The aim of this study is trying to find out a set of items that conform to the assumption of unidimension in Rasch model. The author introduced a procedure combined with a statistic of model fit, . Firstly, all the possible combinations of items were classified. The p value of a statistic of model fit was calculated for each set of items. Then find out which set is with the maximum p value. Secondly, the chosen set was added one item from items to construct a new set. The calculating of p value was performed again to choose next new one set and terminated until predefined upper threshold or cannot be expanded. statistic is somewhat a index of comparing the expected and observed frequencies of persons giving positive or negative response to item i. However, the index requires large sample size to be distribution asymptotically.
The procedure are compared with PCA in simulation study. As expected, the percentage of correct classification was high when sample size increased for both methods. However, the new method gave no striking results compared with PCA. Exception was occurred when sample size is quite small with large number of items. For sample size 500, the performance is better when number of items is large and larger standard deviation of person and item.
Qs:
1. The article is bad written and hard to understand, especially in his core of procedure.
2. The results with the new method are not evident. If we have small sample, how can the procedure guarantee a better results of classification than other methods?
3. It sounds strange that the medium correlation of persons would lead to a worse results than no correlation. No clear explanation was given.