We also tackle the potential for reverse causation directly by estimating equations (1) and
(2) on a sub-sample of the data in which both respondents in a given matched pair have lived in
that neighborhood for at least two years, but one of them was not employed for the full year in the
previous year, defined as having worked less than 45 weeks in 1989. In this case, we can be fairly
certain that if we see the same individuals working together in the current year then the referral
was among residential neighbors rather than work colleagues. Unfortunately the Census does not
contain any direct information on job search activity. Therefore, we use the “not employed for the
full year in 1989” category as a proxy for the set of individuals who are most likely to have been
actively searching for a job last year.28 We also estimate an intermediate specification using the
sub-sample of pairs whose members were both in residence at least two years, and adding
controls for whether one and/or both individuals were not employed for the full year in 1989. The
goal of this analysis is to examine whether evidence of referrals is present in this sub-sample.
Importantly, because this sub-sample is (by construction) very different from the main sample,
we do not expect the resulting model of social interactions to be identical to our baseline results.
As a result, there is no reason to believe that the referral effect will be stronger for matches in this
subsample or even to believe that the estimated parameters will be stable over this subsample.
The strength, rate of utilization, and the form of the local referral network are likely to differ
based on how long an individuals resides in a neighborhood.
Inference. Finally, a word about inference. The sampling scheme, which is based on drawing
matched pairs of individuals who reside in the same block group, makes it very difficult to
compute appropriate standard errors for our estimates. In particular, the observations in our
sample -- pairs of individuals in the same block group -- do not constitute a random sample. In
fact, suppose that individuals a and b work in the same block. Suppose further that individuals b
and c work in the same block. Then, by transitivity, individuals a and c must also work in the
same block. As a consequence, if we compute standard errors via the basic OLS formula, we may
tend to understate their size because we are not taking into account this inherent correlation
structure in the data. There is also the reasonable concern of heteroscedasticity across block
groups that may bias standard errors in fixed effects analyses. In fact, the use of the linear
probability model assures heteroscedastic errors. To address these issues, all standard errors in the
match model are estimated based on pairwise bootstraps. It should be noted that some concerns
28 Note that in estimating earnings and wage equations in Tables 6 and 7 we condition on a set of
individuals that were fully-employed in the previous year defined as having worked at least 40 weeks and at
least 30 hours per week. This definition is different than that for not employed for the full year in 1989
used here, which is not at all based on hours.
16