IRT-based procedures using common items are widely used in test score linking. A critical assumption of linking methods is the invariance property of IRT. This assumption is that item parameters remain the same on different testing occasions when they are reported on the same è-scale. In practice, however, there are occasions when an item parameter drifts from its original value. This study investigated the impact of keeping or removing linking items that were showing item parameter drift. Simulated data were generated under a modified three-parameter logistic model with common item non-equivalent group linking design. Factors manipulated were percentage of drifting items, type of drift, magnitude of drift, group achievement differences and choice of linking methods. The effect of item drift was studied by examining mean difference between true ès and è-estimates, and the accuracy of the classification of examinees into proper performance categories. Results indicated that the characteristics of the drift had little impact upon the performance of the Fixed Common Item Parameter (FCIP) method or Stocking & Lord's Test Characteristic Curve (TCC-ST) method and its influence on the performance of the Concurrent method varied depending on whether the drifting items were removed from the linking. In addition, better estimation was achieved if the drifting items were removed from the linking when the Concurrent method was used.
Read
