particularly prone to measurement error.23 This concern about the full sample is substantiated in
the unconditional correlation estimates, as these are significantly greater in a number of cases.
The correlation estimates that condition on block group fixed effects, however, are generally of
the same magnitude as those reported for the full sample. Moreover, the estimates that condition
in addition on housing characteristics are in many cases even smaller than those reported for the
full sample.
The magnitude of the remaining correlation between individual and neighbor attributes
within block groups provides clear support for the notion that the amount of sorting within block
groups on observables is less extensive than across the neighborhoods of the metropolitan area as
a whole. This evidence is particularly compelling for our identification strategy because a number
of these attributes, such as residents’ race or the presence of families with children, would be the
characteristics of one’s immediate neighbors that might be most observable at the time of moving
into a new residence. Thus, by controlling for these observables, it may be the case that within-
block group sorting on other characteristics is even less extensive. While the correlation estimates
reported in Table 3 are small, however, they are not identically zero. An obvious question, then,
is whether the remaining block-by-block sorting on the basis of observables within block groups,
small though it may be, is enough to significantly increase the propensity of pairs drawn from the
same block within a block group to work together. We provide additional evidence on this
question after first introducing a heterogeneous version of the model.
Heterogeneous Specification. The initial specification shown in equation can easily be extended
to include a set of covariates Xij that describe the pair of individuals (e.g., those summarized in
Table 1) both in levels and interacted with Rijb:
(2) Wijb =ρg+β'Xij+(α0+α1'Xij)Ribj +εij
23 In particular, a bias is induced in the estimated correlations reported here as a result of the fact that the
average block characteristics are constructed from a (1-in-7) sample of individuals rather than a complete
census of neighbors. This bias is present, however, in each specification reported in Table 3 and,
importantly, should not generally be greater in the specification that conditions on block group fixed effects
than in the unconditional specification. We confirmed this with Monte Carlo simulations. The results for
the sample of blocks with five workers or more also is supportive of this notion, as measurement error
should be substantially lower in this sample and yet the decrease in the estimated coefficients from the
unconditional specification to the specification that conditions on block group fixed effects is greater in this
sample.
13