Abstract: There were two main features of this paper: 1) the likelihood based p e rson fit index l z and the residual based caution index, Tatsuoka's ζ ha ve similar mathematical structure except the difference of weight. The wei ght of l z index is a logit of an item response function (IRF), while the weigh t of ζ is residual IRF and true score; 2) a degree of coincidence between any t wo aberrant subjects is defined as a similarity for multidimensional scaling (MD S):S j?k=∑gr=1p j?rp k?r+(1 p j?r)(1 p k?r) 1/2, where g is the number of item groups, p j?r is the proportion of subject j who answer correctl y in the item group r. The larger value of the S j?k, the more similar ity will be between the subject j and k. A dataset of Chinese vocabulary test wa s investigated using the l z index and MDS. The 3.5% of the abe rrant subjects were detected, and classified into 5 types. The result of MDS indicated that it is a useful and effective method for classification of aberra nt item response patterns. Suggestions for future research are provided.

Key words: IRT, person fit index, similarity, multidimensional scaling