In this case, the estimated coefficients on the cross terms, α1, allow us to investigate whether the
social interaction effect is weaker or stronger for specific socio-demographic characteristics of the
matched pair. There are two aspects to this: first, certain pairs are more likely to interact because
of the assortative matching present in social networks: for instance, two individuals of similar
age, education, race, or with children of similar age.24 Second, certain individuals may be more
strongly attached to the labor market and may thus provide better referrals or information on jobs
- for example, college graduates, married males or individuals with children. In this case,
matches between pairs in which one individual is strongly attached to the labor market and the
other generally more likely to need a referral should also lead to an increased number of referrals.
In equation (2), β'X measures how the propensity to work together of two individuals that
reside in the same block group but not the same block varies with the observable characteristics
of the pair. Given an estimate ofβ , this heterogeneous specification provides a way to test
whether the remaining within-block group correlation between observable neighbor attributes
would lead to a significantly higher predicted propensity for pairs on the same block to work
together. Specifically, we compare the average β' X for those pairs that reside on the same block
with those that reside on nearby (but not the same) blocks within the block group.25 Given the β
that we estimate below, the results of this test are as follows. The predicted propensity for pairs
that reside on the same block is 0.343 percent; this is 0.01 percentage points lower than the
observed (and predicted) propensity for pairs that reside in the same block group but not on the
same block (0.355). Thus, the remaining block-level sorting on observables does not predict any
increased propensity for individuals on the same block to work together. This evidence strongly
favors the notion that our research design is credible in the face of the small amount of within-
block sorting that exists in the data.
Another competing potential explanation for the finding of a greater propensity of pairs
to work together at the block versus block group level is that this propensity is simply a declining
function of the distance between any pair of individuals in the metropolitan area. While we do not
address this possibility directly in the analysis, two aspects of the results that follow are important
in ruling out this potential explanation. First, the magnitude of the social effect that we identify is
large relative to the underlying propensity for two individuals in the same block group to work
24 See Marsden (1987), (1988) for a discussion of the evidence from the General Social Survey on
assortative matching in networks.
25This model is also rerun using the housing controls that were used our diagnostic test, the analysis of the
correlation between individual and neighborhood attributed. As will be seen below, results are quite
comparable across specifications.
14