have increased within cities for the past 30 years (Reardon & Bischoff, 2011a). Table 2.6 provides evidence in favor of the hypothesis that the social character of a city contributes to a widening wealth gap between Blacks and Whites. More broadly, this study contributes to an emerging literature that explicitly takes into account the effects of living in unequal or segregated areas (Chetty et al., 2014). I used three metropolitan area variables, income level, income inequality, and income segregation, as indirect measures of the social structure of cities. Although these three variables are obviously related to each other - rising income and growing inequality generally lead to income segregation - they are each theoretically separable factors that contribute to the social character of a city in different ways. For illustrative purposes, I consider two very different cities--Portland, OR and New York City, NY-- in the year 2000. Portland, with a population of 1.8 million, 2.6% Black, has one of the lowest levels of racial segregation of any major city. Importantly, Portland has one of the lowest rates of income segregation, with an H-index of 0.101, a household income per capita of $41,000 per year, a gini coefficient of 0.42, and a 15.8% income share of the top 1%, making it a generally equitable major city. Additionally, Portland has a White/Black dissimilarity index of 47.4 in 2000, making it a relatively racially unsegregated city (Reardon & Bischoff, 2011b). On the other side of the inequality spectrum, I chose New York City as the prototypical example of an unequal city. With an income segregation index of 0.192 and a 67 gini coefficient of .68, it is simultaneously the most unequal and the most segregated city in my sample. With a population of 16 million and 20% Black, New York City is highly racially segregated with a racial segregation Theil index of 0.38 and a White/Black dissimilarity index of 79.5 in the year 2000 (Reardon & Bischoff, 2011b). It is not hard to imagine how these patterns of inequality and segregation change the social character of a city over time. Note, in particular, the concentration of poverty in New York City relative to Portland. The difference is dramatic. In the long run, stereotypes related to where people live, attend school, and their socioeconomic background and race will develop as cities become increasingly stratified. This historical process of the creation, dissemination, and reification of stereotypes has been documented for several cities: Detroit, Chicago, and Atlanta (Kruse, 2007; Sampson, 2013; Sugrue, 2005). A surprising finding of sociological field work is how candid employers were regarding their use of residential location, school location, and race in their hiring and employment decisions (Kirschenman & Nickerman, 2001). Looking at the map for New York City, we must also consider the role that New York City's massive income inequality plays in exacerbating the effects of income segregation. The difference in lifestyle and physical environment between the segregated "high-income" neighborhoods and the low-income neighborhoods will be much starker in New York City than in Portland, where the rich are not as different relative to the poor. It is the combination of extreme segregation and stark resource differences that I hypothesize creates the potential for pervasive stereotypes in New York City. Furthermore, the development of these stereotypes impacts more than just discriminatory job market behavior. It can impact discrimination in the housing market through steering, and in the credit market through mortgage discrimination, thus adversely 68 affecting wealth accumulation of all Black households through multiple channels (Greater New Orleans Fair Housing Action Center, 2014; Ladd, 1998). The indirect impact of network effects on household wealth accumulation works differently in these two cities. The poor and rich of New York City are so separated by space and resources that they may never interact with each other in any meaningful way, thus creating negative network effects for those at the bottom of the distribution, which are disproportionately Black and minority households, and positive network effects for those at the upper-middle and top sections of the income distribution. My results that income inequality benefits White households on average were surprising, but the results make sense because information, business contacts, referrals, and investment opportunities are scarce resources that become more valuable as fewer people have access. Thus, high inequality may be what creates the economic value of network resources. The pecuniary value would be arbitraged if everyone had access to it. Being locked out of high-end networks does not necessarily hurt a household, but it provides a significant boost if it does have access to these limited resources. I interpret my results as indicating that income inequality creates differential access to network resources, to which most Black households do not get access, especially in a city like New York. The results of this study open up an entirely different avenue through which to investigate the economic costs of rising inequality in the United States. In particular, the racial dimension had not been fully explored in the previous literature. For example, New York City's recent moves to increase the minimum wage should help ameliorate racial wealth inequality through several channels, not just through increasing the incomes of the poor. Additionally, New York City's renewed commitment to publicly funded mixed- 69 income housing is a step in the right direction, and this study indicates that the benefits could be far reaching (The Editorial Board, 2016). CHAPTER 3 WHAT'S WEALTH GOT TO DO WITH IT? RACE, RACIAL INEQUALITY, AND WEALTH IN THE UNITED STATES 3.1 Introduction Making sense of racial inequality in the United States requires the integration of household wealth into an analysis of the role that social context plays in the perpetuation of racial inequality. Specifically, wealth plays a central mediating role, connecting lived social experience to the long-term persistence of racial inequality in the United States. This mediating role is complex and multidimensional. In this essay, I propose a model that can explain some of the social dynamics that contribute to the perpetuation of racial wealth inequality in the United States. The first two chapters of this dissertation provide empirical evidence that social context has important effects on racial wealth inequality and explore some of the possible mechanisms through which this effect takes place. In this last theoretical chapter, I discuss the history of racial inequality in the United States, and then present a model to analyze the role of social context and markets in the determination of racial wealth inequality. I conclude by discussing how the first two chapters fit within this framework. 