The Effect of Neighborhood Characteristics on Young Adult Outcomes: Alternative Estimates

We estimate a set of alternative models to examine the effect of neighborhood characteristics on outcomes among young adult women. The models are motivated by a concern that standard estimates of neighborhood effects may in part reflect the characteristics of families that reside in those neighborhoods. In addition to a “standard” model that includes controls for family background, we estimate fixed-effect models that also control for unobservable family characteristics that may affect young adult outcomes. To do this, we use a sample of sisters drawn from the Panel Study of Income Dynamics. In models that control for family background, we find evidence of neighborhood effects consistent with other recent work. In the fixed-effect models, however, there are no statistically significant effects that are consistent with standard hypotheses about neighborhood effects. The findings from this exploratory study suggest that one should be cautious about accepting findings of significant neighborhood effects derived from models that do not account for the possible selection of neighborhood. The Effect of Neighborhood Characteristics on Young Adult Outcomes: Alternative Estimates Spurred by William Julius Wilson’s (1987) theory about the causes of ghetto poverty and underclass behavior, social scientists have renewed their attention to the question of whether neighborhood and peer characteristics affect the social and economic outcomes experienced by adolescent and young adults. Recent studies, such those of as Brewster (1994), Brooks-Gunn, Duncan, Klebanov and Sealand (1993), Case and Katz (1991), Crane (1991), Duncan (1994) and Sucoff (1996), report that a variety of such characteristics influence completed schooling and the likelihood that a teenager will drop out of school, not join the labor force, become sexually active or a parent, become involved in gang or criminal activity, or use drugs and alcohol. As Aaronson (1995), Evans, Oates and Schwab (1992), Jencks and Mayer (1990), Tienda (1991), and several of the papers cited above observe, one may challenge these (and earlier) findings on a variety of methodological grounds. A central methodological concern is the difficulty of separating exogenous neighborhood effects from the effects of unobservable family characteristics that may be causally associated with neighborhood characteristics. Parents choose their children’s neighborhoods; their choice is partly influenced by their observable and unobservable characteristics; their children’s outcomes will be influenced by all of these factors. Hence, findings based upon methods that do not account for selection of the neighborhood may be biased. In a thoughtful and striking demonstration of this possibility, Evans et al. (1992) find significant peer-group effects on teenage pregnancy and the decision to drop out of school with a model that ignores this endogeneity, then show that these effects disappear when the endogeneity is taken into account. In this paper, we use family and neighborhood data from the Panel Study of Income Dynamics (PSID) to estimate alternative models of the effect of neighborhood characteristics on outcomes among young adult women. In addition to a standard model that includes extensive controls for family 2 background, we estimate fixed-effect models that also control for unobservable family characteristics that may affect young adult outcomes. We examine three important behaviors and outcomes among young adult women: whether a woman had a nonmarital birth, whether she obtained any postsecondary education, and her economic status, measured by her income-to-needs ratio. We estimate neighborhood effects in three specifications: with no controls, to obtain gross effects; with controls for observed family background variables, to obtain conventional net effects; and finally, with fixed-effects methods, to control for unobserved family effects. We focus on four measures of neighborhood quality from the 1970 and 1980 censuses, measured at the census tract level. In the models that control for observed family background, we find evidence of neighborhood effects consistent with other recent work. The estimated neighborhood effects from fixed-effect models are much weaker, and we find no statistically significant effects that are consistent with standard hypotheses about neighborhood effects. These exploratory results suggest, therefore, that conventional estimates may overstate the effects of neighborhoods on young adult outcomes. RECENT RESEARCH ON NEIGHBORHOOD EFFECTS Jencks and Mayer’s (1990) literature review identifies several mechanisms through which neighborhood characteristics might affect social behavior and outcomes. “Epidemic” or “contagion” models (Crane, 1991) indicate that a person’s likelihood of engaging in a behavior, be it proor antisocial, is positively related to exposure to others engaged in the same behavior. “Collective socialization” models (Wilson, 1987) indicate that a neighborhood’s adults provide role models for the neighborhood’s children and monitor their behavior. “Institutional” models indicate that schools, police, social service agencies, and other local institutions affect children’s behavior. In contrast, “competitive,” “relative deprivation,” and “cultural conflict” models conclude that more advantaged neighbors might be harmful to less advantaged children. For example, less advantaged children will 3 tend to get relatively worse grades in schools largely populated with more advantaged children, and hence, may be more likely to drop out. Jencks and Mayer (1990) conclude that “the [pre-1990] literature ... does not ... warrant any strong generalizations about neighborhood effects” (p. 176). More recent studies by Brooks-Gunn et al. (1993), Case and Katz (1991), Crane (1991), Duncan (1994), and Osterman (1991) tend to support the view that neighborhood or peer effects characteristics influence some outcomes and some groups and operate in ways consistent with either the contagion or collective socialization models, though broad generalizations have yet to emerge from this line of research. The common methodology in these studies is to regress