Intergenerational Neighborhood-Type Mobility - Bryn Mawr College

Intergenerational Neighborhood-Type Mobility: Examining Differences between Blacks and Whites Thomas P. Vartanian1 Page Walker Buck Bryn Mawr College Graduate School of Social Work Bryn Mawr, PA 19010 and Philip Gleason Mathematica Policy Research, Inc. Princeton, NJ Presented at the Association for Public Policy Analysis and Management November 4, 2005 (DRAFT: Not for citation)

1

All correspondence to Thomas P. Vartanian at [email protected], 610-520-2624, Bryn Mawr College, 300 Airdale Road, Bryn Mawr, PA, 19010.

Intergenerational Neighborhood Mobility

1

Abstract Using sibling data from the Panel Study of Income Dynamics linked with U.S. Census data, this work extends the research on neighborhood effects through an examination of intergenerational neighborhood mobility. We estimate linear and non-linear relationships between childhood and adult neighborhood conditions and use these relationships to determine whether children who grow up in the most disadvantaged neighborhoods are likely to remain in disadvantaged neighborhoods as adults. Further, we examine the extent to which intergenerational neighborhood mobility differs by race. Our results indicate that childhood neighborhood conditions of black and white children are substantially different from one another. Few whites live in the most disadvantaged neighborhoods, and few blacks live in the most advantaged neighborhoods. We also find that childhood neighborhood quality has a positive effect on adult neighborhood quality. For whites, this relationship is nonlinear and strong - those children who grow up in the most disadvantaged neighborhoods are far more likely to live in the poorest neighborhoods during adulthood relative to those children who grow up in only slightly better neighborhoods. The relationship between childhood neighborhood quality and adult neighborhood quality is much less strong for blacks. However, our simulations show that if poor blacks in the most disadvantaged black neighborhoods instead grew up in poor white neighborhoods, their chances of living in these poorest neighborhoods as adults would dramatically decline.


2

Introduction The issue of neighborhood effects has been a focus of much social science inquiry over the last decade, spurred largely by William Julius Wilson’s (1987) The Truly Disadvantaged. Wilson argues that contextual effects play a vital role in preventing low-income African Americans from getting good jobs and escaping poverty, with chronic unemployment and social isolation from the mainstream economy causing a breakdown of social function. Poor neighborhood conditions have been hypothesized to have deleterious effects on virtually all forms of human and social capital, including education, fertility, income, welfare use, health outcomes, social support networks and civic engagement (Borrell, Diez Roux, Rose, Catellier, & Clark, 2004; Duncan, Connell, & Klebanov, 1997; Epstein, 2003; Jencks & Mayer, 1990; McClintock, 2005; Vartanian & Buck, 2005; Wilson, 1987). While most of this literature has focused on child and adolescent outcomes, there is a body of evidence on the effects that neighborhood characteristics have on adults. For example, Mendenhall, DeLuca and Duncan (2005) find that the level of neighborhood resources has significant effects on women’s welfare receipt and employment. Similarly, Weinberg, Reagan and Yankow (2002) find that neighborhood characteristics are significantly related to annual hours of work. McClintock (2005) and others (Epstein, 2003; Borrell et al., 2004) have explored the link between neighborhood conditions and health outcomes, suggesting that high crime rates and poor housing conditions significantly increase the stress levels of residents. Such stress, caused by the need to be hypervigilant given dangerous living conditions, is believed to wear down the immune system to such an extent that residents in such neighborhoods may be more prone to illness. One important question that arises from this literature is what childhood factors lead people to live in these types of disadvantaged neighborhoods as adults. More specifically, how likely are children who grow up in poor neighborhoods likely to end up in the same kinds of neighborhoods as


3

adults? In what follows, we analyze data that provides a basis for answering this question through an examination of the intergenerational qualities of neighborhood-type mobility. Based on a review of the existing literature, we posit that childhood neighborhood conditions and adult neighborhood conditions are positively related, and we estimate this relationship to examine whether it is linear or non-linear. These estimates tell us whether growing up in a good neighborhood makes someone more likely to be in a good neighborhood as an adult. Conversely, this also tells us whether those children who grow up in the poorest neighborhoods are likely to be “trapped” in similar neighborhoods as adults. 2 Further, we examine the extent to which intergenerational neighborhoodtype mobility differs by race. The challenge of such a study is confounded by the need to control for factors that lead certain families to live in certain types of neighborhoods. In particular, individuals and families are sorted into neighborhoods for unobserved, and thus unmeasured, reasons. Failing to account for these unobserved determinants of neighborhood residence may lead to biased estimates of neighborhood effects. For example, families with the best networks of social connections may live in the best neighborhoods, and these connections may lead to a variety of other positive outcomes. This creates a positive correlation between neighborhood conditions and other outcomes, even if there is no true neighborhood effect. To address these issues of selection bias, we using sibling fixed effect models with data from the Panel Study of Income Dynamics linked with U.S. Census data.

Previous Literature

2

The use of the term “trapped” is taken from Gramlich, Laren and Sealand’s (1992) theory of entrapment and refers, in this case, to the likelihood that children raised in disadvantaged neighborhoods will end up in similar types of neighborhoods as adults.


