workers receiving tips and referrals about residential locations from their co-workers. These
specifications examine pairs of individuals that have been in their current residence for at least
two years and focus on the estimated interaction effects for individuals who were not employed
for the full year in the previous year. As noted above, the goal of this analysis is to examine
whether evidence of referrals is present in this sub-sample. Again, because this sub-sample is
very different from the main sample, we do not expect the estimated social interactions to be
identical to our baseline results.
For reference, the first panel in Table 8 reports results for the sample of pairs that have
been in their current residence for at least two years, again restricting attention to the sample of
blocks with at least five workers. The estimated coefficients in this case are broadly consistent
with those reported for the full sample in the second column of Table 4; the correlation in the
predicted measure of match quality from these specifications is 0.71. The estimated coefficients
are qualitatively similar although generally smaller in magnitude to those in the baseline
regression for education, age, the presence of children, gender and marital status, and
immigration.
The middle panel of Table 8 adds controls in both levels and interactions with bmatch
based on whether the workers in the pair were not employed for the full year in 1989, defined as
having worked 45 weeks or less. While failing to rise to the level of statistical significance, social
interactions are stronger for matches in which one of the individuals was not employed for the
full previous year while the other individual was (0.02 percentage points greater), whereas
interaction effects are dramatically weakened when both members of the pair were not employed
for the full previous year (0.12 percentage points smaller) relative to pairs in which both were
employed for the full previous year. Since these are workers who have resided in the same
location for at least two years, these findings do not lend support to the reverse causation
hypothesis (co-workers giving referrals about desirable residential locations to new employees).
The last set of columns in Table 8 focuses on the sub-sample of pairs with both
individuals in residence at least two years, but with only one member employed for the full
previous year. Again, this sampling scheme reduces the possibility of reverse causation, since we
are considering workers who are more likely to have made a transition to full employment during
the past year and whose residential tenure is longer than two years. At the same time, by looking
at pairs in which one was employed for the full year while the other was not, we are focusing on
instances in which it is most likely that a referral or information exchange actually took place.
As in the other specifications, the estimated social interaction effect is strongly positive
and statistically significant for the version without covariates. When we introduce covariates, the
29