Variance Estimation for Finite Populations with Imputed Data

One way of handling survey nonresponse is to impute data for each nonrespondent. When estimating sampling variances, however, treating the imputed data as a complete set frequently leads to underestimates of the true sampling variance. Techniques have been recently developed to yield valid variance estimates in the presence of imputed data for some estimators and sample designs. Economic surveys frequently deal with highly skewed populations and employ high sampling rates, including selection with certainty for large units. It is also common for economic surveys to have administrative or historical data available for use in imputation. This paper describes a Monte Carlo study of variance estimation on skewed populations with high sampling fractions in some strata. We examine a variety of imputation techniques and patterns of nonresponse. We extend the Rao-Shao technique to finite populations and nearest neighbor imputation, and compare the resulting estimators to the true variance.