Does Imputation Increase Statistical Power?

• Activity Number: 320
• Type: Contributed
• Date/Time: Tuesday, August 6, 2013 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Wenyaw Chan*+ and Xiaoying Yu and Elaine Symanski
• Companies: The University of Texas-Health Science Center at Houston and The University of Texas- Health Science Center at Houston and The University of Texas-Health Science Center at Houston
• Keywords: imputation, missing data, power

Abstract

It is widely believed that imputation increases the power of a statistical test when missing observations occur. In this research, we use three well-known statistical tests as examples and demonstrate that imputation does not increase power even when data are missing completely at random (MCAR). The three scenarios chosen are (1) the one-sample t test, (2) the two-sample t test when variances are equal and (3) simple linear regression. Under each scenario, we examine three different imputation methods: (a) imputed by a sample from the observed, (b) imputed by the predicted mean and (c) imputed by drawing a sample from the predicted distribution. For scenario (3), (a) is not used because it is obviously biased. Analytic reasoning and computer simulations are used to evaluate whether imputation improves the power of statistical tests. For all scenarios and methods of imputation, the results show no gains in power when data are imputed relative to that achieved using only the complete-case data. The power for imputation method (b) is comparable with the power for no imputation, while the power for each of the other two imputation methods is slightly lower.