The magnitudes of the labor force participation and employment effects estimated in
Table 6 are generally consistent with the increased propensity to work with at least one neighbor
in the same block estimated using the match specification above. In particular, the estimated true
standard deviation of match quality for our preferred sample (5+ workers on the block) is about
0.18 percentage points.46 This change in the propensity to work with each neighbor raises the
probability that an individual works with at least one neighbor by approximately 9.2 percent at
the mean. Given that one person in a match is providing the referral, this in turns implies an
increase in the propensity to find a job through a neighborhood referral of 4.6 percent. This
number corresponds to the increased propensity to work with someone on exactly the same block
and, therefore, provides a lower bound on the number of neighborhood referrals more generally.
When compared to the employment effect estimated for the corresponding sample (1.8
percent), this then suggests an upper bound of about 40 percent of referrals (1.8/ 4.6) that result in
employment for an individual who would not be employed in the absence of the referral, while
the other 60 percent of neighborhood referrals go to individuals who would find employment
through another search method. Again, because the denominator in this calculation is expected to
be understated while the numerator is not, the actual fraction of referrals that result in a non-infra-
marginal employment is likely much less.47
Table 7 reports the coefficient on match quality for labor market outcome regressions
corresponding to the work match regressions based on the block groups that exhibit the least
block-by-block sorting reported in Table 5. In general, the results are qualitatively similar to the
ones obtained using the full sample, thereby confirming the robustness of our analysis to block-
level sorting. One interesting aspect of this analysis, however, is that the labor force participation
and employment results are smaller for each of these sub-samples, while the wage results are
larger suggesting that referrals may be useful largely for finding a good job rather than for finding
any job.
Reverse Causation. Table 8 provides estimates of specifications designed to address the
possibility that the estimated social interaction effect may be due to reverse causation, i.e.,
46 As discussed in the last sub-section in Section 5, match quality is measured with error due to the 1-in-7
nature of the Census sample. As a result, the measured standard deviation is significantly greater than the
true standard deviation, which we estimated through Monte Carlo simulations.
47 Recall that we expect the labor market outcome regressions to provide an estimate of the ultimate impact
of all actual referrals from the neighborhood including individuals in both the same and nearby blocks. In
particular, with limited sorting within block groups, expected match quality for individual with others in the
same block group is the same as their actual block match quality. Consequently, the block level index for
match quality is likely to capture the effect of referrals both within the block and from neighboring blocks.
28