are most likely to be capitalized - such as college-educated neighbors. While there is some slight
evidence of this, the same pattern generally holds for this specification. Given these
complications, however, we treat the specification shown in Column 2 as our preferred
specification. The correlation of predicted match quality across these specifications exceeds 0.95
in each case, so this choice has little impact on the second stage of our analysis.
Robustness - Sorting within Block Groups. While the correlation analysis presented in Section
4 and the results of the specifications reported in Table 4 provide a great deal of re-assurance
regarding the robustness of our analysis to concerns about the sorting of households across blocks
within block groups, we seek to provide additional evidence that such sorting is not
fundamentally driving the results. To this end, as described in Section 4, Table 5 reports the
results of estimates based on sub-samples based on the 50 percent of block groups that exhibit the
least amount of block-by-block sorting in four dimensions: education, race, the presence of
children in the household, and immigration status. It is important to note, of course, that these
restrictions on the sample change the nature of the set of households for which social interaction
effects are identified so that there is no reason to expect the results to be identical to the full
specification. In our minds, then, this exercise serves mainly as a broad check regarding block-
level sorting.39
The first row of the table again summarizes the results for specifications that do not
include any covariates - either in the levels or interacted with bmatch. In each case, the results
remain similar to the initial regression reported in Table 4, ranging from 0.09 to 0.14 percent.
When covariates are included in the analysis, the main findings related to age, the presence of
children, gender and marital status from our baseline specification are confirmed and, in some
cases, strengthened. Matches between high school graduates continue to lead to strong referral
effects relative to other categories.40 Again, the match quality indices for these specifications have
correlations with the match quality index from specification 2 in table 4 as well as with each other
in excess of 0.90.
In sum, our estimated social interaction effects persist, even in areas that do not
experience a significant degree of sorting below the block group level with respect to
characteristics most likely to be observed at the time a household moves into a block. We believe
39 It should also be noted that these estimates are run using the sample that drops small blocks, but does not
include the housing variables since they had only a minor impact on the estimate correlations in Table 3.
40 Again, the effects for race and immigration status are a bit difficult to evaluate across samples as by
construction, these samples differ significantly in the number of immigrants and racial minorities included
26