multiple variables are measured with error, this result does not necessarily follow immediately
because of the possibility of correlation across regressors. To address this concern, we also
consider robustness to the omission of all block average attributes other than match quality in
these labor market regressions. A finding of similar results for these alternative specifications
provides some confidence that the results are not driven by measurement error.
6 RESULTS
Having described the research design for each portion of our analysis above, we now
present the results. We begin by examining the propensity for two individuals to work together,
first reporting some summary statistics and then the estimated coefficients of the baseline
regression specifications given in equations (1) and (2). We then present results for the
alternative specifications based on sub-samples drawn from the most homogeneous block groups
along various sociodemographic dimensions. Having presented these estimates of the work match
regressions, we proceed to discuss the corresponding labor market outcome regressions for each
of these specifications. A final sub-section explores both employment location match and labor
market outcome specifications that address the possibility of reverse causation, examining sub-
samples that condition on residential tenure and on whether individuals were fully employed in
the previous year.
Table 1 contains summary statistics for our matched pairs sample. As described above,
the first column reports the fraction of pairs that fit the description in the row heading. The
second column reports - for each category - the empirical frequency that two individuals that
reside in the same block group but not the same block work together. The third column reports the
probability that two individuals that reside on the same block work together. In this way, the first
row indicates that the baseline probability of working together for two individuals that reside in
the same block group but not the same block is 0.36 percent; this figure rises to 0.94 percent for
two individuals that reside on the same block. As we will see below, much of this increased
propensity for individuals residing on the same block to work together results from the fact that
the sample of individuals that reside on the same block is disproportionately weighted to larger
blocks - i.e., dense block groups. The inclusion of block group fixed effects in our main empirical
specification ensures that our social referral effects are estimated purely on the basis of
comparisons within the same block group.
The remaining rows of Table 1 reveal how these patterns vary with the characteristics of
the pair of individuals. First, notice that individuals residing on the same versus nearby blocks
show an increased propensity to work together across all of the types of pairs characterized in the
21