An Application of James-Stein Estimation to Survey Data

Abstract

A procedure is examined for adjusting observed values of a sample of entities toward the overall sample mean. This procedure provides estimators with lower expected mean square errors than the ordinary least-squares estimators, at the cost of accepting a slight bias. The procedure was applied to a variety of analysis problems, using data from two sample surveys, and the performance of the adjusted estimators was compared to that of maximum likelihood estimators. While generally the adjusted estimator was somewhat superior to the ordinary estimator, the improvement was too slight to warrant recommending widespread use of the technique.