neighborhood.19 Secondly, that it may be difficult for individuals to identify block-by-block
variation in neighbor characteristics at the time of purchase or lease. That is, while an individual
may have a reasonable sense of the socio-demographic structure of the neighborhood more
generally, that variation across blocks within a neighborhood is less easily observed a priori.
The second key assumption is that interactions with neighbors are very local in nature -
i.e., occur mostly among individuals on the same block. To the extent that individuals do have
some interaction with neighbors on surrounding blocks, our design will provide only a lower
bound on the overall strength of local interactions - measuring only the difference between these
very local and broader effects. In this way, the design will allow us to detect interactions provided
that they are significantly stronger at closer distances, but may still understate the strength of
those interactions.
A Diagnostic Test of the Identifying Assumption. To examine whether our first key assumption
- that there is no correlation in unobserved factors affecting work location among individuals
residing on the same block within a block group - is reasonable, we analyze the correlation
between observable individual and neighbor characteristics at the block level. While this kind of
test does not prove anything with respect to the importance of potential correlation in unobserved
factors, it certainly provides an indication of whether this assumption is at all reasonable.20 In
particular, for each block in the sample, a single prime age adult is selected and the characteristics
of other individuals that reside in the same block but not the same household are used to construct
a measure of average neighbor characteristics.21 The first three columns of Table 3 report the
the existence of social effects is based on comparisons across different geographic scales rather than on
correlations with group averages, the simultaneity issue does not arise in our context.
19 In fact, only 11 percent of the blocks in our sample have an owner-occupied unit that changed owners in
the 2 years prior to the Census. Given that the Census is a 1-in-7 sample and assuming a uniform
probability for a house to be on the market in this two year period, this implies that the chances that any
owner-occupied unit is available on a given block within a given 3 month period is only about 11 percent.
Thus, it may be difficult for households searching in a given timeframe to select a house on a particular
block. The comparable figure for renter-occupied units for blocks that contain at least one rental unit in our
sample is 45 percent. This suggests that it is generally easier, although far from certain, for renters to find
housing on a specific block within a particular search window.
20 This is in the same vein of Altonji et al. (forthcoming): their approach to correct for selection bias
suggests that selectivity in terms of unobserved heterogeneity is in some sense proportional to selectivity on
observables.
21 By sampling only one individual per block, we avoid inducing a mechanical negative correlation that
would come about if all individuals were used in estimating the correlation between individual and average
neighbor characteristics. This negative correlation arises because each individual is counted as a neighbor
for all of the others in the same block, but not for herself. For estimates of the correlation that do not
condition on block group fixed effects, this bias is inconsequential because an individual’s own
characteristics contribute very little to the average neighborhood characteristics of others in the full sample.
For estimates that condition on block group fixed effects, however, this negative bias is quite large in
11