factors related to significant changes in neighborhoods that are not easily disentangled (see
Moffitt (2001) for a detailed discussion).
A second broad approach seeks to deal with the difficulties induced by correlation in
unobserved attributes at the neighborhood level by aggregating to a higher level of geography.
Evans, Oates, and Schwab (1992), Cutler and Glaeser (1997), Ross (1998), Weinberg (2000,
2004), Ross and Zenou (2004), and Card and Rothstein (2005) identify the effect of location on
outcomes using cross-metropolitan variation. For example, Cutler and Glaeser (1997) analyze the
impact of segregation within a metropolitan area on a variety of outcomes including education,
labor market activity, and teenage fertility, and Evans, Oates and Schwab use metropolitan area
poverty rates as an instrument for neighborhood level poverty. Again, the advantages of this
approach are clear - aggregation certainly eliminates the problem of correlation in unobservables
among neighbors (although potential correlation in unobservables at the metropolitan level
becomes an issue). The effects identified through aggregation, however, include not only the
average neighborhood effects operating in a metropolitan area but also any broader consequences
of living in a segregated or high poverty metropolitan area.6 Thus, the strict interpretation of the
estimated effects as neighborhood effects requires the assumption that metropolitan segregation
does not directly affect outcomes.7
An interesting way to view the research design developed in this paper is as the converse
of designs based on across metropolitan area variation. That is, instead of aggregating to the
metropolitan level, we disaggregate below the level of the neighborhood to isolate block-level
variation in neighbor attributes. While the strict identification of neighborhood effects with the
across metropolitan area design requires the assumptions of no metropolitan effects and no
correlation in unobservables at the metropolitan level, strict identification with our design
requires the assumptions that social interactions among neighbors are very local in nature -
operating at the level of the block - and that there is no correlation in unobservables across blocks
within block groups.8 In this way, we view the current paper as offering a complementary
approach to the existing literature that allows researchers to identify a wide range of causal
neighborhood effects using an alternative set of assumptions (testable on the observables) than
have been used in previous studies.
6 More residentially segregated metropolitan areas might be associated, for example, with increased racial
taste-based discrimination in the labor market, in the application of criminal justice, etc. due to decreased
levels of regular inter-racial contact in residential neighborhoods.
7 It is important to point out that Cutler and Glaeser (1997) do not claim that the effects identified in their
analysis are strictly a neighborhood effects.
8 To the extent that interactions occur among neighbors at greater distances, our estimates reflect only the
increased intensity of interaction at the block level.