The following definitions of classification consistency and classification accuracy are quoted from Lee (2010).
Classification consistency is the degree to which examinees would be classified into the same performance categories over parallel replications of the same assessment. Classification accuracy is the degree to which observed classification would agree with ‘true’ classifications assuming known cut-scores on a single assessment.
This paper tried to evaluate two approaches for classification accuracy, one is Lee's that used the total score to calculate the probability for classification, and the other is Rudner approach that used latent trait estimates. By manipulating IRT model (1PL, 2PL, 3PL and GRM), sample size, test length and cut score location, it was found that both the approaches performed well on the conditions for empirical CAs. Moreover, although the Rudner approach is general good for most of the conditions, while the Lee approach good at the conditions of 1PL model.
The comparison between Lee and Rudner is somewhat strange, cause both approaches are involving item response models and the total score and latent trait estimates are derived from IRT models. I guess the difference between the two approaches is whether normal distribution of ability parameters assumed. In order to compare these two approaches, the distribution of ability parameters should be manipulated.
Since the TRUE is always unknown(if not impossible) in practices, the consequence of using the wrong model (rather than the presence of misfit items only) should be investigated.