Kayıp verinin test eşitlemeye etkisinin incelenmesi
Özet
The purpose of this study is to compare equating error (RMSE) and equating bias (BIAS) values with regard to estimated item and ability parameters obtained as a result of equating scores of tests in different conditions which created according to missing data location, mechanism and handling methods using IRT-based Stocking-Lord method within anchor test design.
In accordance with this purpose, a data generation process which consisted three stage were followed in order to examine the effects of missing data problem on test equating. In the first stage, complete data sets which composed binary items fit to 3PL model were generated to create each test form. In the second stage, missing data sets which have three different missing data location (new test-NT, both test-BT, anchor test-AT), three different missing data mechanism (missing completely at random-MCAR, missing at random-MAR, missing not at random-MNAR) and three different missing data rate (10%, 20%, 40%) were created with using data deletion algorithms on complete data sets which generated in the first stage. In the third and last stage, data sets which handled missing data problem were reached with using four handling methods (treating as not administered-TNA, treating as incorrect-TI, logistic regression-based multiple imputation-LRMI, discriminant function-based multiple imputation-DFMI) on missing data sets.
After data generation, equating process were conducted. In this context, equatings were carried out separately on reference condition whose test forms that have complete data sets and 108 (3x3x3x4) different simulation conditions whose test forms that were obtained by manipulating of complete data sets. To sum up, total 109 equating process were performed. 50 replication were done one by one to entire analysis process which conducted through R package. At the stage of evaulating of analysis results, average equating error (RMSE) and average equating bias (BIAS) values which obtained by equatings on every simulation condition were reported. And then these values and the values which obtained by equatings on reference condition were compared and looked how close they were.
The results obtained from study showed that BT had the least equating error and the least equating bias values while NT had the most equating error and the most equating bias values in terms of missing data location factor, MCAR had the least equating error and the least equating bias values while MNAR had the most equating error and the most equating bias values in terms of missing data mechanism factor, 10% had the least equating error and the least equating bias values while 40% had the most equating error and the most equating bias values in terms of missing data rate factor. The results about missing data handling methods showed that DFMI had produced the closest and the lowest average RMSE and BIAS values to the average values of reference condition. Results also showed that TNA and TI methods which frequently used to handle missing data problem in binary items produced inaccurate and biased estimations on test equating.