On Recommending a Single Imputation Method for Economic Census Products

• 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): Katherine Jenny Thompson* and Xijian Liu
• Companies: U.S. Census Bureau and U.S. Census Bureau
• Keywords: response propensity analysis ; rank-based statistics ; imputation error ; nonparametric tests

Abstract

The U.S. Census Bureau conducts an Economic Census every five years, producing key measures of American business and the economy. Besides collecting a set of common items from all eligible establishments, the Economic Census collects detailed information on each establishment's products. Beginning in 2017, the Economic Census will use the North American Product Classification System (NAPCS) to produce economy-wide product tabulations from cross-sector collections. This marks a major departure from the current collection - which explicitly links products to industry - and makes the trade-area specific missing data adjustment practices impossible. Instead, we sought a single imputation method for all industry products that is statistically defensible and operationally practical. The research is complicated by the nature of product data, which are characterized by poor item response rates, few available predictors, additivity-within-establishment requirements, and different units of collection. In this paper, we describe our decision-making process, briefly presenting selected empirical analyses before focusing on the evaluation methods used for a comprehensive simulation study.