138 – Nonresponse Adjustment - 2
Evaluating Imputation Techniques in the Monthly Wholesale Trade Survey
Martin Klein
U.S. Census Bureau
Joanna Fane Lineback
U.S. Census Bureau
Joseph L. Schafer
U.S. Census Bureau
The Monthly Wholesale Trade Survey (MWTS) provides estimates of the dollar value of sales and inventories for wholesale businesses in the United States. In this longitudinal survey, missing values for sales and inventories in any month are imputed via a ratio adjustment applied to data from the prior month. In this article, we describe ongoing research to evaluate the performance of the current imputation method and to investigate possible alternatives. Using information from the MWTS and the sample frame, we generated an artificial population of wholesale businesses with two years of monthly sales and inventory data. We repeatedly drew samples from this artificial population, imposed patterns of nonresponse on the samples, and filled in the missing values by two methods: the current ratio-based procedure and a new model-based multiple-imputation procedure. Preliminary results from this simulation are challenging to interpret because, apart from missing data, the inferential procedures (complete-data point estimates, variance estimates and interval estimates) do not behave as large-sample normal theory suggests they should. Based on these results, we recommend further research on improving the quality of the complete-data inferences, using methodologies that are better suited for stratified sampling from populations that are highly skewed.