an outcome on personal and family background variables and one or more neighborhood variables. The coefficients on the neighborhood variables are the neighborhood effects. Such an approach, however, may not adequately capture the effect of unobservable characteristics that affect the way parents choose their communities. If, for example, parents who choose to live in disadvantaged neighborhoods tend to be parents who do not strongly encourage schooling or have few contacts in the labor market, an estimated neighborhood effect may in part reflect omitted family background variables. Estimation methods that do not take account of this will attribute the effect of the omitted background variables to the neighborhood. This is more likely when only a restricted set of background variables is available to the investigator. Parallel comments apply to the estimation and interpretation of observed peer-group variables. Three recent studies suggest we take seriously the possible spuriousness of estimated neighborhood effects. Corcoran et al. (1992) use PSID data, with zip-code-level census data on four community characteristics, to analyze 25to 32-year-old men’s earnings, wage rates, hours of work and family income. After controlling for an unusually wide array of personal and family background variables, they find that the percentage of community families receiving welfare is negatively 4 associated with most outcomes, but that the other three community characteristics generally have negligible effects on all outcomes. The rich set of background variables may have washed out spurious neighborhood effects by largely eliminating the bias likely to arise when there is a limited set of background variables (i.e., owing to omitted variables). In another study using the PSID, but with 1 census tract data, Ginther, Haveman, and Wolfe (1993) find that neighborhood characteristics lose statistical significance as models include increasingly more family background variables. Evans et al. (1992) argue that the choice of peer group is likely to be endogenous and, hence, simple regression models produce biased estimates. Their single-equation models of teenage pregnancy and school dropout show significant peer effects. With a simultaneous-equation model, the effects disappear. Recent research, then, provides tantalizing evidence that the social and economic characteristics of neighborhoods have important impacts on youths’ lives and the likelihood that they will become productive adults. Yet there are plausible methodological grounds to suspect the evidence. EMPIRICAL PROCEDURES, DATA, AND MODEL SPECIFICATION Methods. The model typically used to estimate neighborhood effects is of the general form Y = Z + N + , (1) i i i i where Y is some outcome measure of interest, Z is a set of individual and family variables that affect Y, N is a measure of neighborhood quality, is an error term, and and are the corresponding parameter estimates. There are a number of potential problems in this simple model. As Evans et al. (1992) note, N may be endogenous, since it reflects the location decisions of the parents. A related problem, on which we focus, is the presence of unobservable family effects that may affect the outcome and be correlated with neighborhood characteristics. Let F be the effect on outcome Y of growing up in family j, net of j 5 the observed family characteristics included in Z. For example, in a model of educational attainment, F might represent educational aspirations and encouragement of the parents. If F affects Y but is unobserved and thus omitted from the model, estimates will be biased as long as F and N are correlated. Suppose that more effective families (those with a higher value of F) and neighborhoods of better quality both affect an outcome in the same direction. Then, if more effective families tend to locate in better quality neighborhoods, the estimate of from a conventional model will be biased upward in absolute value, since it incorporates some of the impact of the unobserved family effect, F. A standard approach to a problem like this is to use fixed-effects methods to eliminate the effect of F , and estimate the model by using multiple observations on Y, Z, and N for a given F . j j Fixed-effect models have been used, for example, to measure the effect of a teen birth on socioeconomic outcomes (Geronimus and Korenman, 1992; Hoffman, Foster, and Furstenberg, 1993) as well as in other contexts (Ashenfelter and Krueger, 1994, Griliches, 1979). There are, however, no published applications of this method to the estimation of neighborhood effects. In this application, 2 fixed-effect estimation requires information on two or more siblings per family. Identification of is then achieved by differences in N between or among siblings who share a common value of F but j different values of N. Data and Model Specification Our data come from the Panel Study of Income Dynamics (PSID). The PSID is a nationally representative survey of approximately 7,000 families, who have been interviewed annually since 1968. The PSID is well-suited for sibling analyses because all individuals who resided in PSID households in 1968 are sample members and remain part of the sample when they form independent households. As a result, the sample includes a large number of young adults with siblings who were originally sampled as children in PSID households. Although the data set has some limitations that we note below, it does 6 include the key elements necessary for our analyses—information on neighborhood characteristics and on socioeconomic outcomes for a sizable sample of sisters. The measures of neighborhood quality come from the 1970 and 1980 censuses, which have been matched with addresses in the PSID at the census-tract level. We use four measures of 3 neighborhood characteristics similar to those considered in previous research: the percentage of families with children headed by a single female; the percentage of families receiving public assistance; the percentage of low-income families (< $5,000 in 1970, $10,000 in 1980); and the percentage of middle/upper income families (> $15,000 in 1970, $30,000 in 1980). These income levels identify 4 approximately the bottom 20 percent and the top 25 percent