On the impact of missing values on item fit and the model validness of the Rasch model

A crucial point regarding the development and calibration of an aptitude test is the presence of missing values. In most test administrations, examinees omit individual items even in high-stakes tests. The most common procedure for treating these missing values in data analysis is to score these responses as incorrect; however, an alternative would be to consider omitted responses as if they were not administered to the examinee in question. Previous research has found that both procedures for dealing with missing values result in bias in item and person parameter estimation. Regarding test construction, not only is there an interest in item parameter estimation, but also in global and item-specific model tests as well as goodness-of-fit indices. On the basis of such statistics, it will be decided which items constitute the final item pool of a test. The present study therefore investigates the influence of two different procedures for dealing with missing values on model and item-specific tests as well as item fit indices for the Rasch model. The impact of these different treatment alternatives is shown for an empirical example and, furthermore, for simulated data. Simulations reveal that the global model test, as well as the item test, is affected by the procedures used to deal with missing values. To summarize, the results indicate that scoring omitted items as incorrect leads to seriously biased results.