4

While there is no body of literature on intergenerational neighborhood quality in particular, we locate our study within the research on intergenerational socioeconomic mobility, neighborhood effect theories and residential mobility. What follows is a review of these literatures as they relate to this study of how childhood neighborhood quality helps to predict adult outcomes.

Intergenerational Socioeconomic Mobility Intergenerational socioeconomic mobility, and the lack thereof, has been the longstanding focus of much scholarly work, prompted initially by Blau and Duncan’s (1967) seminal work, American Occupational Structure. While there is no consensus among social scientists about how much socioeconomic mobility exists in the U.S., and even less about its theoretical basis, researchers continue to try to capture those factors that enhance or impede people’s chances in life (Piketty, 2000). Studies have shown that intergenerational income correlations typically fall somewhere in the range of 0.35 to 0.49 (Chadwick and Solon, 2002; Lee and Solon, 2005; Solon, 1992; Zimmerman, 1992). Earnings correlations have been shown to be as high as .60 (Mazumder, 2001), while wealth correlations before bequests have been found to be .37 (Charles & Hurst, 2003). These pooled estimates of intergenerational mobility, however, disguise the highly differential effects of race. Hertz (2005) suggests that average mobility correlations of .40 are driven primarily by low upward mobility of poor African Americans. He found that African American children born into the poorest families had a 42 percent chance of ending up in similarly poor circumstances as adults, compared to a 17 percent chance among white children. Similarly, African American children in the bottom income quartile were only half as likely as White children to end up in the top income quartile. Absent from this mobility literature is the study of intergenerational neighborhood mobility – the extent to which escape from disadvantaged childhood neighborhoods is possible in


5

adulthood. While the probability of this type of escape has been studied contemporaneously (Crowder & South, 2005; Massey, Gross & Shibuya, 1994; South & Crowder, 1997; Quillian, 1999), we know of no studies that examine the intergenerational nature of neighborhood mobility. As Bowles and Gintis (2002) suggest, geographic location is one of the important, yet understudied, elements of the intergenerational transmission of economic success. That the potential effects of neighborhood residence range from education and employment to mental and physical health outcomes (Borell et al., 2004; McClintock, 2005) suggests that this study of residential context across the lifespan is vitally important to the larger study of socioeconomic mobility. It may well be that the ability to move out of a distressed neighborhood is a significant factor in predicting one’s chances to move up the economic ladder.

Neighborhood Effect Theories The ways in which neighborhood effects exert influence on individuals are best described by two theories: neighborhood advantage theory and epidemic theory. The first, the neighborhood advantage theory (Vartanian & Buck, 2005), combines several aspects of the theories originally defined by Jencks and Mayer (1990): collective socialization theory, social isolation theory, and neighborhood institutional resource theory. Neighborhood advantage theory suggests that the greater the resources and other advantages of good neighborhood conditions during childhood, including exposure to more positive role models and more abundant institutional resources, the better the adult outcomes. At the same time, negative neighborhood conditions resulting from social isolation from such positive role models and community resources, coupled with a higher exposure to crime, have negative intergenerational effects. For children, disadvantaged neighborhoods may lack types of formal community resources such as good schools, while for adults they may lack the social networks often necessary for employment (Mendenhall et al., 2005). Crime rates also tend to


6

be very high in these areas, draining personal and community-level resources. Evidence of neighborhood advantage has been shown in several studies (Aaronson, 1998; Harding, 2003; Kling, Liebman & Katz, 2005; Vartanian & Buck, 2005). Second, the epidemic theory stems from Wilson (1987) and has been extended and tested by Crane (1991). The theory suggests that neighborhood effects act in non-linear ways similar to epidemics. In most types of neighborhoods, the effects of changing conditions are modest. When poor neighborhood conditions reach a certain level or tipping point, however, the negative effects of such conditions become highly contagious. A number of studies find evidence of this type of nonlinear neighborhood effect on a range of adolescent and adult outcomes (Duncan, Connell, & Klebanov, 1997; Galster, Quercia, and Cortes, 2000; Mendenhall, DeLuca & Duncan, 2005; Vartanian, 1999; Vartanian & Buck, 2005). The investigation of non-linear neighborhood effects has significant policy relevance if, as Galster (2004) suggests, overall social well-being can be improved when the well-being of only a small fraction of households is increased.

Methodological Issues Estimating causal inference in neighborhood studies is confounded by the related issues of simultaneity, omitted variable bias, and endogenous membership (Leventhal & Brooks-Gunn 2000; Duncan & Raudenbush 2001; Dietz 2002). The basic challenge is that individuals and households are not exogenously placed in particular types of neighborhood, but live there either out of choice or because of their characteristics. Thus, when differences in later outcomes are observed between individuals who live (or who lived) in different types of neighborhoods, it is difficult to distinguish whether these differences in outcomes are caused by the neighborhood itself or by the pre-existing differences in the types of individuals and households that live in different neighborhoods (sometimes referred to as selection bias). Failing to account for such issues may lead to biased


