Evaluating Performance of Missing Data Imputation Methods in IRT Analyses

Missing data is a common problem in datasets that are obtained by administration of educational and psychological tests. It is widely known that existence of missing observations in data can lead to serious problems such as biased parameter estimates and inflation of standard errors. Most of the missing data imputation methods are focused on datasets containing continuous variables. However, it is very common to work with datasets that are made of dichotomous responses of individuals to a set of test items to which IRT models are fitted. This study compared the performances of missing data imputation methods that are IRT model-based imputation (MBI), Expectation-Maximization (EM), Multiple Imputation (MI), and Regression Imputation (RI). Parameter recoveries were evaluated by repetitive analyses that were conducted on samples that were drawn from an empirical large-scale dataset. Results showed that MBI outperformed other imputation methods in recovering item difficulty and mean of the ability parameters, especially with higher sample sizes. However, MI produced the best results in recovery of item discrimination parameters.