155 – Implementing Research-Based Recommendations in Ongoing Programs
Variance Estimation for Product Sales in the 2017 Economic Census: Challenges in Implementing Multiple Imputation-Based Variance Estimation
Matthew Thompson
U.S. Census Bureau
Katherine J. Thompson
U.S. Census Bureau
The U.S. Census Bureau conducts an Economic Census every five years, producing key measures of American business and the economy. In addition to a core set of items collected from all establishments, the Economic Census requests information on the revenue obtained from products sold from all large businesses and a probability sample of smaller businesses. The 2017 Economic Census will – for the first time – publish variance estimates for product sales estimates. A research team was established to recommend a variance estimation method that accounted for variance both due to sampling and imputation. The team's evaluative approach relied on simulation, using empirical data from a purposively selected, small subset of industries as the basis for the study. The research was complicated by the nature of product data, which are characterized by poor item response rates, few available predictors, additivity-within-establishment requirements, and many rarely reported products in an industry. The research team considered several alternative variance estimators on this limited number of industries and a subset of reported products, ultimately recommending a multiple imputation method that utilizes the Finite Population Bayesian Bootstrap (FPBB) to address the sampling variance and the Approximate Bayesian Bootstrap (ABB) to incorporate variance due to imputation. The implementation of this proposed approach unveiled a number of modifications and enhancements needed to accommodate the complete set of variables. This paper describes the recommended variance estimation method and how this method is being implemented into the 2017 Economic Census production system. Examples are provided to illustrate implementation issues and the modifications and enhancements needed to fully implement the research-based recommendations.