A Comparison of Linking Methods for Multidimensional IRT Calibrations

• Main Authors: Yuan-sung Wang, 王苑菘
• Other Authors: Hueying Tzou
• Format: Others
• Language: zh-TW
• Published: 2010
• Online Access: http://ndltd.ncl.edu.tw/handle/92908781964782209346

碩士 === 國立臺南大學 === 測驗統計研究所碩士班 === 98 === When item responses do not satisfy the assumption of unidimensionality, one should consider the multidimensional item response theory (MIRT) for modeling the relationship of item-examinee interaction. With the application of MIRT, however, the fact of establishing a common scale based on multiple dimensions is still an issue. In this study, the concurrent MIRT calibration and separate MIRT calibration with linking are examined. The purpose of this study is to investigate the performance of the concurrent calibration and two separate MIRT linking methods, using simulated data under the conditions of different sample sizes, test structures, and ability distributions. More specifically, the study compares the linking results from these three methods by evaluating the accuracy of the item parameter estimates, where the average differences between the true parameters and the corresponding estimates (Bias) as well as the root mean square errors (RMSE) are examined. The results of this study show that the concurrent calibration performs better than separate calibration with linking given the same ability distribution for both groups. When the average abilities of two groups are different, the recoveries of discrimination parameter from the three methods are fairly good and the TCF and the Procrustes methods perform better than the concurrent calibration for the difficulty parameter. Under the condition of two related dimensions, the concurrent calibration with Promax rotation performs better than the separate calibration with linking for the discrimination parameters; the concurrent calibration performs better for the difficulty parameter. As for unequal variance between two groups, the separate calibration with linking performs better for the discrimination parameter and the concurrent calibration performs better for the difficulty parameter. The larger the sample size is, the better the performance of theseparate calibration with linking is. Given the results, the test structure does not affect the performances of the multidimensional linking methods under the onditions of different sample sizes. However, the complex structure obviously results in worse recovery.