Accounting for Imputation When Estimating Variances in the Economic Surveys at the Census Bureau

• Activity Number: 16
• Type: Contributed
• Date/Time: Sunday, August 11, 2002 : 2:00 PM to 3:50 PM
• Sponsor: Section on Survey Research Methods*
• Author(s): Patrick Cantwell*+
• Affiliation(s): U.S. Census Bureau
• Keywords: inflation factor ; pseudo-values ; regression imputation ; ratio imputation

Abstract

Imputation is commonly used in surveys to address item nonresponse. However, variance estimates are often computed as if the imputed values were valid responses. The true variance of such a survey estimate tends to increase with the amount of nonresponse (Cochran 1977). Depending on the type of imputation applied, the expected value of the resulting variance estimate is often smaller than that without imputation (Särndal 1992; Korn and Graubard 1999).

Processing for the Census Bureau's economic surveys is handled by a generalized system. Within this system, we investigate approaches that apply an inflation factor to the usual variance estimate to approximate the true variance under several imputation methods. We then consider procedures where the usual variance estimator is applied to pseudo-values created according to the imputation used, as suggested by Kim (2001). When comparing alternatives, several factors are weighed: 1.) the accuracy of the resulting variance estimates; 2.) the ease of applying the procedure to compute the variance; and 3.) the ability to generalize the method to various types of imputation.