Robustness of item parameter estimation programs to assumptions of unidimensionality and normality

The effects of test dimensionality (one- or three-dimensional), theta distribution shape (normal, positively skewed, or platykurtic), and estimation program (BILOG, MULTILOG, or XCALIBRE) on the accuracy of item and person parameter estimates were assessed. The criterion was the root mean squared error of the difference between estimated and true parameter values. There was an interaction between program and dimensionality, indicating that the robustness of the unidimensionality assumption was a function of the estimation program. With the sample size and test length used, unidimensional estimation programs were insensitive to different shapes of the underlying theta distribution. BILOG consistently produced the smallest root mean squared error under most conditions. However, MULTILOG and XCALIBRE showed less variance in parameter estimation due to the violation of unidimensionality, with the exception of estimating the discrimination parameter in MULTILOG. Guidelines for estimating parameters of multidimensional test items using unidimensional item response theory models are suggested.