7

estimates of neighborhood effects (Ginther, Haveman, & Wolfe, 2000). In particular, it is possible that previous studies that find strong positive neighborhood effects—with good neighborhood conditions leading to positive outcomes—suffer from a positive selection bias, leading to questions about the existence of neighborhood effects at all. Despite criticism that much of the research on neighborhood effects does not adequately address these issues (Durlauf, 2003; Ellen & Turner, 1997), recent studies have used a variety of interesting techniques to address selection bias. The strongest design has been based on experimental data, but experiments involving individual or household residential locations are typically costly and difficult to implement. Non-experimental studies have used basic linear and non-linear regression models, propensity score matching, and sibling fixed effect models to attempt to obtain unbiased estimates of neighborhood effects. While there is no broad consensus on the effects of neighborhoods on individual outcomes, studies employing these methods have continued to find evidence of neighborhood effects for a range of outcomes (Harding, 2003; Kling, Liebman & Katz, 2005; Vartanian & Buck, 2005). Residential mobility projects using experimental data provide evidence for the effects of neighborhoods on child and adult outcomes. The Moving to Opportunity (MTO) project is the largest and most studied of the mobility experiments, where public housing residents entered into a lottery to receive vouchers for housing in non-poor neighborhoods. The most recent evidence using these data (Kling et al., 2005) shows that female youth benefited the most from moving to advantaged neighborhoods, with significant positive effects on mental health, physical health, educational outcomes, and risky behaviors. For male youth, the effects of moving to a better neighborhood were negative, suggesting that gender is a significant factor in the ways in which youth respond to new neighborhood conditions. The effects for adults were limited to improvements in mental health status.


8

Observational, or non-experimental, studies using data from longitudinal datasets such as the Panel Study of Income Dynamics (PSID) also show that neighborhood effects are robust in models that control for endogenous neighborhood selection. Harding (2003) uses propensity score matching with PSID data to estimate neighborhood effects on educational and fertility outcomes. He finds that of two groups of children who are matched on a large number of characteristics at age 10 but who experience different neighborhood conditions, those who live in disadvantaged neighborhoods during adolescence have worse educational and fertility outcomes than those who live in better neighborhoods during adolescence. To assess the role of selection bias in this estimate, he conducts a sensitivity analysis in which a hypothetical selection effect is simulated and its influence on the neighborhood effects estimates is observed. This sensitivity analysis shows that the basic neighborhood effects are relatively robust to moderately strong selection bias. Other studies have attempted to account for selection bias using instrumental variables (Duncan, Connell and Klebanov, 1997; Evans, Oates & Schwab, 1992), in which neighborhood characteristics are “instrumented” by factors correlated with neighborhood residence but not related to the outcomes of interest. While this type of model addresses selection bias in theory, its assumptions are strong and difficult to test, and the results of selection bias models are often quite sensitive to the particular assumptions being made (Ellen & Turner, 1997; Harding, 2003). Fixed effect models with sibling data use residential migration and changes in neighborhood conditions to compare outcomes among siblings who experienced different neighborhood conditions while growing up (Plotnick & Hoffman 1995, 1999; Aaronson 1998; Vartanian & Buck, 2005).3 By comparing siblings, these models hold time invariant observed and unobserved family characteristics constant, or fixed. Other than the Plotnick and Hoffman (1999), who find no effects in their fixed effects models, results from other such studies show positive neighborhood effects. 3

Also see Levy & Duncan 2000 for a comparison of OLS and FE estimates in a model of the relationship between family income and educational outcomes.


9

Vartanian and Buck (2005), for example, find that the effects of childhood neighborhood residence on adult economic outcomes are positive and similar (or larger) in fixed effect models than in conventional OLS regression models. Aaronson (1998) also finds evidence of this type of negative selection -- effects are larger in the fixed effects models that control for selection bias than in OLS models that do not. The potential drawbacks of using fixed effect models with sibling samples are worth noting. In particular, only certain unobservable factors can be controlled in fixed effect models, including factors that do not vary among siblings. Unobservable factors that do vary among siblings, such as child ambition, child IQ, and parental treatment of the child, are not controlled in these models. This is a potentially large drawback to fixed effect models, given recent sociological work that finds that sibling inequality is sometimes quite great, especially in poor families (Conley, 2004). Fixed effect models are also vulnerable to endogenous effects if the reasons a family chooses a neighborhood are related to factors that affect their children’s outcomes (Duncan & Raudenbush, 2001). Finally, these sibling fixed effects models can be estimated using only data from families with multiple children and in which neighborhood conditions vary from child to child. Families with a single child or in which all children experienced similar neighborhood conditions do not contribute to these estimates.

Residential Mobility Theories of residential mobility suggest that the ability to move to a more desirable neighborhood is a function of three primary factors: human capital, life-cycle development, and place stratification (Verma, 2003). Human capital, including education, income, and stable employment, as well as life cycle factors such as age, marital status and fertility are often considered to be the foundations of residential mobility. Whereas increases in socioeconomic status


