Methods for Handling Missing Data

Abstract

This chapter describes a general approach to handling missing data in psychological research. It provides a theoretical background in readable, nontechnical fashion. Our overall goal was to give practical, usable advice, rather than to give a detailed statistical treatment of issues surrounding analysis of incomplete data. We give an overview of the older, unacceptable methods for handling incomplete data, so that readers will know what approaches to avoid; although analysis of complete cases is sometimes an acceptable solution, we argue that pairwise deletion and mean substitution should be avoided. With respect to newer, acceptable methods, we give a general overview, including a brief discussion of the full information maximum likelihood structural equation modeling procedures (such as Amos, Mx, LISREL 8.5, and Mplus), but focus primarily on multiple imputation as a general solution. We give specific guidelines for making use of state‐of‐the‐art multiple imputation software and step‐by‐step instructions for using multiple imputation with Schafer's (1997) NORM program. Empirical examples of exploratory and data quality analyses and a substantive illustration involving multiple linear regression demonstrate the use of multiple imputation in practice. The chapter concludes with a discussion of some practical issues that often arise in connection with the analysis of incomplete data.

Interest development: Arousing situational interest affects the growth trajectory of individual interest

“…There were two types of missing data: (1) missing data for individual items (e.g., failed to respond to one or more items) and (2) missing data due to absence of students at one or more lessons (e.g., illness). The first type of missing data constituted less than 2% for the individual interest measurements and less than 1% for the situational interest measurements and therefore did not constitute a problem (Graham, Cumsille, & Elek-Fisk, 2003). The magnitude of the second type of missing data was as follows: Week 1 = 1%, week 2 = 5%, week 3 = 4% and week 4 = 2%.…”

General practitioner empathy, patient enablement, and patient-reported outcomes in primary care in an area of high socio-economic deprivation in Scotland—A pilot prospective study using structural equation modeling

“…It is one of the recommended methods for preventing biases caused by not completely random missing data processes [22,23]. The imputation was performed with the software NORM [24].…”

The role of positive psychology constructs in predicting life satisfaction in adolescents with and without cognitive disabilities: An exploratory study

“…The total percentage of missing data values was 1.1% for students with disabilities and 2.1% for students without disabilities. Because of the potential deleterious effects of not including all available data in the analysis process, we used the EM imputation algorithm using the PROC MI procedure within the SAS program (Graham, Cumsille, & Elek-Fisk, 2003). In so doing, we used the totality of information within our data set to impute the missing data, and therefore maintained important characteristics of the data set, improving our ability to calculate unbiased and efficient parameter estimates (Graham et al, 2003).…”