Multiple Imputation for Nonresponse

Abstract

For many datasets, especially for nonmandatory surveys, missing data are a common problem. Deleting units that are not fully observed and using only the remaining units is a popular, easy-to-implement approach in this case. However, using only fully observed observations will generally lead to reduced efficiency for the estimates. But even more problematic, this approach can possibly lead to severe bias if the strong assumption of a missing pattern that is missing completely at random (MCAR; see Section 5.2) is not fulfilled. Imputing missing values can help handle this problem. However, imputing missing values only once (single imputation) generally doesn’t account for the fact that the imputed values are only estimates for the true values. After the imputation process, they are often treated like originally observed values, leading to an underestimation of the variance in the data and from this to p values that are too significant. Multiple imputation was suggested by Rubin (1978) to overcome these problems.

Most of this chapter is taken from Drechsler and Rässler (2008) and Drechsler (2011a).