model.30 For all other outcomes, such as weeks worked, hours-per-week worked, wages and
earnings (in logs), we use a simple linear regression.
We then add - for each model specification - a ‘network quality’ proxy variable for each
individual, which is constructed by examining that individual’s matches with other adults in her
block, using the coefficient estimates α1 from the estimation of equation (2). Specifically, the
average match quality for individual i, Qi, is constructed using a sample of all possible pairings of
individual i with other individuals who reside in the same block and do not belong to the same
household. For each pair, a linear combination Mij of the pair's covariates is created using the
estimated parameters from the interaction of these variables with Rijb in equation (2):
Mj = α1X lj. Then, Ql is computed as the mean value ofMij- over all matches for individual l:
(3) Ql=à ∑M
where Ni is defined as the set of other individuals that reside on the same block but not in the
same household as individual i.
We would generally expect individuals with good matches in their block - high value of
Qi - to have better labor force outcomes on average, after controlling for the direct effect of their
attributes, the average attributes of their block, and block group fixed effects. We repeat the
analysis for each of the various specifications described in Section 4 to address the sorting and
reverse causation issues. In particular, by using a sub-sample of individuals that were not fully
employed last year, we focus on the group that was most likely to have been looking for work in
the past year. The effect of Qi on labor market outcomes cannot be driven by residential referrals
from coworkers if the sample and match quality model is conditioned (to the extent possible
using census data) on a residential location match that arose before the employment location
match. As mentioned previously, we have no a priori expectations concerning how the strength
of the referral effect varies depending upon whether the employment referral occurred recently or
sometime in the past. The specification used for this second stage of our analysis is given by:
(4) Ei = θg + δ'iXi + δ'2 Xl+ δ3 Qi + ui
30 We have also performed our analysis using a multinomial logit specification, with three discrete
outcomes: out of the labor force, unemployed, and employed. The results are qualitatively very similar.
18