Though calibration of an achievement test within a psychological and educational context is very often carried out by the Rasch model, data sampling is hardly designed according to statistical foundations. However, Kubinger, Rasch, and Yanagida (2009, 2011) suggested an approach for the determination of sample size according to a given Type-I and Type-II risk and a certain effect of model contradiction when testing the Rasch model. The approach uses a three-way analysis of variance design with mixed classification. For the while, their simulation studies deal with complete data, meaning every examinee is administered with all of the items of an item pool. The simulation study now presented in this paper deals with the practical relevant case, in particular for large-scale assessments, that item presentation happens to use several test-booklets. As a consequence, there are missing values by design. Therefore, the question to be considered is, whether this approach works in this case as well. Besides the fact, that data are not normally distributed but there is a dichotomous variable (an examinee either solves an item or fails to solve it), only a single entry for each cell exists in the given three-way analysis of variance design, if at all, due to missing values. Hence, the obligatory test-statistic's distribution may not be retained, in contrast to the case of having no missing values. The result of our simulation study, despite applying only to a very special scenario, is that this approach works, indeed: Whether test-booklets were used or every examinee is administered all of the items changes nothing in respect to the actual Type-I risk or to the power of the test, given almost the same amount of information of examinees per item. However, as the results are limited to a special scenario, we currently recommend any interested researcher to simulate the appropriate one in advance by him/herself.
|Number of pages||11|
|Journal||Journal of Applied Measurement|
|Publication status||Published - Dec 2015|