Accuracy of dichotomous logistic IRT models to estimate ability with multiple choice items

This presentation focuse son multiple-choice items without any control or correction for guessing. Most often these items are recoded as binary outcomes (TRUE-FALSE) and logistic IRT models are calibrated using these recoded data. The main purpose of this study is to highlight the limits of this approach by (a) proposing a simulation model to allow for guessing at various levels of ability, and (b) comparing the dichotomous approach to the use of nominal response model. Preliminary simulations indicate that none of the models are very accurate to estimate latent abilities accurately, though the nominal response model has fewer bias then the dichotomous models.