of all families. To analyze the effect of neighborhood attributes on outcomes, we focus on a sample of young women with sisters who are also part of the PSID sample. The outcome measures we use are standard 5 measures of attainment for young women: whether a woman had a nonmarital birth, whether she obtained any postsecondary schooling, and her income-needs ratio, which measures her own earnings and those of her spouse, if she is married, relative to the poverty threshold. The contagion and socialization models suggest that higher values on the measures regarding female family head, public assistance, and low income will be associated with a higher likelihood of a nonmarital birth, a lower likelihood of obtaining postsecondary schooling and a lower income-needs ratio, while higher values on the income measure will be associated with each outcome in the opposite way. Ideally, we would like to have information on neighborhood characteristics over a substantial portion of each woman’s childhood, so that we could construct a measure of average childhood neighborhood quality or even examine at what ages neighborhood quality mattered most. Unfortunately, we are constrained by the data available in the PSID and also by the need to measure outcomes at a meaningful young adult age. Because the PSID does not contain any locational information prior to 1967, it is not possible to construct neighborhood information at young ages for 7 many individuals. Even if that information were available, the relevant years would be sufficiently distant from 1970 to make the 1970 census tract data misleading. Restricting the sample to persons for whom earlier ages fall after 1967 is possible, but this would reduce the sample size substantially and many of them would still be quite young at the point outcomes are measured. We have, therefore, chosen to measure neighborhood quality in the later teen years—as an average across the census tracts in which each woman lived when she was aged 16 to 18. Further, we limit the sample to women who were between ages 25 and 36 in 1987 (the year in which we measure outcomes) and who had one or more sisters in the sample. This age range means that ages 16–18 fall 6 between 1967–69 for the oldest women and 1978–80 for the youngest women. With these restrictions, we are left with a sample of 614 women from 265 families. Within-family variation in neighborhood characteristics of sisters can exist either because the family moved or because the neighborhood changed over time, and as a result a younger sister experienced a different neighborhood from that of an older sister. To estimate the effect of neighborhood characteristics on adult outcomes, we use three specifications. First, we estimate a model using only a constant and one of the neighborhood measures to obtain a gross effect. Then we add a set of major personal and family background variables, including whether the mother and father were high school graduates, whether the respondent grew up primarily in a two-parent family, her race (black or not), her age in 1987, and the parental family’s average income-to-needs ratio, measured over the teenage years. This yields an estimated neighborhood effect, net of measured background. Finally, we specify a family fixed-effects model that estimates the neighborhood effect based on the within-family differences in outcomes, neighborhood characteristics, and personal characteristics that can vary among sisters (i.e., age, mean income-to-needs during her teen years, and being raised primarily in a two-parent family). 8 To keep the exposition manageable, we present results for a model in which the neighborhood effects enter in a linear fashion. This constrains the effect of a one-unit change in a neighborhood characteristic to be independent of the level of the neighborhood characteristic. We have also estimated models with nonlinear effects, using a spline specification in which the effects are permitted to vary above and below the mean. Because the basic findings are so similar across models, we do not present those estimates here. We estimate the models using OLS regression and fixed-effects regression methods for the income-needs ratio and logit and fixed-effects logit for the two dichotomous variables. This fixedeffects logit model is Chamberlain’s conditional logit model. As Chamberlain (1980) shows, only families in which there is within-family difference in outcomes contribute to the likelihood function. This reduces the sample size, because observations from families in which all sisters have the same outcome must be dropped. The computational formulas for the elements of the likelihood function are shown in Maddala (1987), and differ according to the number of sisters in the family and the number of sisters with the particular outcome. In the simplest case of a family with two sisters, the probability that the second sister has a particular outcome and the first does not is Prob (0,1) = exp[ (Z Z ) + (N N )] / [1+exp( (Z Z ) + (N N ))], (2) 1 2 1 2 1 2 1 2 where the subscripts denote the two sisters. The corresponding terms in the likelihood function for families with three or more sisters are constructed in a similar, but more complex, way, with a distinct term for each possible combination of outcomes. Note that this is a standard logit probability in which 9 the explanatory variables are in difference form. The effect of any variable that does not differ between sisters (for example, race or mother’s education) cannot be estimated, nor are there estimates of the fixed effects. In most previous fixed-effect logit estimations (see, e.g., Geronimus and Korenman, 1992; Hoffman, Foster, and Furstenberg, 1993), researchers have selected two sisters per family, regardless 9 of family size, constructed the terms in (2) manually, and then estimated the model via standard logit. Using two sisters in larger families causes two potential problems: first, the choice among sisters is often arbitrary; second, restricting the number of sisters tends to reduce within-family variation in the dependent variable, which is important when the dependent variable is dichotomous. We take advantage of new software to estimate fixed-effect logit models using all sisters in each family.