A modified graded response model to account for guessing in multiple-choice items

This presentation focuses on multiple-choice items without any correction or control for guessing. Most often, item responses are recoded as binary outcomes so that dichotomous logistic item response models (such as the 2PL or 3PL models) can be used to estimate ability levels. This is however at the cost of loosing information from collapsing all distractors into a single FALSE category. Polytomous IRT models could be considered to overcome that drawback but none of the classical models are suitable to incorporate guessing into account. The purpose of this talk is to introduce a modification to the graded response model (Samejima, 1969) that can take guessing into account. The modification consists in introducing lower and upper asymptotes to the cumulative probability curves. After a brief theoretical and technical description, practical implementation using the R package mirt (Chalmers, 2012) is discussed. Primary results from a simulation study are eventually presented and compared to traditional IRT scoring methods for scoring multiple-choice items.