Evaluating Imputation and Estimation Procedures in a Survey of Wholesale Businesses

• Activity Number: 163
• Type: Topic Contributed
• Date/Time: Monday, August 10, 2015 : 10:30 AM to 12:20 PM
• Sponsor: Government Statistics Section
• Author(s): Joanna Lineback and Martin Klein* and Joseph L. Schafer
• Companies: U.S. Census Bureau and U.S. Census Bureau and U.S. Census Bureau
• Keywords: Business survey ; confidence interval coverage ; estimation ; longitudinal survey ; modeling ; simulation

Abstract

The Monthly Wholesale Trade Survey (MWTS) provides estimates of change in inventories and sales for wholesale businesses in the United States. In previous work, we developed a new procedure for multiply imputing values of sales and inventories for nonrespondents based on a multivariate linear mixed model. We conducted a simulation study, generating an artificial population with frame data and two years of monthly sales and inventory figures, and we observed how the current MWTS Horvitz-Thompson and random group estimators performed over repeated samples with and without missing data. We discovered that the complete-data point estimates, variance estimates and interval estimates did not behave as large-sample normal theory suggests they should. Over repeated samples, point estimates were highly skewed and interval estimates had poor confidence coverage. In this paper, we review our findings to date and describe ongoing research into new estimation procedures for Bayesian finite-population inference with stratified samples from highly skewed populations.