Advances in Longitudinal Survey Methodology

17Proper Multiple Imputation of Clustered or Panel Data

Martin Spiess¹, Kristian Kleinke², and Jost Reinecke³

¹Department of Psychological Methods and Statistics, University of Hamburg, Hamburg, Germany

²Department of Education Studies and Psychology, University of Siegen, Siegen, Germany

³Faculty of Sociology, University of Bielefeld, Bielefeld, Germany

17.1 Introduction

Most datasets are affected by missing values, i.e. values intended to be surveyed but are not observed. In contrast to these unintended missing values there may also be planned missingness like, e.g., in split questionnaire designs. These are examples of item nonresponse. An extreme form of item nonresponse is unit nonresponse where selected units are completely missing. In addition, panel data are affected by attrition, i.e. units observed in wave t are not observed in later waves.

In this chapter, we consider multiple imputation (MI) as a method to compensate for missing items. If the number of cases dropping out is large, however, compensating using weights might be more convenient in this situation but will not be treated here. Combining both methods is, for example, considered in Spiess (2006) and some promising results are presented in Seaman et al. (2012b).

If a dataset is affected by missing values then, in the analysis step, assumptions about the process leading to nonresponse are made either implicitly (implicit strategy), e.g. by accepting the default techniques of many software packages like analysing ...