10

have been found to promote mobility into advantaged neighborhoods (Rossi, 1980; South & Crowder, 1997), welfare receipt, public housing (Kasarda, 1988), and homeownership (South & Crowder, 1997) have been found to be limiting. These frequently cited factors, however, do not operate equally across races. African Americans, for example, are less likely to translate human capital into residential mobility (South & Deane, 1993) due to what some suggest is place stratification (South & Crowder, 1997). That is, neighborhoods are sorted along racial and ethnic lines, such that minorities end up living in the most disadvantaged areas because of social preferences, hierarchies, and discrimination (Logan & Alba, 1993). African Americans, in particular, live in significantly worse neighborhoods than do whites (Crowder & South, 2005; Massey et al.,1994; Logan & Alba, 1993; Quillian, 2003; Timberlake, 2002) – a trend that has persisted despite a significant decline in the concentration of poverty during the 1990s (Jargowsky, 2003). As Timberlake (2003:1) explains, “black and white children on average live in neighborhoods with vastly unequal levels of social, economic and physical resources.” Research has shown that African Americans tend to live in poor neighborhoods, due in part to low likelihoods of exiting and high likelihoods of reentering (Crowder & South, 2005; Gramlich, et al. 1992; South & Crowder, 1997; Quillian, 2003). Residential segregation is further maintained in large part by the fact that whites overwhelmingly avoid predominantly African American and racially mixed neighborhoods (Massey & Denton, 1993; Quillian, 2002), and are more able and more likely to move out of poor neighborhoods (South & Crowder, 1997). However, while considering both individual and structural factors in the study of residential mobility, none of this research takes into account the impact of childhood residence on adult neighborhoods. Net of individual and structural level factors, to what degree does childhood neighborhood quality affect where one lives as an adult? We suggest that the answer to this intergenerational research agenda


11

adds an important new dimension to the existing body of work on socioeconomic mobility in the U.S., with implications for both policy and practice.

As indicated earlier, this study seeks to answer three primary questions. First, is there an intergenerational component to neighborhood quality whereby the type of neighborhood a person lives in as a child influences the type of neighborhood he or she lives in as an adult? Second, does the relationship between child and adult neighborhood quality vary by race? Third, is this relationship non-linear? More specifically, is there any evidence of the epidemic theory of neighborhood effects, with those in the worst neighborhoods most strongly influenced by these neighborhood conditions?

Methods The following equation shows a model of the determinants of adult neighborhood conditions:

Yij = α j + β 1 FPj + β 2 FIVij + γN ij + µ ij Here, FPj represents a set of observed permanent characteristics of family j, FIVij represents a set of time varying characteristics of family j and individual i, Nij represents characteristics of the neighborhood that individual i lived in as a child, µij is the error term, β1, β2, and γ are the coefficients for the respective permanent family, varying family and individual, and neighborhood variables, and αj is the family-specific intercept. This intercept, or a fixed family effect, represents unobserved time invariant family characteristics that influence adult neighborhood conditions The dependent variable in the model, Yij, is a measure of the quality of the neighborhood that individual


12

i from family j lives in as an adult. The key parameter that we will be estimating in this model is γ, the neighborhood effect4. Under the assumption that there are no family fixed effects (in other words, that αj = α ), this model can be estimated using ordinary least squares (OLS).5 If, however, there are family fixed effects, OLS estimates of the neighborhood effect will be biased. To estimate this fixed effects model, we use data from individuals in families with more than one child, and subtract the family mean values from each individual sibling’s observation, as shown below:

Yij − Y. j = (α j − α j ) + β 1 ( FPj − FP. j ) + B2 ( FIVij − FIV. j ) + γ ( N ij − N . j ) + ( µ ij − µ . j ) . Because the fixed effect does not vary across siblings, it drops out of the model, and we can effectively control for this fixed effect even though it is unobserved.6 In addition, the observed permanent family characteristics also drop out of the model. In the model described above, childhood neighborhood conditions are entered into the model linearly. We estimate this linear model, but also estimate models in which childhood neighborhood conditions are allowed to influence adult neighborhood conditions non-linearly. In particular, we estimate an alternative model that includes quadratics and splines for the childhood neighborhood index variable.

4

Appendix Table A1 shows the within-family variation in the neighborhood variables for Blacks and Whites. The first column for each race indicates the percentage of siblings that had within-family differences in each variable. The second column indicates the variables’ mean differences and standard deviations for siblings. The third column indicates what percentage of the variance in the variable is accounted for by differences within the family. Results show that almost all families report some variation in their neighborhood variables during childhood, but that the differences generally are not large. Likewise, results suggest that only a small proportion of the total variance in any of the neighborhood conditions during childhood is explained by within-family differences. A far larger proportion of the variance of adult neighborhood conditions is explained by within-family differences. 5 The OLS models use robust standard errors for coefficient estimates to account for the non-independence of the observations (i.e., siblings). 6 As noted previously, this model does not deal with the potential issue of unosbserved time-varying differences between siblings that are both related to their childhood neighborhood conditions and that influence the outcome of interest (Aaronson, 1998; Levy and Duncan, 2000; Vartanian and Buck, 2005). Differences may arise from differences in parental aspirations for their children or changes in parents’ emotional states. It may also be that parenting is a learned process (Aaronson, 1998) whereby younger children may benefit from parental experience. Alternatively, parents may favor one child over another or use scarce resources for the benefit of the most promising child (Conley, 2004), which could influence both the type of neighborhood the child lives in and the later outcome for that child. To partly address this issue, we follow Aaronson (1998) and include controls for the birth order of the child.


