Abstract
Answering essentially any question with sample data requires variance estimates and these estimates depend critically on the sample design. The design information necessary to estimate variances for sample statistics from the US Current Population Survey (CPS) is not publicly released in order to protect respondent confidentiality. To circumvent this problem, the US Census Bureau provides a variance estimation methodology but it is only valid for a few specific point estimates. This paper discusses shortcomings of the Census Bureau methodology and proposes an alternative, general approximation methodology that produces variance estimates for a significantly wider class of statistics, including regression analysis. The proposed approach is based on resorting the data and assigning subsequent observations to synthetic clusters in a manner that creates similarities with the actual CPS sample. The synthetic design approach successfully approximates a baseline for comparison in 34 of the 37 sample estimates considered.