Bootstrap Variance Estimation in the Presence of Imputed Data

• Activity Number: 483
• Type: Contributed
• Date/Time: Wednesday, August 4, 2010 : 8:30 AM to 10:20 AM
• Sponsor: Section on Survey Research Methods
• Author(s): Christian Léger*+ and Zeinab Mashreghi and David Haziza
• Companies: Université de Montréal and Université de Montréal and Université de Montréal
• Keywords: variance estimation ; bootstrap ; imputation ; sampling weights

Abstract

In the presence of imputed data, Shao & Sitter (1996) introduced a bootstrap method where the non-respondents in the bootstrap sample are reimputed with the same method, which leads to a valid variance estimator if the sampling fraction is small. We argue that this method can be justified by the reverse framework for variance estimation developed by Shao and Steel (1999). When the sampling fraction is small, this framework suggests a second valid bootstrap variance estimator that can be obtained using complete data bootstrap methods (e.g., the method of Rao, Wu and Yue, 1992) and does not require reimputing. Finally, we obtain a third variance estimator, valid even when the sampling fraction is not small, by modifying the weighting to account for non-response in the bootstrap weights approach and resampling the data and the response status independently rather than jointly.