71 3.2 Why Race, Why Black? The Relevance of Race in the United States A well-established literature exists that probes the idea of race and outlines the historical development of racial categories, and in particular Black identity (Fields & Fields, 2014; Wilson, 2012). Although authors disagree about how race is socially constructed or the significance of the social construct, they agree that race is a social construct and has real, tangible consequences. Two unique aspects of the Black experience in the United States set it apart from that of any other group, perhaps save Native Americans. First, Blacks have occupied a position firmly at the bottom of the social hierarchy since emancipation, consistently ranking at the bottom in terms of pay, access to resources, health, and political power (Heilbroner & Singer, 1998; Margo, 1994; Wilson, 2012). Although rarely subject to the same degree of social stigma, other groups suffered discrimination and poverty upon arrival, such as Slavic and Italian immigrants during the 19th century, but they would eventually assimilate into the U.S. economy within a matter of decades (Ferrie, 1994; Heilbroner & Singer, 1998). These two divergent experiences beg the question, what social or economic processes enable some groups to assimilate while leaving other groups to languish? The second unique aspect of the Black experience in the United States is the degree and longevity of social isolation Blacks have faced. Social isolation generally takes three forms: geographical isolation, economic isolation, and ideological isolation. Geographic isolation involves residential isolation, a physical separation between Blacks and Whites. Segregation can take place at the macro level, where Blacks and Whites physically live far apart from each other in homogenous neighborhoods, or at the micro level, where Blacks and Whites may live in proximity but never in the same neighborhood. Different cities in 72 the United States exhibit these two forms of residential segregation (Massey, 2001; Reardon & Bischoff, 2011; Sampson & Sharkey, 2008). Economic isolation involves the segregation of Blacks and Whites into different socioeconomic groups. One clear example is the disproportionate segregation of Blacks into menial labor or low-status positions within a corporate hierarchy and Whites into skilled or high-status positions. Economic isolation perpetuates norms of authority and power, reinforcing notions of White superiority and Black inferiority. The growing numbers of well-educated, middle-class Black household may imply that the degree of economic isolation is diminishing, but it may in fact be changing forms. The definition of socioeconomic status is complex. This dissertation makes the point that household wealth is a central factor in the determination of socioeconomic status. Middle-class Black men and women may be working in skilled or high-status jobs, but on average, they occupy a lower rung in the social hierarchy due to persistently low levels of household wealth (Conley, 2010). Ideological isolation involves the set of cultural notions, social institutions, and ideology that justify socioeconomic differences between Blacks and Whites (Anderson, 2013). Even if geographic isolation or economic isolation is not prevalent in certain areas, the idea that Blacks are somehow intrinsically different from Whites creates a degree of social separation between the two groups. Many authors argue that ideological isolation is merely the end result of geographic and economic isolation or historical discrimination (Anderson, 2013; Fields & Fields, 2014). However, ideological isolation, once established as a norm, may take on a life of its own, exacerbating the effects of isolation and feeding on itself (Kirschenman & Nickerman, 2001; Sugrue, 2005). I consider ideological isolation 73 an important factor to consider in the analysis of the long-term persistence of racial inequality. I argue that these three dimensions of social isolation, taken together and maintained over a long period of time, have created a uniquely deep and pervasive degree of social isolation for Black households in the United States. It is the connection between social isolation and the long-term persistence of racial inequality in which I am interested, and I believe that household wealth plays a central role in this connection. In this chapter, I propose a conceptual model that explains some of the connections between social isolation and racial wealth inequality. This model explicitly details how the network isolation of Black communities impacts their ability to accumulate wealth, thus perpetuating racial inequality and differences in social outcomes. Racial differences in social outcomes are very real, even if race is a social construct. Three competing hypotheses attempt to explain why race is still important. First, unobserved class or socioeconomic factors related to race explain the persistent disparities in social outcomes between Blacks and Whites. This is the position taken by a large number of scholars who find that specific Black-White gaps, such as educational outcomes or home ownership, shrink to negligible size once disparities in socioeconomic status are fully accounted for (Conley, 2010; Hilber & Liu, 2007). According to this hypothesis, it is the lack of resources and human capital that is to blame for racial inequality. The second popular hypothesis holds that subtle forms of institutional racism and discrimination remain active and important throughout the economy. For example, one form of institutionalized racism that has gained attention as of late is mass incarceration and the criminal justice system (Coates, 2015; Temin, 2015; Wacquant, 2010). Other forms of 74 discrimination are assumed to take place via discriminatory preferences, statistical disc