The remainder of Table 4 reports results for the specification in equation (2) that includes
the full set of covariates shown in Table 1 in both levels and interacted with whether the pair
resides on the same block. The rows are assembled by groups of variables, such as educational
attainment or race/ethnicity of workers in the pair, where the parameter estimates for the level
coefficients are listed for the entire set of variables followed by the parameter estimates for the
variables when interacted with whether the two workers live on the same block, bmatch.
Focusing first on the results for the full sample, the bmatch interaction estimates are
statistically significant for most of the included socio-demographic categories in Xij.34 The
interaction effects vary by pair characteristics in a number of interesting ways. With respect to
education, stronger interactions occur for matches where both individuals are high school
graduates while the weakest interactions occur for matches between high school dropouts.
Matches between individuals with children, and especially those with elementary or secondary
school-aged children of the same age also result in strong referral effects. Similar evidence of
assortative matching among neighbors can be seen in the age interactions, where the size of the
referral effect is also largest for matches between the youngest adults in the sample.
Across gender and marital status categories, referral effects are weakest for matches
between married females relative to all other combination, while matches where at least one of
the members is a married male result in especially strong referral effects. 35 The results for high
school dropouts and married females suggest that referrals happen less frequently in matches
where both individuals share characteristics that are associated with particularly weak attachment
to the labor force. In general, then, our findings are broadly consistent with two common
empirical findings in the existing literature on social networks and on informal hiring channels:
(i) that there is strong assortative matching within social networks and (ii) that referrals can only
occur when at least one member of the pair is well-attached to the labor market.36
Four additional aspects of these heterogeneous results are worth mentioning. First, the
results for race and immigration status show strong estimated coefficients among pairs where
both members are recent immigrants and among pairs where both members are either Asian or
Hispanic. This is not surprising given the propensities for recent immigrants residing on the same
34 The negative intercept for the specification with covariates means that the effect is negative (but barely
statistically significant) for the left out category: this is for matches between Asians/Hispanics and Blacks,
where one person is a high-school graduate and the other is a college graduate, and one person is 25 years
old while the other is 35, etc. Such a category is a very tiny portion of all pairs in the sample. The
estimated social interaction effect is estimated to be positive for over 99 percent of pairs observed in the
data for each specification shown in Table 4.
35 Note, however, that the decreased referral effect for pairs of married females will be balanced by the
increased effect for pairs with (especially similarly-aged) children.
24