Effect on Secondary Data Analysis of Common Imputation Methods

We investigate the properties of seven commonly studied imputation methods from the perspective of the secondary data analyst who assumes that the data set to be analyzed has only observed responses. Because of the complexity of the algebra, we consider a simple specification: The data analyst wishes to make inferences about the slope of a linear regression, and any missing data are missing at random. We investigate the biases of the customary estimators of the slope and residual variance and the quality of the usual confidence interval for the slope. We compare the seven imputation methods and discuss the implications of our findings for secondary data analysis.