Discussion - comment on Richard Freeman, in this issue - Special Issue: Earnings Inequality

Katherine O'Regan and John Quigley have written an excellent paper, using intra-urban spatial variation to try to isolate the connection between neighborhood and employment and schooling outcomes for teenagers. They find strong effects of neighborhood poverty and unemployment on teenage employment and idleness (not being at school or at work). They also find that actual physical access to jobs is relatively unimportant. It seems like another victory for the loose form of the spatial mismatch hypothesis (segregation by race and income affects employment outcomes), although something of a loss for the strict form of the spatial mismatch hypothesis (location matters because of transportation distance to work)? The paper is clear and well done, and it is truly "state of the art" in using cross-neighborhood variation within cities to identify the effects of neighborhood on outcomes.

My thoughts on this topic will be arranged in two categories: (1) discussion of the implications of these results for policy and for future research, and (2) discussion of the basic approach of using intra-urban variation to identify neighborhood effects. The first section accepts O'Regan and Quigley's results and discusses what they mean for policy; the second section discusses the perils of using within-city data for these purposes and how one might eliminate some of those dangers.

The Implications of O'Regan and Quigley's Results

The empirical issues involved in estimating the importance of space are extremely dense and often daunting. Issues of omitted variables, endogeneity, and measurement error plague the research in this area (including, of course, my own work). O'Regan and Quigley's research represents a superb effort, but there can be no doubt that we are still far from being able to establish conclusively (1) a firm connection between neighborhood and outcomes, or (2) the neighborhood mechanisms that really matter, or (3) the way these neighborhood mechanisms influence childhood development and employment outcomes. However, these issues are so important, and they extend to so much of the work in social science and to such a wide range of policy-making, that we must welcome truly significant contributions like that of O'Regan and Quigley quite warmly. We also must hope that this conference represents a renewed commitment to continue the quest for more understanding and better methods for dealing with these problems. This first section of my comments presents a brief description of why these issues are so important and what O'Regan and Quigley's results in particular mean, both for policy and for social science.

Implications of a Connection between Neighborhood and Outcomes

The very existence of a strong causal connection between neighborhood and individual outcomes immediately implies the existence of strong, spatially related externalities, especially if that connection does not work through the provision of local public, goods. If a person's identity influences, to even a small degree, the outcomes of his entire neighborhood, then private, free market outcomes there may be not only inequitable but also, quite possibly, highly inefficient. A classic externality exists, because an individual's skill level and work habits influence his neighbors' outcomes in a way that is not regulated through the market.

Location-specific spillovers stemming from the effects of concentration of poverty may suggest, among other things, a need for strongly subsidized education for the poor. As the education of one member of the neighborhood will benefit all of his neighbors, that person's education choice will not internalize all of the neighbors' benefits, and that person will underinvest in education relative to the social optimum. Individual migration decisions will also fail to internalize effects on local neighborhoods. In principle, such results could provide a rationale for a federal government role in reducing white flight, for example, or subsidizing other migration decisions. Once we have clearly established the connection between neighborhood attributes and outcomes, the floodgates have been opened for justifying a myriad of governmental policies. Of course, the standard cautions (which this author believes strongly) about the tendencies of governmental policies to exacerbate rather than improve existing market failures also apply in this case.

A particular example of this last point occurs when local governments take actions that change the neighborhood composition of adjoining areas. One locality may create attractive zoning regulations that draw the wealthy from another area and thus impose significant externalities on the poor remaining in the other area. While I believe strongly in the benevolent effects in many cases of local competition among governments, just as I believe in the benefits of local competition among firms, the presence of substantial externalities may limit the extent to which we want to decentralize certain types of power to local hands. In particular, local control over redistributional activities is known to lead to sorting by income classes. If neighborhood effects are real, then this income sorting may be highly inefficient and socially costly in a way that will not be internalized by local governments.

