Abstract:
|
The Economic Census collects information on the revenue obtained from product sales. Product sales are only collected for a sample of establishments, and variance estimates are needed for these sample-based estimates. In any given industry, establishments can report values from a wide variety of potential products. Often, product descriptions are quite detailed, and many products are mutually exclusive. Consequently, legitimate missing values occur frequently. The reported product dollar values are expected to sum to the total receipts reported earlier in the questionnaire. Variance estimation of these product values will need to account for a number of factors including sampling variance, variance due to imputation, and the effects of calibration weighting. After reviewing available literature, three bootstrap methods adjusted for complex survey designs and a Bayesian model-based approach were identified as candidate variance estimation techniques. This paper will describe each of the candidate methodologies. A simulation study will be used to evaluate the application of each method in a complete response scenario for Horwitz-Thompson estimates and ratio estimates. Future research will evaluate each method in the presence of non-response and measure the impact of variance due to imputation.
|