13

Spline regressions fit a regression equation into a series of linear segments, and each segment may have a different slope (Galster et al. 2000; Marsh and Cormier, 2001). We have set cutoff points at the top 10 percent of neighborhoods, the 11th to the 25th percentiles, the 26th to the 50th percentiles, the 51st to 75th percentiles, the 76th to the 90th percentiles, and at the 91st to 100th percentiles. There would be evidence for the epidemic theory if the poorest neighborhoods (i.e., 91st to 100th percentile) were to have the most detrimental effects on future neighborhood conditions. Childhood neighborhood characteristics might influence adult neighborhood conditions directly, or their influence may be indirect, through a preliminary impact on such outcomes as education or work experience. For example, bad neighborhood characteristics in childhood could lead to fewer years of education, leading in turn to living in more disadvantaged neighborhoods as an adult. To assess whether the estimated relationship between childhood and adult neighborhood conditions is direct or indirect, we estimate the basic model both with and without control variables that reflect adult outcomes other than neighborhood conditions. In the models that exclude these adult outcomes, we estimate overall neighborhood effects–both direct and indirect. In the models that include these other adult outcomes, we estimate only direct neighborhood effects. Finally, we estimate each of our models separately for white and black individuals. Separate models are estimated because of the vastly different types of neighborhoods they experience as children, as well as previous research showing different types of neighborhoods for the two groups.

Data and Variables

We use data from the Panel Study of Income Dynamics (PSID), a nationally representative data set that began with interviews of approximately 5,000 families in 1968. The heads of the original households have been interviewed every year since 1968, as have the heads of households containing members who were part of one of the original households (in 1968) and who have since left that household to join another or to start one of their own. Among the original households, the


14

poor and African Americans were oversampled.7 This work uses sample waves through 2001. The 2001 PSID contains over 7,400 families. A secondary source of data is the PSID Geocode File, which allows for the linking of census data with PSID respondents. Census data is the source of information on the characteristics of PSID respondents’ neighborhoods, operationally defined as their census tracts.8 This file contains 1970, 1980, 1990, and 2000 census data on factors such as the poverty rate, the proportion of female headed households, and the proportion of households receiving public assistance income, for the census tract in which each PSID respondent lived during each year of the survey. To be included in the sample, respondents and their sibling(s) must have at least four years of childhood (age 0 to 18) data and at least one year of adult (over age 24) data. In some cases we examine all 18 years of childhood data, while in others it is only four. Kuntz, Page, and Solon (2001) find that because childhood neighborhood characteristics are highly correlated across childhood years, even having a single year of neighborhood information produces only small errors-invariables bias. Respondents must have at least one year of adult data, although some have more than 20. In all cases, we average childhood and adult variables over the period of observation to minimize the effects of single-year outliers. In models that include adult variables, we include the

7

Weights adjust for this oversampling and for differential rates of attrition in reported mean values and standard deviations. 8

Although census tracts are not necessarily synonymous with neighborhoods, they are generally regarded as the best available proxy for neighborhoods. Neighborhoods have been defined as census tracts in a number of previous studies (e.g., Brooks-Gunn et al. 1993; Ginther et al. 2000; and Plotnick and Hoffman 1999). The average number of people living in a census tract is approximately 4,000. To the extent that census tracts do not correspond to true neighborhoods, the resulting measurement error will lead to a downward bias in the estimate of neighborhood effects. When census tract data are not available, the neighborhood is defined as the next lowest level of geography. Minor civil divisions are first examined for valid data, and if missing, zip code data are used. Neighborhood characteristics from the 1970 Census are used for the years 1968-1975. Those from the 1980 Census are used for the years 1976-1985. Those from the 1990 Census data are used for years 1986-1999. The 2000 Census is used for PSID year 2001. The PSID does not contain links for 2000 Census for years before 2001. We used samples that did not contain neighborhood information for PSID years 1996-1999, to consider the possible confounding effects of using the 1990 Census for years closer to the 2000 Census than the 1990 Census, to determine if results changed markedly without these years. We did not find large differences when we deleted these years from our adult neighborhood averages.


15

maximum age as an adult (including its quadratic) to control for the effects that work history, experience, and earnings potential could have on the ability to live in a more affluent neighborhood. The white sample consists of 2,265 observations and the black sample consists of 1,863 observations. All observations have siblings in the sample in order to make the fixed effect and OLS models comparable. Several neighborhood variables are used in the analysis to measure neighborhood quality. These variables include the neighborhood poverty rate, the percentage of households receiving public assistance income, the percentage of households headed by females, the percentage of households with income below $15,000 (in 2001 dollars), the percentage of households with income above $60,000 (in 2001 dollars), the percentage of households with income above the respondent’s average family income, and the percentage of households with income at the same level of income as the respondent’s average family income.9 These types of variables have been used by previous researchers to reflect the conditions and quality of the neighborhood (Plotnick and Hoffman, 1999; Ginther et al., 2000). We use these same variables to characterize both childhood and adult neighborhood quality. To serve as summary indicators of neighborhood quality, we created index variables through principal components analysis for both childhood and adulthood neighborhood quality. Each index takes into account information from the seven neighborhood variables for all observations. Details of the construction of the neighborhood quality index are shown in the Appendix Table A2. High quality neighborhoods, both in childhood and adulthood, have a high proportion of residents with high income, a low proportion of residents with low income, a low poverty rate, few households receiving public assistance, and primarily two-parent households. Because of the way both the 9

Each variable is measured as the average value of the characteristic over the childhood years. Among those children who moved away from their parents before age 19, however, only the years that they lived with their parents are used in this calculation.