Of course, these are primarily efficiency issues, and much of the discussion in this area relies on equity concerns. If neighborhood effects are clearly established, then it becomes tempting to ask whether we cannot use these neighborhood effects to achieve equity goals of redistribution between races or between income groups. In other words, if we believe in neighborhood effects, then by altering where the poor live or who their neighbors are, we can improve their lives. Of course, it is still a matter of debate whether space-based redistributional methods (which might include programs helping minorities relocate or community-based redevelopment projects) are particularly efficient means of achieving equity goals. It may well be that cash or simple in-kind transfers are cheaper and more effective means of achieving equity goals than attempts to guide which neighborhood people choose to live in. Naturally, even the most recalcitrant opponent of space-based programs would be forced to accept that it would be of clear social benefit to eliminate spatial distortions created by government policies, such as greater availability of or more access to AFDC payments in high poverty areas, or police discrimination in white neighborhoods.

Finally, it is worthwhile mentioning that documenting the kinds of neighborhood connections that operate is of huge relevance to economics and other social sciences. Much of modern growth theory hinges on externalities in the production of knowledge. Issues in labor economics and macroeconomics are also possibly related to the presence of spillovers across workers in the accumulation and use of human capital. This type of research is invaluable in helping us to document the presence or absence of such forces.

Implications of How Neighborhoods Change Outcomes

One of the strongest implications of O'Regan and Quigley's work is that local poverty matters, while physical distance to jobs does not matter nearly as much. The authors do not really try to distinguish between different forms of poverty or joblessness, but rather restrict themselves to distinguishing between the two fairly different hypotheses. They make a strong case for the importance of local poverty, relative to local job access.

Ideally, we would be able to sort out which types of poverty matter most in creating neighborhood effects. Do neighborhood effects work through the percentage of the population relying on the government? In that case, the admittedly odd implication would be to eliminate all social programs. Is the important attribute the raw income level of the neighborhood? Then, the implication might be to hand out cash. Is the important attribute the concentration of the adult idle? In this case, the goal of policy should be getting people to work. Alternatively, racial composition or some other variable might represent the crucial neighborhood effect.

As pleasant as it might be to believe that we can simply use multiple variable regressions to distinguish among these hypotheses, I am dubious at best about the possibilities for this type of work (O'Regan and Quigley are, too). These poverty-related neighborhood characteristics are tightly correlated in the data. Distinguishing between the effects of unemployment versus poverty versus single-parent families is enormously difficult. The selection and endogeneity problems differ for each one of these variables and further complicate the analysis, and I am not sure that we ever will believe anything we see that differentiates between these forces.

Given these problems, I believe that Quigley and O'Regan adopt the right approach. They basically look at two hypotheses: Is it neighborhood composition that is the major difficulty in poor neighborhoods? Or is it the lack of proximity to employment? These two variables, neighborhood composition and physical location of employment, are not tightly connected, and the authors do seem to be able to effectively distinguish between them. They reject the idea that the lack of proximity to jobs is the major problem. My own work in this area (Cutler and Glaeser 1995) has also found that proximity to jobs is not a particularly large determinant of neighborhood effects. It does seem that the problem of poorer neighborhoods is not the absence of local employment but rather the presence of broader social problems that leave lasting scars on youths growing up in poverty-stricken areas.

The most straightforward interpretation of O'Regan and Quigley's work is that job access, the variable that relates to immediate benefits from legal-sector employment, is relatively unimportant. The variable that appears related to a culture of poverty and its effect on long-term human capital development is important, however. The findings suggest that we should not expect that an individual who is whisked away from a poor neighborhood and dropped into a high-employment neighborhood will immediately see an improvement. More likely, the children of this individual will have better peers and better role models and will eventually learn from the new location. The implication is that neighborhoods are about long-term accumulation of skills or attributes and not about an immediate return to paid work.

Such a policy conclusion casts doubts on the effectiveness of employment zones, enterprise zones, improvements in transportation for inner-city residents, or any policy focused primarily on cutting the costs of moving between ghettos and jobs. While such programs surely will not hurt, and may even be of some benefit, they will not address the primary problems of inner-city neighborhoods. The benefits of these programs will show up only gradually, and through an indirect effect of employment levels on long-run human capital accumulation. Neighborhoods will change, if the O'Regan and Quigley results are right, only if the cycles of poverty are broken and if their residents become employed and acquire human capital. Unfortunately, this type of policy implication goes against any kind of quick fix. Urban policy must be about changing long-run human capital accumulation and altering the patterns of family responsibility.

