INVESTIGATION OF THE EFFECT OF MISSING DATA ON DIFFERANTIAL ITEM FUNCTIONING IN MIXED TYPE TESTS

Abstract

This research is aimed to analyze the impact of Markov chain Monte Carlo (MCMC), multiple imputation (MI) and expectation maximization (EM) on differential item functioning (DIF), one of the techniques for coping with missing data in mixed type tests including dichotomous and polytomous items. The study was implemented on a complete dataset consisting of the scores of 1160 students who received booklet number 9 and answered all the questions in the booklet in the science test at TIMSS 2019. The conditions to be examined for the effectiveness of the methods were defined as missing data mechanism (MCAR and MAR), DIF level (A, B and C) and missing data rate (10% and 20%). Over the dataset that mentioned, data have been imputed to the missing data set created with the aid of using deleting data at distinctive rates under missing completely at random (MCAR) and missing at random (MAR) mechanisms by using MCMC, MI and EM methods. DIF analysis was performed by poly-SIBTEST method on all datasets obtained. Accordingly, by the reference of the implications obtained from the complete dataset were compared with the implications of other datasets. As a result of the research, in terms of all the conditions examined the EM and MCMC methods performed better for the C-level DIF than the A and B levels. It has been seen that the most accomplished method for determining DIF in DIF indicative items in 10% and 20% MCAR mechanisms is MI. Compared to the complete dataset, all three methods manifests similar results in the 10% MAR mechanism, while MCMC displayed the closest results in the 20% MAR mechanism compared to results of the other methods.