Alternative Variance Estimates in the Current Population Survey and American Community Survey

Michel Boudreaux, University of Minnesota
Michael Davern, University of Minnesota
Peter Graven, University of Minnesota

The Current Population Survey (CPS) and American Community Survey (ACS) are based on complex samples. As such, variance and standard error estimates need to be adjusted. Assuming a simple random sample (SRS) (the default of most statistical packages) will result in biased standard errors and statistical tests of significance. We describe and evaluate alternative variance estimation methods that accommodate sample design complexity. Using successive difference replication (SDR) as a gold-standard we evaluate various alternatives including a Taylor series (TS) approach, Census Bureau variance parameters, and the SRS method. For most variables considered, in both surveys and across geographic domains, the TS preformed better than design factor method. These results show that analysts wishing to avoid the cumbersome computation of design factor based standard errors can use a Taylor Series approach based on publically available geographic variables.