Method of successive dichotomizations: An improved method for estimating measures of latent variables from rating...

The most commonly used models for estimating measures of latent variables from polytomous rating scale data are the Andrich rating scale model and the Samejima graded response model. The Andrich model has the undesirable property of estimating disordered rating category thresholds, and users of the model are advised to manipulate data to force thresholds to come out ordered. The Samejima model estimates ordered thresholds, but has the undesirable property of estimating person measures on a non-invariant scale-the scale depends on which items a person rates and makes comparisons across people difficult. We derive the rating scale model logically implied by the generally agreed upon definition of rating scale-a real line partitioned by ordered thresholds into ordered intervals called rating categories-and show that it estimates ordered thresholds as well as person and item measures on an invariant scale. The derived model turns out to be a special case of the Samejima model, but with no item discrimination parameter and with common thresholds across items. All parameters in our model are estimated using a fast and efficient method called the Method of Successive Dichotomizations, which applies the dichotomous Rasch model as many times as there are thresholds and demonstrates that the derived model is a polytomous Rasch model that estimates ordered thresholds. We tested both the Method of Successive Dichotomizations and the Andrich model against simulated rating scale data and found that the estimated parameters of our model were nearly perfectly correlated with the true values, while estimated thresholds of the Andrich model became negatively correlated with the true values as the number of rating categories increased. Our method also estimates parameters on a scale that remains invariant to the number of rating categories, in contrast to the Andrich model.