Missing data for longitudinal surveys occur in a variety of patterns which can be sorted and categorized into different classes of missingness depending on the survey unit. For this study, the survey unit is a person. Therefore the missingness that occurs in the data can be person nonresponse, whereby no data are available for a person at any given time period in the survey, record-type nonresponse where an entire module of related data for a person is unavailable, and item nonresponse in which data are missing sporadically throughout the person record. For this study we focused on record-type nonresponse for a single continuous variable. It is important that this type of nonresponse be addressed as it occurs generously throughout a longitudinal survey. Also, simulation of record-type nonresponse provides reasonably sized data files to study and manipulate. It is important to note that the techniques investigated can be employed to compensate for both item and record-type nonresponse.

The objective of this study is to evaluate seven different methods of imputation for continuous data in a longitudinal survey. The methods compared are described below as are the procedures to compare them. In our comparisons, we employed a variety of summary statistics and graphic techniques. The particular findings are detailed in the body of the text and a number of graphs and tables are included in the Appendix to support these findings. No information was observed to support any assumptions of normality in the data studied, and the analysis proceeds using a variety of nonparametric techniques.

In Section II we describe the data used in this study and discuss how it was used. In Section III we discuss each of the alternative imputation strategies that are compared against one another. In Section IV the methods used to compare the different procedures are described and the results of our analyses are presented. Findings are summarized in Sections V and VI, and an Appendix contains the tables, graphs, and summary statistics used in our analyses.