Akademik Veri Yönetim Sistemi
JOURNAL OF MEASUREMENT AND EVALUATION IN EDUCATION AND PSYCHOLOGY-EPOD, cilt.8, sa.2, ss.200-223, 2017 (ESCI)
The purpose of this study was to examine the effects of missing data handling techniques on model data fit and item model fit in the one parameter logistic Item Response Theory Model. For this purpose, data sets with sample sizes of 500, 1000, and 1500 and with 20 items that fit to one parameter logistic model were generated. Item difficulty values of the items in the generated data sets ranged from -2 to +2 and item discrimination was fixed as 1.5. The generated complete data sets were exposed to deletion at %5, %10, and %15 under missing at complete random (MCR) and missing at random (MR) conditions. Missing at complete random mechanism was obtained as a result of random values deleted among the total number of cells in the data set. A particular percentage of random units (individuals) were deleted for listwise deletion method. Missing at random mechanism was reached as a result of random deletion of cells pursuant to defining a three level variable in the data set at the following percentages: 20% at Level 1, 30% at Level 2 and 50% at Level 3. The generated missing data were resolved using listwise deletion method (LM), regression imputation, and expectation maximization algorithm (EMA). -2 log lambda, AIC, and BIC evaluation criteria were used for model data fit estimation and chi(2) statistics were used for item model fit estimation. Values obtained from the complete data sets were taken as reference for predictions in the data sets that were completed with the effect of missing data handling techniques. As a result of the examinations, it was concluded that expectation maximization algorithm had good performance in missing at random mechanism but partially good in missing at complete random mechanism. It was also seen that regression imputation had good performance under certain conditions but the performance of listwise deletion method was poor. In all missing data mechanisms, the performance of the effect of missing data handling techniques declines as missing data increase. It is certain that a single method to give best results in all mechanisms and under any conditions is unlikely to be assumed.