out the long form of the Census, these data contain the complete set of variables that are available
in the public-use version of the Census PUMS, but, in addition, detail each individual’s
residential and employment locations down to the Census block level. In addition to these
geographic variables, the Census also provides a wide range of sociodemographic information:
age, gender and marital status, education, race, family structure, and duration in the residence as
well as information on labor market outcomes including labor force status, salary and wage
income if employed, occupation, and industry.
With regard to the geographic structure of the data, Census blocks correspond roughly to
actual city blocks; they are typically rectangular regions delimited by the four intersections that
constitute the corners of the block.12 Our sample consists of approximately 25,500 Census blocks
arranged into 2,565 block groups, i.e., an average of 10 blocks per block group. The distribution
of blocks per block group is depicted in Figure 1; the median number of blocks per block group is
8, and about 95 percent of all block groups have 20 blocks or fewer.
It is the precise geographical information for each individual in these restricted Census
data that provides the backbone of our research design, permitting us to isolate the block-level
variation in neighbor exposure by conditioning on block group fixed effects. The first stage of our
analysis considers the propensity of a pair of individuals to work in the same location, comparing
this propensity for a pair that live on the same versus nearby blocks. For this portion of our
analysis, we construct a sample that contains of individuals that (i) are currently employed, (ii)
are between 25 and 59 years of age, (iii) do not live and work in the same block, and (iv) for
whom the Census data on place of work has not been imputed. 13 The total number of workers in
the Census sample that meet these criteria is 129,175 (5.1 per block, 50 per block group). Figure
2 reports the corresponding histogram of workers meeting these criteria across blocks.14
12 Notice that this definition implies that Census blocks are not constituted as the set of buildings that face
each other on the same street. To the extent that social interactions are also strong between residents on
opposite sides of the same street, a comparison of interactions between individuals that reside on the same
Census block versus other blocks in the same block group will tend to understate the increased effect of
immediate neighbors as those on the opposite side of the same street will count in the control group. For
some blocks, however, one may argue that the opposite holds: streets may effectively act as dividers of
local communities, and interactions may be strongest in the alleys and courtyards connecting the rear sides
of buildings on the same block. In either case, our research design should detect (although may understate)
particularly local interactions provided that the block group contains a reasonable number of blocks.
13 Currently employed refers to the reference week in the calendar year 1990 used by the Census. We focus
on prime-age adults in this paper so as to avoid empirical issues related to the labor market participation
versus continued schooling of youths and young adults. We drop all individuals for which place of work is
imputed for obvious reasons. We also drop all individuals that work in the same block in which they reside
to avoid any overstatement of referral effects due, for example, to the clustering of small businesses and
other retail shops on commercial blocks within block groups.
14 In the analysis below, we consider specifications that limit the analysis to blocks with five or more
sample workers.