Implications of How These Mechanisms Work

Quigley and O'Regan do not really begin to tell us how locational unemployment levels actually drive youth idleness. The reader can immediately imagine several mechanisms by which this effect could work through the public sector or the provision of locational services. For example, even when different census tracts are part of the same school district, they may not have access to the same schools, and it may be school quality that is driving this effect.

Are crime levels higher in these poverty areas? The number of poorer youths involved in some form of crime is quite high: About 35 percent of the National Longitudinal Survey of Youth's 20-year-olds have committed crimes recently. Are these young adults just avoiding the legal sector? Is drug use and availability important for this group? Is gang membership important?

Are the poverty effects working through an absence of role models? An easy way to test this is to see whether neighborhood effects become important for children who have both parents present and "successful," where successful may just mean employed. These mechanisms should not be seen as alternatives to a basic poverty effect; rather, analysis should first document the role of poverty and then try to decompose the ways that poverty drives poor outcomes.

It would also be helpful to know if neighborhood effects are seen for younger children, and at what age these forces start to be important. When do we begin to see school dropout rates respond to local area attributes? Naturally, all of these questions form an agenda for many future papers and go far beyond the scope of this work, but they are important if we are to formulate policy on the basis of these types of results. For example, if we found that all the neighborhood effects worked completely through school quality, and school quality was a function of spending, then it would make sense to consider equalizing school spending. If neighborhood effects worked through school quality, but the relevant school quality effect worked through peer interactions at the school, then busing or, alternatively, a measure for paying children with high human capital to go to school with children with low human capital might be preferable. If neighborhood effects worked through high crime rates, and these crime rates discouraged legal activities and encouraged illegal activities, then altering the policing structure might be appropriate.

The point is not that a clear mandate exists on what should be done, but rather that determining how neighborhoods affect outcomes, if indeed neighborhoods do affect outcomes, is critically important for determining our overall policy approach. We cannot even begin to think about the right steps to take to eliminate the problems of the inner city without first being convinced that neighborhoods, rather than individuals, are important factors in creating social problems; without knowing which types of neighborhood characteristics drive poor outcomes; or without understanding the mechanisms by which they drive these outcomes.

A Discussion of the Intra-Urban Approach

As I have argued elsewhere (Cutler and Glaeser 1995), using intra-urban variation to identify the effects of neighborhood characteristics on individual outcomes poses two major problems. O'Regan and Quigley are aware of both, but it is worthwhile discussing the assumptions needed to avoid these problems and whether or not we think that these assumptions are palatable.

The first problem is that omitted family and child characteristics surely are highly correlated with neighborhood choice. Neighborhoods are endogenously chosen, and individuals select into different locations based on their characteristics. Some of these characteristics will be the observables that O'Regan and Quigley do use in their work. Other relevant characteristics relating to neighborhood choice might be the willingness to sacrifice for future benefit (patience), unobserved human capital and skills, or connections with and attitudes toward mainstream society. If negative attributes are correlated with choices to live in poorer neighborhoods, then our estimates of neighborhood effects will be biased upward, since neighborhood characteristics will be correlated with omitted variables that work in the same direction (as long as bad neighborhoods attract low-potential individuals).

The second problem, which also stems ultimately from the endogeneity of neighborhood choice, is that identical individuals must in equilibrium be indifferent between neighborhoods. Thus, the marginal individual making the decision about neighborhood location must be indifferent between living in a poor neighborhood and living in a rich neighborhood. (Housing prices surely go a major part of the way to induce this indifference.) This effect will mean that we should not see neighborhood differences in utility levels of the decision-makers, if we are able to control for all individual attributes.

My approach to these problems has been sheer cowardice. David Cutler and I avoided using intra-urban variation entirely and identified neighborhood effects from inter-urban variation. We were able to use governmental and topographic features of different urban areas to instrument for the degree of segregation within the area. Unfortunately, we had only weak methods of dealing with inter-urban mobility, which is also endogenous. More important, the price of going to inter-urban variation (also the approach used in O'Regan and Quigley 1995b), is a tremendous loss of variation. In the extreme case, where every urban area was identical but had huge neighborhood differences, inter-urban variation would yield no evidence whatsoever. While the world is less extreme than that, all researchers lose a large amount of information when they give up the information contained in within-city data, and a huge cost is attached to adopting that type of strategy. I think that in the long run we will be better off figuring out ways to use the intra-urban data than we are relying solely on inter-urban variation.

