Americans. Ioannides and Loury (2004) provide a detailed survey of the literature on job
information networks.
Only a small number of studies attempt to quantify the impact of specific social
interactions or exposures on outcomes,11 and these studies tend to be outside of the labor market
context. Bertrand, Luttmer, and Mullainathan (2000) examine the relationship between an
individual’s own welfare participation and the welfare participation rate among those who speak
the same language as this individual. They find a strong positive relationship suggesting the
existence and impact of language specific social networks. Similarly, Aizer and Currie (2004)
find evidence that the prenatal care use of pregnant women is most likely to be influenced by the
behavior of new mothers in the same ethnic group as compared to mothers in different ethnic
groups who reside in the same neighborhood. In the labor market context, Topa (2001) finds that
employment in a given Census tract is positively affected by average employment in neighboring
tracts when these tracts are located within a common, larger local community (as defined by their
residents).
Our paper adds generally to the body of evidence suggesting that social networks have a
substantial impact on labor market outcomes, and more specifically to this small, but very
important literature, on the heterogeneous use of social contacts by individuals and how that use
differs with respect to their observed characteristics. Our analysis indicates that there is
considerable variation in the likelihood of successful referrals across different pairs of neighbors.
Further, this heterogeneity in referral effects enables us to construct a proxy for match quality at
the block level that we use in the second stage of our analysis to quantify the economic impact of
referrals. Our research design and unique dataset allow us to focus very closely on a specific
mechanism through which social interactions at the local level may operate, namely referrals and
information about job opportunities, while still carefully addressing methodological concerns
arising from sorting across neighborhoods.
3 DATA
The data for our analysis are drawn from a restricted version of the 1990 US Census of
Population for the Boston metropolitan area. For the full (1-in-7) sample of individuals that filled
11 Even if such analyses were conducted using referral data, the results would quite likely be based on self-
reported networks that arise from individual choices. The studies cited below look at the impact of
exposure to possible social networks, which are presumably less endogenous than the actual networks
accessed by the individual. The key exception to this statement is Marmaros and Sacerdote (2002), who
base their analysis on exposure to a randomly assigned roommate in a college dormitory. See Arcidiacono
and Vigdor (2004) and Weinberg (2004) for recent studies that document sorting/assortive matching in the
process of forming social networks using data on college and high school students, respectively.