16

childhood and adult neighborhood index variables were created, low values of the index represent high quality neighborhoods, and vice versa.10 The dependent variable is the index of adult neighborhood conditions at age 25 and beyond. The dependent variable is created using neighborhood values from all observations and is then examined in separate white and black samples (see Appendix Table A2). The key independent variable in the model is a similarly constructed index of childhood neighborhood conditions. Other independent variables in the model include controls for factors such as region of the country, marital status, the log of family income-to-needs, housing status, work limitations of the head of household, age of the youngest child, gender, age when the child entered the sample, year the child entered the sample, number of years as a child in the sample, percentage of years the family moved residence, and birth order. We also control for variance within childhood for both income and neighborhood conditions (see Levy & Duncan, 2000) to determine the effects of these fluctuations. In the first set of multivariate models, only childhood variables are included. In a second set of multivariate regressions, variables from the respondents’ adult years are included in order to control for a number of personal, educational, and economic outcomes. See Appendix Table A3 for a complete list of control variables and the models in which they are used.

Simulations

To help us interpret the estimation results and illustrate the implications of these results, we conduct a series of simulations. These simulations are designed to address a set of hypothetical questions regarding how a given individual’s adult neighborhood conditions might be expected to change if they had experienced different neighborhood conditions while growing up.

10

We also created an index for the level of variation in neighborhood conditions over the childhood years, based on variation in each of the seven neighborhood variables across the childhood years of each individual in the sample. See Appendix Table A2.


17

In particular, we conduct two types of simulations. The first type of simulation examines how the average sample member would be affected by growing up in different types of neighborhoods. For example, if the average white child grew up in a very bad neighborhood (such as the worst 1 percent of neighborhoods of white children in the sample), what would be the likelihood that that child would, as an adult, live in an equally bad neighborhood? What would be the likelihood that he or she would live in very good neighborhood as an adult (say, a neighborhood at the top 5 percent of the distribution)? Conversely, we could conduct a simulation that would place that same child in the best of all neighborhoods while growing up, and ask the same set of questions about the type of neighborhood he or she would likely live in as an adult. If there is no relationship between childhood and adult neighborhood conditions, the answers to the questions about the likelihood of living in a very good neighborhood as an adult will be the same regardless of the type of childhood neighborhood into which our simulations placed the child. If there is a strong positive relationship, on the other hand, then the likelihood of living in a very good adult neighborhood will be much higher for the child simulated to have grown up in a very good neighborhood. The second type of simulation we conduct addresses the implications of neighborhood segregation. As described below, white children and black children currently grow up in very different types of neighborhoods. To assess how this influences the types of neighborhoods these groups live in as adults, we perform the following thought experiment. Suppose that white children grew up in the very same neighborhoods that black children now grow up in. How would this influence the types of neighborhoods that the white children lived in as adults? And, what would happen if black children grew up in the same neighborhoods that white children grow up in? These simulations were conducted using the following five-step procedure: 1.

Generate estimates of the model parameters, or coefficients.

Intergenerational Neighborhood Mobility 2.

18

For each individual in the sample, multiply those coefficients by the individual’s values of all the associated variables in the model, except for the individual’s childhood neighborhood quality index. Instead of multiplying the individual’s value of this variable by its associated coefficient, use the simulated childhood neighborhood index. Sum the products of the variables in the model and their associated coefficient estimates to generate a predicted value of the outcome variable (the adult neighborhood index) under this simulation.

3.

Simulate a value of the error term for each individual by selecting a random draw from a normal distribution with a mean of 0 and variance of sigma squared (σ2) (the error variance obtained from the model estimation).

4.

Sum the predicted value of the outcome variable under this simulation (#2) and the simulated error term (#3) for each individual. This is the simulated value of the adult neighborhood index for each individual.

5.

Calculate statistics based on the distribution of this simulated neighborhood index across all individuals in the sample.

The procedures for conducting the first and second types of simulations differ only in terms of how the simulated childhood neighborhood index value is created. In the first type of simulation, we simply pick a single neighborhood index value and use it for all sample members. For example, if we want to simulate the implications of all children growing up in very bad neighborhoods, we might choose a childhood neighborhood index value that is in the worst percentile of the distribution and assign each individual in the sample that index value. In the second type of simulation, the simulated childhood neighborhood index value varies across children. In particular, for the simulation in which we assign to white children the


19

neighborhood conditions experienced by black children, we assign to white children living in the very worst neighborhoods (among whites) the neighborhood index value of the corresponding black children living in the very worst neighborhoods among blacks. For example, all white children in the bottom 5 percent of the distribution of the childhood neighborhood index among whites are assigned the 2.5th percentile of the neighborhood index among black children. Conversely, white children living in the best neighborhoods are assigned the neighborhood index values of the corresponding black children living in the best neighborhoods.

Results

In this section, we present our results in three parts. First, we describe types of neighborhoods in which black and white sample members live, both as children and as adults. Second, we present the results of our multivariate analyses, showing our estimates of the relationship between childhood and adult neighborhood characteristics separately by race. This presentation includes examining differences between our model specifications, such as OLS and fixed effects (FE) models, as well as linear versus non-linear specifications. The results from the FE models are then used in simulation models that further illustrate the implications of our model estimates. Table 1 shows the distribution of neighborhood types for black and white sample members in both childhood and adulthood. Neighborhood type is defined according to the distribution of the neighborhood index variables across the full sample (including both black and white sample members). For example, a given child is defined to live in the top decile of neighborhoods if the neighborhood index for that child is among the top ten percent of all sample members.


