|
Abstract:
|
The Economic Census collects data on the revenue obtained from products (product data) from all sampled units. In the 2017 Economic Census, missing product data will be imputed using hot deck imputation, and variance estimates for product data will be published for the first time. The majority of Economic Census sectors utilize a probability subsample of establishments and employ post-stratification to account for industry classification and sampling deficiencies. Product data pose unique challenges. Often sampled establishments elect not to provide any values (complete nonresponse) and many products are rarely reported. Consequently, the variance estimator must account for sampling variance, post-stratification, and imputation variance. Previous research examined the variance estimation effects of sampling variance and post-stratification separately from imputation variance. The recommended multiple imputation variance estimator combines the finite population Bayesian bootstrap (FPBB) with the approximate Bayesian bootstrap (ABB). Using a simulation study, we evaluate the performance of this variance estimator with two hot deck imputation methods (nearest-neighbor and random),
|
