Ability Estimation and Item Calibration Using the One and Three Parameter Logistic Models: A Comparative Study.

摘要

The literature on latent trait calibration procedures was reviewed to determine the methods available to calibrate dichotomous items for tailored testing applications. From the procedures, the most promising techniques for the calibration of items using the one- and three-parameter logistic models were selected for comparison. The maximum likelihood procedure developed by Wright and Panchapakesan was selected for the one-parameter model; and the estimation procedure for use with omitted responses, developed by Wood, Wingersky and Lord, was selected for the three-parameter model. The two procedures were then compared on their suitability for use with multivariate item pools, the sample size required for calibration, the effects of item quality, and the cost of calibration. Sixteen data-sets were used for these evaluations; eight live testing data-sets, and eight simulation data-sets generated to match specified factor structures. The one- and three-parameter models were found to estimate different components when the tests measured several independent factors. The three-parameter model estimated parameters from one of the factors, ignoring the others, while the one-parameter model estimated the sum of the factors.