20

African Americans and whites live in very different types of neighborhoods. Whites are very unlikely to live in the most disadvantaged neighborhoods while blacks are very unlikely to live in the most advantaged neighborhoods, both in childhood and as adults. For example, almost half of all whites (49%) live in the most advantaged quartile of neighborhoods during childhood, while less than 1 percent of blacks live in such neighborhoods. During adulthood, whites have around a 44 percent chance of living in the most advantaged quartile of neighborhoods while blacks have a 4 percent chance of living in such neighborhoods. The opposite patterns emerge in the most disadvantaged neighborhoods. Whites have a 3 percent chance of living in the most disadvantaged quartile of neighborhoods during adulthood, while blacks have a 51 percent chance of living in such neighborhoods. Almost no whites live in the most disadvantaged decile of neighborhoods and almost no blacks live in the most advantaged decile of neighborhoods, either as children or adults. Table 2 shows the simple relationship between childhood and adult neighborhood characteristics by race. In particular, the columns represent the types of neighborhoods children grow up in, and within each column, the rows show the distribution of neighborhood conditions that this group of children experience as adults. The initial column shows, for example, that among white children who grow up in the top decile of neighborhood quality, 31 percent are in the top decile of neighborhoods as adults, while 35 percent are in the next best category of neighborhoods (11th to 25th percentiles), 29 percent are in the next category (26th to 50th percentiles), and 4 percent are in a neighborhood with an index value below the median. Table 2 shows a modest relationship between childhood and adult neighborhood conditions. Among both blacks and whites, those who grow up in relatively good neighborhoods are more likely to end up in relatively good neighborhoods as adults than those who grow up in lower quality neighborhoods. However, interpreting this table is made somewhat more challenging by the fact


21

that white sample members are concentrated almost entirely in the top half of the distribution while black sample members are concentrated in the bottom half. To address this issue, we created separate white and black neighborhood quality indexes. The index created using the white sample defines neighborhood quality only as it is experienced by white children and adults. The index created using the black sample analogously captures neighborhood conditions experienced by African American children and adults. Thus, while 10 percent of white children have values in the bottom decile of the white childhood index and 10 percent of black children have values in the bottom decile of the black index, the actual neighborhood conditions represented by the bottom decile of the white and black indexes are very different from one another. Table 3 shows the relationship between childhood and adult neighborhood quality using these race-specific neighborhood indexes. The table shows that for both groups, those who grow up in the most advantaged neighborhoods have a relatively high likelihood of staying in those neighborhoods as adults. This pattern also holds for those who grow up in both average and more disadvantaged neighborhoods. However, there are some notable differences between whites and blacks. Among whites who grow up in the most disadvantaged decile of neighborhoods (based on the white neighborhood index), 41 percent end up in that type of neighborhood as an adult. For blacks, the analogous likelihood is 24 percent. Similarly, whites who grow up in the top 76th to 90th percentile of neighborhoods according to the white neighborhood index have a 34 percent chance of ending up in such neighborhoods as adults, while for blacks, this value is only 16 percent. These results give some evidence that blacks are less likely to be “trapped” in the most disadvantaged race-specific neighborhoods, but it should be noted that if we defined bad neighborhoods by using the white (or overall) neighborhood index for the black sample, we would


22

conclude that the vast majority of African Americans are in bad neighborhoods as both children and adults. Table 4 shows the weighted characteristics of the residents of different race-specific neighborhood-types during childhood. One characteristic that stands out in this table is the likelihood that the family owns a home for the majority of a sample member’s childhood. Blacks in the top decile of neighborhoods have a 57 percent likelihood of owning their own home, compared with 93 percent among whites. The likelihood of owning a home decreases as neighborhood-type becomes more disadvantaged, and reaches 13 percent for blacks and 63 percent for whites in the most disadvantaged neighborhoods. Differences in income levels are also notable. Even in the most advantaged neighborhoods (for black sample members), black income levels are on average only 70 percent above the poverty line, compared to whites who have incomes that are on average 278 percent above the poverty line in the best neighborhoods among white sample members. Looking at these income discrepancies another way, we find that whites living in the most disadvantaged white-type neighborhoods have income levels (81% above the poverty line) that are comparable to the income levels of blacks living the most advantaged black-type neighborhoods. Neighborhood conditions for whites and blacks also differ considerably for those living in the most advantaged white-type and black-type neighborhoods. For example, the poverty rate in the best white-type neighborhoods is around 3 percent, while for blacks, the poverty rate in the best black-type neighborhoods is around 12 percent. The poverty rate for white-type neighborhoods does not reach close to 12 percent until the 51st to 75th percentile. The percentage of households who are headed by women is roughly two to three times as high for blacks as whites, in comparable neighborhood categories.


