How Many Imputations Are Needed to Stabilize Imputation-Based Study Results?

• Activity Number: 132
• Type: Contributed
• Date/Time: Monday, July 30, 2012 : 8:30 AM to 10:20 AM
• Sponsor: Biopharmaceutical Section
• Author(s): Kaifeng Lu*+
• Companies: Forest Labs
• Keywords: conditional variance ; multiple imputation ; randomized clinical trials

Abstract

Multiple imputation procedure replaces each missing value with a set of plausible values based on the posterior predictive distribution of the missing data given the observed data. In many applications, as few as 3-5 imputations are sufficient to achieve high relative efficiency. However, empirical evidence suggests that substantially more imputations are often needed to stabilize the results. In this talk, I present the conditional variance for the multiple imputation estimate given the observed data. In the context of randomized clinical trials, I express the number of imputations needed to control the conditional standard deviation within a threshold as a function of the fraction of missing information and the study power to detect the treatment difference. I show that hundreds of imputations may be needed so that imputation-based study results are virtually independent of the random seed that initializes the imputation procedure.