However, using intra-urban variation requires dealing seriously with all the potential biases that such data create. Consider the following earnings equation:

[Mathematical Expression Omitted]

where E reflects some outcome variable (perhaps earnings, or some propensity towards idleness), X represents observed individual characteristics, [Beta] the returns to those characteristics, Z observed neighborhood characteristics, [Theta](Z) the average returns to those characteristics, [[Theta].sub.i] the individual specific returns to those characteristics, [[Alpha].sub.i] omitted ability, and [[Epsilon].sub.i] an independently distributed error term. The potential problems with using ordinary least squares to estimate the equation, and the possible solutions, are discussed below.

Case One - Garden Variety Omitted Variables

In this case, [Theta](Z) = [Theta], [[Theta].sub.u] = 0, and the covariance of [[Alpha].sub.i] and Z is not equal to zero. Ordinary least squares will yield biased coefficients, because neighborhoods are correlated with unobserved attributes. O'Regan and Quigley (1995a) are aware of this problem and handle it by implicitly assuming that parental job attributes determine location and that these attributes are orthogonal to teenage attributes. In their words, household choice is "made by the parent(s), using the standard transportation-housing costs calculus. Household choice is exogenous to the transport demands of youth." As the equation illustrates, the necessary condition for unbiased estimates is not the exogeneity of location choice with respect to youth's employment concerns, but rather the orthogonality of location with respect to youth's employment concerns.

The authors assert, perhaps correctly, that households do not choose location based on what will make employment more probable for their children. I am skeptical of this comment in many cases, especially given what we know about how sensitive parents are to school quality in their location choice. Nevertheless, even accepting this assertion, the parental factors that induce parents to locate in high-poverty areas are surely correlated with the characteristics of youth that determine employment probabilities. Indeed, O'Regan and Quigley assert that, in their data, family characteristics "really 'matter' in the empirical results." If the observables matter so much, surely the unobservables matter too, and the results are biased.

How can we work to improve this problem? First and most classical is the instrumental variables approach. The goal is to find a parental characteristic that determines location but is clearly orthogonal to omitted youth characteristics that drive location. One possibility is that the industrial or occupational training of parents might influence locational choice.

Naturally, we would have to control for the overall quality of industry or occupation as well. The method would involve creating a location measure for each industry/occupation pair and also an average wage and average skill measure for each industry/occupation pair. The location measures (where the industry/occupation employment is located in the city) might be clean instruments if the industry/occupation quality measures are also included in the regression. Alternatively, in data samples where we know when the parents came to the city, we could use the areas of the city being built then to get a sense of where the parent would have been attracted to initially, and use that as an instrument. Ideally, we could use randomized data (such as the Gautreaux or Moving to Opportunity experiments) to get better instruments as well.

A second approach is to get a sense of how big the selection problems are. How much is sorting by parental observables? How strong is the correlation between parents and children? How big would the unobservables need to be, relative to observables, to invalidate the results? These kinds of sensitivity analyses are made possible by Quigley and O'Regan's use of Census variables with a battery of parental background data, and I believe that the authors should exploit this information as much as possible.

In a final approach, the authors could separate individuals into long-term and short-term residents of the community. Presumably location choice would be less of an issue for long-term residents. If the data showed that neighborhood was most important for long-term residents, this would lead us to believe that it is neighborhood that drives outcomes. If neighborhood is more important for short-term residents, then we would have to believe that outcomes drive neighborhood choice.

