Using an original administrative dataset in the context of a scarcity induced-natural experi-ment in New York City, I find that families placed in shelters in their neighborhoods of origin remain there considerably longer than those assigned to distant shelters. Locally-placed families also access more public benefits and are more apt to work. A fixed effects model assessing multi-spell families confirms these main results. Complementary instrumental variable and regression discontinuity designs exploiting policy shocks and rules, respectively, suggest difficult-to-place families – such as those that are large, disconnected from services, or from neighborhoods where homelessness is common – are especially sensitive to proxi-mate placements. Better targeting through improved screening at intake can enhance pro-gram efficiency. The practice of assigning shelter based on chance vacancies ought to be replaced with a system of evidence-based placements tailored to families' resources and constraints.