This article provides an overview of the issues related to missing data in research in higher education. The authors point out that much of the literature on handling missing data is in the fields of psychology and sociology and the majority of research in higher education does not mention methods for dealing with missing data.
The authors suggest that traditional methods of listwise deletion, pairwise deletion, imputation, and dummy-variable adjustment are all inadequate as they lead to biased results that underestimate the variance in parameters and standard errors.
They suggest that maximum likelihood (ML) and multiple imputation (MI) be used whenever possible to provide more accurate estimates of the population. If possible, Full-Information Maximum-Likelihood (FIML) procedures should be used if interaction variables are of interest. The authors recommend that all studies use MI as the default method of dealing with missing data if possible unless the context calls for a different approach.
They then complete all of these missing data procedures on an actual dataset from higher education, arriving at the conclusion that EM was the most appropriate approach due to their context, but that MI is usually the most appropriate approach, particularly if the data will be subject to regression analysis.