Detecting Item Misfit in Rasch Models

Psychometrics have long relied on rule-of-thumb critical values for goodness of fit metrics. With powerful personal computers it is both feasible and desirable to use simulation methods to determine appropriate cutoff values. This paper evaluates the use of an R package for Rasch psychometrics that has implemented functions to simplify the process of determining simulation-based cutoff values. Through six simulation studies, comparisons are made between information-weighted conditional item fit (“infit”) and item-restscore correlations using Goodman and Kruskal’s γ. Results indicate the limitations of small samples (n < 500) in correctly detecting item misfit, especially when a larger proportion of items are misfit and/or when misfit items are off-target. Infit with simulation-based cutoffs outperforms item-restscore with sample sizes below 500. Both methods result in problematic rates of false positives with large samples (n >= 1000). Large datasets should be analyzed using nonparametric bootstrap of subsamples with item-restscore to reduce the risk of type-1 errors. Finally, the importance of an iterative analysis process is emphasized, since a situation where several items show underfit will cause other items to show overfit. Underfit items should be removed one at a time, and a re-analysis conducted for each step to avoid erroneously eliminating items.