Does the Rasch Model Really Work? A Discussion for Practitioners. ERIC/TM Report 67

Issues of concern to test developers interested in applying the Rasch model are discussed. The current state of the art, recommendations for use of the model, further needs, and controversies are described for the three stages of test construction: (1) definition of the content of the test and item writing; (2) item analysis; and (3) test calibration. The Rasch model is defined as specifying that the probability of a person's correctly answering a test item is a function of two parameters: the person's ability and the item's difficulty. Calibration refers to the process of estimating the two sets of parameters, and evaluating the fit of these data to the model. The Rasch model has been found to work for achievement tests, both criterion- and norm-referenced, as well as other types of tests. The model helps to identify poor test items, but does not help to write better items. Mean square fit statistics are recommended for item analysis. Questions are raised regarding unidimensionality, the robustness of the model, sample size issues, and the definitions of sample-free and item-free. The authors conclude that test developers can feel comfortable, albeit cautious, in using the Rasch model. (Author/GDC)