Differentiated Effects of Data Analyses on Between- and Within-Imputation Variances in Multiple Imputation

• Activity Number: 472 - Imputation and Nonresponse Bias
• Type: Contributed
• Date/Time: Wednesday, August 2, 2017 : 8:30 AM to 10:20 AM
• Sponsor: Survey Research Methods Section
• Author(s): Qiyuan Pan* and Rong Wei and Yulei He
• Companies: NCHS/CDC and NCHS / CDC and National Center for Health Statistics
• Keywords: Multiple imputation ; Impact of multiple imputation ; Missing data ; Between-imputation variance ; NAMCS

Abstract

In multiple imputation (MI), the total variance (T) is estimated by U+(1+1/m)B, where U is the within-imputation variance, B the between-imputation variance, and m the number of imputations. The expected value of U is not affected by a proper MI, whereas the extra variance B can be captured only by MI but not by single imputation (SI). Whether B is large enough to cause a meaningful change in T may have an effect on people's perspective towards the value of MI as compared to SI. This paper evaluates how data analysis affects the impact of MI (IMI), measured as IMI = 100(B/T)1/2. MI trials were conducted using the data of the 2012 Physician Workflow Mail Survey. Difference in analytic models had differentiated effects on B and U. Our results suggest that, for the same MI and the same data, IMI may be negligible (<1%) in one analysis but substantial (>5%) in another.