The economic census is the U.S. Census Bureau’s official five-year measure of American business and the economy. It provides industry and geographic detail not typically available from other economic statistics sources benefitting businesses, policymakers and the American public.

The term “census” in this case is actually a slight misnomer. The Census Bureau requests data from most large businesses and a sample of small businesses. We ask each of these businesses to provide data on sales, shipments, and receipts or revenues for each of its establishments (i.e. for each single physical location)—as shown in Figure 1.

We also ask for the revenues obtained by each establishment from the types of products likely to be produced or sold based on its primary industry. Product statistics are needed by the Bureau of Economic Analysis to benchmark the national accounts, as well as by the Bureau of Labor Statistics in constructing producer price indexes. The North American Product Classification System defines over 8,000 different products that can be reported across the entire census.

As an example, Figure 2 provides a short extract from the product collection for establishments in the “Automobile Dealers” retail trade industry from the 2012 Economic Census. Notice that, on the surface, these products don’t seem to be related to automobile dealers, but they are products that could be found at automobile dealerships, and that is why they are included on the questionnaire. The product list for some establishments can span more than 50 potential products. Additionally, for certain industries Census designates “must-have” products. For example, an automobile dealer should report revenue from automobile sales.

In most industries, only a few products are frequently reported and many sampled establishments do not report any data on products. This makes it difficult to produce good product statistics and measures of reliability.

For the past two years, the Census Bureau has conducted extensive research into product statistics. Initial research by the team focused on determining a single missing data treatment method for products in the 2017 Economic Census. The research, presented in a topic contributed session entitled “Evaluating Alternative Imputation Methods for Economic Census Products: The Cook-Off” was reported at the 2015 Joint Statistical Meetings.

This year, we have been exploring how to estimate the variance for product sales. Besides the sampling, imputation and post-stratification components, there are additional challenges caused by the lack of good predictors and high expected zero rates for many products, compounded by the high product nonresponse rates. We believe that it is possible to find a variance estimator with good statistical properties for the well-reported products, but we remain concerned about the others. So far, the team has conducted two separate simulation studies that investigate the possibility of finding a variance estimator that performs well on many different products considering only (1) sampling variance and post-stratification, and (2) product nonresponse and hot deck imputation. We will share these results on August 1, 2016, at the JSM. The next phase of our research will combine the findings from the two separate studies to develop a single variance estimator for products.