Case Two - Random Coefficients

In this case, [Theta](Z) = [Theta], and [[Theta].sub.i] [not equal to! 0, but [[Alpha].sub.i]i = 0, and the covariance of [[Theta].sub.i] and the Z variables is not equal to zero. This is a version of the standard Roy model, where individuals have different returns to different neighborhoods and will select into the neighborhoods that give them higher returns. While the relative returns may be parental returns, so long as they are parental returns, ordinary least squares will yield biased coefficients, because neighborhoods are correlated with unobserved attributes. The problem here is not that omitted variables are present that positively affect employment and are also correlated with neighborhood, but rather that the returns to neighborhood location itself differ across neighborhoods. A particular, real world example of this concern is the fact that the minorities who have selected to live in rich neighborhoods are minorities for whom that neighborhood is particularly valuable, so that it is impossible to translate from information about those people to general results about the importance of location for minorities.

This version of the problem has two approaches. The first tends to be highly parametric and involves assumptions about the distribution of the returns to neighborhood. Luckily a large literature exists on this topic, stemming from Heckman's work in the 1970s, and well-worked-out techniques are available for dealing with this problem parametrically. However, while the robustness of the neighborhood results to Heckman-type corrections would be an extremely pleasant thing to see, I am not sure that skeptical readers would be completely convinced by this type of approach.

A second approach to this topic examines whether the returns to neighborhood location differ much, using observable characteristics. This type of test is readily performable and amounts to looking at the cross-effects between individual and neighborhood characteristics. These cross-effects are in fact intrinsically interesting, as well as useful in providing evidence about the extent to which returns to neighborhood differ over varying types of people. Of course, it is worthwhile remembering that even if little difference is found in the returns to neighborhood variables by observables, significant differences still might exist in the returns to neighborhood by unobservables.

Case Three - Endogenous Average Returns

In this case, [Theta](Z) = [Theta](Z), [[Theta].sub.i] = 0, and [[Alpha].sub.i] = 0. Here the returns to different neighborhoods are the function of market forces, and in equilibrium the same people will be indifferent between neighborhoods; that is, the distributions of populations will select to the point where individuals are indifferent between different neighborhoods. In part, this issue is the most easily resolved by O'Regan and Quigley's argument that parents select on the basis of their own needs, not the needs of their children. If they are right, then parents will be indifferent but children need not be, and identification still makes sense. In this case, it is enough that location be exogenous, and we are not concerned about the correlation of location with unobservables.

While the argument that they use is both technically correct and quite possibly true, the authors could take this issue much more seriously. It would help to show the factors that parents select on and try and predict what determines the parent's choice of location, and to show that it has little to do with variables that affect children's outcomes. More generally, to the extent that the authors are able to indicate compensating differentials in other areas - high housing costs in the areas where children benefit most - it will be more plausible to believe that the equilibrium does not rely completely on children being indifferent. Indeed, in some ways this problem is the least troublesome, because it does not involve any estimation bias. Instead, what is involved here is the question of why we would expect to find neighborhood effects, if the ability to migrate between neighborhoods exists. Much of the answer assuredly lies in the nature of the equilibrium and of the forces that equilibrate the system.

These three problems with intra-urban data are potentially quite serious. I have presented them separately, but further problems arise if all three problems occur at once. However, approaches to these problems can be developed and O'Regan and Quigley have made invaluable steps forward, both by formalizing some of their responses to these criticisms and by using such a rich, strong data source.

1 Kain (1968) is the father of the spatial mismatch hypothesis. I have taken to splitting the hypothesis into strong and weak forms, where the strong form states that minority problems are related to distance from jobs and the fact that minorities are constrained to live in their neighborhoods, whereas the weak form argues that segregation, which is a result of discrimination, leads to poor minority outcomes.

References

Cutler, David and Edward L. Glaeser. 1995. "Are ghettos good or bad?" National Bureau of Economic Research working paper no. 5163.

Kain, John F. 1968. "Housing Segregation, Negro Employment and Metropolitan Decentralization." Quarterly Journal of Economics, vol. 82, pp. 175-97.

O'Regan, Katherine and John Quigley. 1995a. "Spatial Effects upon Employment Outcomes: The Case of the New Jersey Teenagers." New England Economic Review, this issue.

-----. 1995b. "Teenage Employment and the Spatial Isolation of Minority and Poverty Households." Mimeo. Forthcoming, 1996, Journal of Human Resources.

Edward L. Glaeser, Associate Professor of Economics, Harvard University.

COPYRIGHT 1996 Federal Reserve Bank of Boston
COPYRIGHT 2004 Gale Group