Several important social science literatures hinge on the functional relationship between
neighborhood characteristics and individual outcomes. Although there have been numerous
non-experimental estimates of these relationships, there are serious concerns about their
reliability because individuals self-select into neighborhoods. This paper uses data from HUD’s
Moving to Opportunity (MTO) randomized housing voucher experiment to estimate the
relationship between neighborhood poverty and individual outcomes using experimental
variation. In addition, it assesses the reliability of non-experimental estimates by comparing
them to experimental estimates.
We find that our method for using experimental variation to estimate the relationship
between neighborhood poverty and individual outcomes – instrumenting for neighborhood
poverty with site-by-treatment group interactions – produces precise estimates in models in
which poverty enters linearly. Our estimates of nonlinear and threshold models are not precise
enough to be conclusive, though many of our point estimates suggest little, if any, deviation from
linearity. Our non-experimental estimates are inconsistent with our experimental estimates,
suggesting that non-experimental estimates are not reliable. Moreover, the selection pattern that
reconciles the experimental and non-experimental results is complex, suggesting that common
assumptions about the direction of bias in non-experimental estimates may be incorrect.