Methods for Handling Missing Data

An eight-step process for performing basic missing data procedures (single and multiple imputation) using NORM and SAS and including auxiliary variables to the missing data model that are highly correlated with variables containing missingness is a good way to reduce the standard errors of estimated parameters.

Abstract

The field of missing data analysis has evolved quickly; outdated strategies have been replaced by missing data procedures that make use of partial data to provide unbiased and efficient parameter estimates. Here, we describe an eight-step process for performing basic missing data procedures (single and multiple imputation) using NORM (Version 2.03; Schafer, 1997) and SAS (Version 9.2; SAS, 2011a). We make use of the automation utility, MIAutomate, to facilitate the analysis of NORM-imputed data with SPSS. We also perform a basic multiple regression to compare the performance of the following five missing data approaches: (1) complete cases analysis, (2) mean substitution, (3) FIML-based Amos, (4) multiple imputation with no auxiliary variables (MI), and (5) multiple imputation with auxiliary variables (MI+). Mean substitution showed clear bias in parameter estimates. Although the complete cases model yielded parameter estimates similar to those obtained with MI and FIML, the standard errors were consistently larger because the estimation was based on fewer cases. The MI and FIML models yielded reasonably unbiased estimates of the b-weights; however, they had noticeably lower t-values compared to the MI+ analysis. Based on prior research and the empirical results shown in this chapter, including auxiliary variables to the missing data model that are highly correlated with variables containing missingness is a good way to reduce the standard errors of estimated parameters. The prescribed missing data procedures will always perform at least as well as more traditional approaches (e.g., complete cases analysis), and in most circumstances, there will be a clear advantage of the prescribed procedures, in terms of estimation bias and/or statistical power. Useful and accessible missing data procedures are now an integral part of many mainstream statistical packages, making it increasingly difficult to justify not using them.