23

Also notable is the fact that blacks living in the most disadvantaged neighborhoods tend to live in highly populated areas, whereas whites who live in the most disadvantaged neighborhoods tend to live in rural areas. For example, 58 percent of whites who live in the lowest decile of neighborhoods live in cities/towns outside of standard metropolitan statistical areas (SMSAs), with populations of less than 25,000 people. Comparatively, only 8 percent of blacks live in these rural types of areas. Both blacks and whites who live in the most advantaged neighborhoods tend to live in highly populated areas. For example, 98 percent of whites and 78 percent of blacks who live in the top decile of race-specific neighborhood types live in SMSAs with city populations of 100,000 or more.

Correlations and Multivariate Regression Results

We present correlations between childhood and neighborhood characteristics in Table 5, followed by key results from estimation of the OLS and fixed effect models for blacks and whites in Tables 6 and 711. In interpreting these results, it is important to keep in mind that higher values of the neighborhood index represent more disadvantaged neighborhood conditions (higher poverty rates, fewer two-parent families, and so on). Thus, the 91st to 100th percentile of the childhood neighborhood index shown in the spline specification represents the worst 10 percent of neighborhood conditions. In the regression models, positive linear neighborhood effects can be interpreted to mean that the more disadvantaged the neighborhoods in childhood, the more disadvantaged the neighborhood in adulthood (ceteris paribus). For the quadratic models, we use two squared quadratic terms, one positive (for those who have positive [worse] neighborhood index

11

For a full list of control variables used in our models, see Appendix Table A3.


24

values as children) and one negative (for those who have negative [better] neighborhood index values as children).12 Table 5 shows the relationship between the childhood and adult neighborhood indexes in simple correlations and partial correlations that include a single control for age at the end of the sampling period. These results indicate that childhood and adult neighborhood conditions are more highly correlated for whites than for blacks. The simple correlation approaches 0.50 for all whites regardless of whether age is controlled. For blacks, the overall correlation is 0.30, with a slightly higher correlation once age is controlled. The correlations for lower income sample members (regardless of race) are similar. These correlations reinforce the results of our earlier crosstabulations (see Table 3) in suggesting that there is somewhat less intergenerational neighborhoodtype mobility for whites than for blacks. Compared with white sample members, in other words, among the neighborhoods in which black sample members are likely to live, a black child is more likely to move from a bad neighborhood in childhood to a better neighborhood as an adult. These simple correlations, however, do not control for individual and family characteristics that could potentially influence adult neighborhood type more strongly than childhood neighborhood type. Table 6 shows the OLS and FE regression results, using linear, spline, and quadratic neighborhood models, which do control for these individual and family characteristics. The top panel of results excludes variables representing adult outcomes such as family structure and educational attainment that could mediate the effects of childhood neighborhood conditions on adult neighborhood conditions. Among both white and black sample members, the results of the FE models for the spline and quadratic models are substantially different than the results of the OLS models, suggesting that the unobserved permanent family characteristics controlled for in the FE 12

The neighborhood index takes on a roughly equal number of positive and negative values across the full sample. Among white sample members, however, the index is negative for most sample members, indicating better neighborhood conditions. Conversely, most black sample have positive values of the neighborhood index, indicating more disadvantaged conditions.


25

models but not in the OLS model are important. Thus, we focus primarily on the results of the FE models here. For whites, the results suggest that childhood neighborhood quality has a nonlinear effect on adult neighborhood quality. Although the coefficient on the childhood neighborhood quality index is positive and significant in the linear model, evidence from both the spline and quadratic specifications indicate that this neighborhood effect is actually nonlinear. In particular, the spline specification shows little evidence of a relationship between childhood and adult neighborhood conditions among whites, except for those who grew up in the bottom decile of neighborhoods. For this group, the spline coefficient is positive and highly significant. The quadratic FE results for whites are consistent with these spline results. For blacks, we also find some evidence of nonlinear neighborhood effects, although these effects are somewhat hard to interpret. In particular, there is evidence of both positive and negative effects as neighborhoods become more disadvantaged. We find similar results when using controls for other adult outcomes in the bottom panel of Table 6. The effects of childhood neighborhood conditions and their general significance levels stay about the same as in the top panel of the table. This suggests that the neighborhood effects shown in the top panel are not operating through the types of adult outcomes we include in the model in the bottom panel. These estimated effects are either direct effects of childhood neighborhood conditions on adult neighborhood conditions, or they operate through other adult outcomes not controlled for in the bottom panel of Table 6. Table 7 shows the same models as Table 6, but only for a sub-sample of those with incomes at or below 150 percent of the poverty line. This sub-sample of the poor/near-poor includes relatively few of the original white sample members (421 out of 2,265) but most of the blacks from our full sample (1,464 out of 1,863). Not surprisingly, the results for blacks are about the same as in


26

Table 6, with the FE models generally showing evidence of nonlinear, though difficult to interpret, childhood neighborhood effects. For whites, the nonlinear FE estimates of the effects of child neighborhood quality increase when we limit the sample by income in Table 7. These effects remain strong relative to the effects among blacks, again demonstrating the differential effects of race on the relationship between child and adult neighborhood quality. Because so few blacks dropped out of the models when we limited the sample to those with incomes less than 150 percent of the poverty line, we estimated an additional set of models for black sample members with incomes at or below the poverty line (Table 8).13 In the spline model, we find effects only for those blacks who grew up in the most disadvantaged neighborhoods. We also find positive linear effects (p