Evaluation of Alternative Imputation Methods for Economic Census Products: The Cook-Off

• Activity Number: 409
• Type: Topic Contributed
• Date/Time: Tuesday, August 11, 2015 : 2:00 PM to 3:50 PM
• Sponsor: Survey Research Methods Section
• Author(s): Jeremy Knutson* and Jared Martin
• Companies: U.S. Census Bureau and U.S. Census Bureau
• Keywords: fraction of missing information ; ranking ; imputation error

Abstract

In preparation for the 2017 change to the North American Product Classification System (NAPCS), Economic Census staff was tasked with determining a single imputation method to treat missing product data collected from all trade areas. To objectively compare four proposed imputation methods, we conducted a simulation study to obtain two evaluation measures: imputation error (IE), to measure the accuracy of the overall estimate, and the fraction of missing information (FMI), to measure the precision of the imputed estimate. For the "cook-off," we generated complete pseudo populations by applying each imputation method to missing sample data, inducing product nonresponse in each population, and applying each imputation method to the missing data. Nonresponse was induced independently in each pseudo population, yielding 50 replicates. Each imputation procedure was multiply-imputed within replicate. Imputation methods ("treatments") are evaluated within trade area using the averaged IE and FMIs. This evaluation approach is generalizable to other programs with similar missing data problems.