Assumption-Lean Shrinkage and Model Averaging for Spatial Parameters

Abstract

Economic decisions often depend on many noisy estimates of quantities such as neighborhood effects, school quality, and hospital performance. Shrinkage estimation can improve decisions by pooling information across related units, but geography, adjacency, and shared characteristics each define a different notion of relatedness, and each implies a different way of pooling. We treat the choice of relatedness as part of the estimation problem, using Stein's Unbiased Risk Estimate (SURE) to form a weighted average over a library of flexible shrinkage estimators. This comparison among the candidate estimators treats no prior or latent covariance structure as a correctly specified model for the parameters being estimated. Each candidate is judged by its SURE value. Under smoothness conditions on the estimators, the SURE-weighted average performs nearly as well as the best fixed weighted average of trained candidates, including nonlinear rules whose reported values use the full vector of noisy estimates. In an application to Opportunity Atlas economic mobility data from 20 commuting zones, the best individual spatial specification varies across zones, yet the SURE-weighted average tracks the best in each zone and reduces estimated mean squared error by about 27% relative to the best-performing non-spatial empirical Bayes baseline in our library of estimators.