block to work together reported in Table 1. Again, as noted there, because it is very likely that
many recent immigrants simultaneously receive referrals for both residences and homes at the
time of immigration, we do not interpret the resulting coefficient as a causal neighborhood effect
but include immigration status only as a control. Again, all of the results reported in the paper are
robust to dropping immigrants from the sample.37
Second, the results also reveal that social interaction effects are declining with population
of a block (i.e., decreasing in density). That our estimated referral effects are driven by blocks
with a smaller number of housing units is encouraging because the housing market for such
blocks will naturally be thinner - hence with less scope for sorting within block groups.38 Notice
that this is another case where our estimated social effect has the opposite sign when compared to
the baseline propensity for two individuals residing in the same block group to work together.
That is, while individuals that reside in dense block groups are generally much more likely to
work in the same location, we estimate that referrals from neighbors are less likely in dense
places.
A third important aspect of the results presented in Table 4 is that there are significant
differences between the level and the interaction coefficients associated with the Xij covariates.
For example, conditional on the other attributes in the model, pairs of married females within the
same block group are each the most likely to work in the same block (as discussed above, perhaps
because they tend to work close to home) and have the weakest referral effects among all gender
and marital status categories. A similar pattern obtains for high school dropouts. As discussed in
Section 4 above, such substantial differences between the estimated α1 and β coefficients provide
additional assurance that the estimated referral effects are not simply capturing additional sorting
at the block level.
Finally, a comparison of the results across the three specifications reported in Table 4
reveals a very similar pattern as blocks with fewer than five sampled workers are dropped and
housing characteristics for each pair are included as controls. Again, because these housing
controls, which include price measures, might absorb out too much of the variation in the
underlying effect that is actually attributable to neighbor characteristics (due to capitalization) we
expect that this specification may understate the strength of the interaction for characteristics that
36 See, for example, Corcoran et al. (1980).
37 Note that matches between pairs where both are non-US born individuals having immigrated in the past 8
years represent only 0.22 percent of the overall sample. Thus, the magnitude of this effect is not
responsible for the overall average referral effect of 0.12. In fact, the estimated average effect falls by less
than 0.02 percentage points when all immigrants are dropped from the sample.
38 Alternatively, one could think that social interactions are weaker in larger blocks because it is more
difficult to establish and maintain a social contact in such a block.
25