The objective of this research is to utilize measures of distance between observed item response vectors and ideal response patterns that can provide effective nonparametric classification, under a wide variety of possibilities for the underlying cognitive diagnosis model responsible for generating
the data. A simulation study is used to compare nonparametric classification to maximum likelihood estimates derived under the true model as well as misspecified models. A real data analysis of fraction subtraction data is presented to show how nearly model-based and nonparametric classification agree.
For simulation study, the pattern-wise agreement rate (PAR) and Attribute-wise agreement rate (AAR) are used for judgment. The results show that for DINA model, the nonparametric technique
appears to be quite competitive in comparison with MLE; for NIDA model, the nonparametric
method appears generally less tolerant of larger slipping and guessing parameter settings and depend
on the size of k.
in equation 4, the author said the weighs can be specific to items, what about the specific value, since the value will affect the results?
How can we sure the slipping or guessing parameter is under 0.5 in real situation?