results for the full sample; column 2 reports results for the sample that drops blocks with fewer
than five workers in the sample; column 3 includes a series of controls that characterize the
housing stock. These latter specifications relate directly back to the correlation analysis shown in
Section 4. Given the results of that analysis, which show that the correlation between observable
individual and neighbor characteristics falls to near zero with the dropping of blocks with small
numbers of sampled workers and the inclusion of block group fixed effects, column 2 reports our
preferred specification. While the inclusion of housing characteristics in that analysis moved the
estimated correlations even closer to zero, the fact that house value and rent may in part capitalize
some components of neighbor characteristics lead us to believe that this specification provides a
lower bound on the interaction effects. As we will see, all three specifications yield quite similar
results.
Starting with the results for the specifications without covariates summarized in the first
row, the estimated social interaction effect is positive and statistically significant in each case,
indicating a strong additional propensity for two workers living in the same block to also work in
the same block, over and above the estimated propensity for matches in their block group. The
magnitude is 0.12 percentage points for the full sample and the sample based on blocks with at
least five workers, falling to 0.11 percentage points when housing controls are added. This effect
is sizeable: it is roughly 33 percent the size of the baseline propensity to work together for two
individuals that reside in the same block group but not the same block (0.355 percent).31
An increased propensity to work with a given neighbor implies a much larger propensity
to work with at least one neighbor. For our preferred sample, which restricts the sample to blocks
with at least five sampled workers, given the average of 80 individuals per block,32 an estimated
referral effect of 0.12 percentage points translates to approximately a 6.9 percentage point
increase in the propensity that an individual works with at least one individual on the same
block.33 Thus, the referral effect estimated here is certainly economically meaningful.
31 As noted above that this effect is less than the mean difference reported in Table 1 suggests that a portion
of the differences in mean between those residing in the same block versus those in the same block group
but not the same block was driven by variation across block groups related to population density. See
Section 4 for a discussion of this issue.
32 While the average number of workers meeting our sample criteria for the match model is only 5.1
workers, the fact that the Census is a 1-in-7 sample and that many workers are excluded from our analysis
due to the presence of imputed data accounts for the larger average number of actual prime-age workers per
block.
33 For computational ease, this calculation treats the likelihood of working with each neighbor as an
independent event. The reported 0.069 = (1 - 0.00355)^80 - (1 - (0.00355+0.0012))^80, where 80 is the
average number of adults on the same block, 0.00355 is the baseline propensity for individuals to work
with someone in the same block group and 0.0012 is our estimated social interaction effect.
23