The Power Of Polytomous Person-fit Statistics With Grm To Detect Typical Bias Behaviors In Psychological Test

Purpose:In Psychological Test aberrant response behavior detection and control has received a lot of concern by psychometricians and practitioners. Person-Fit Statistics (PFS) are statistical index developed by researchers to detect bias responses in psychological tests. But so far all researchers only have focused on PFS which are based on dichotomous Item Response Theory (IRT) models. We still don't know whether the calculation formulas of the PFS can be extended to polytomous Item Response Theory (IRT) models and how much the statistical powers are.In this study we first extended the formulas of dichotomous-item-based PFS to polytomous items based on the Graded Response Model(GRM), and then throw Monte Carlo procedure we generated normal item-score patterns and aberrant item-score patterns. Type I andⅡerrors ratios are calculated under different conditions.Method: Three IRT-based dichotomous PFSs are involved in this study, they are ECI4,LnU and LnW, which performed well in previous researches, whose formulas are extended to polytomous items. And three kinds of aberrant response patterns are generated by Monte Carlo procedure according to their meaning in psychological tests. These aberrant response patterns are social desirable responding (e.g. spuriously high response), faking good (e.g. spuriously low response) and indifference (e.g. always choosing middle answer).Other two conditions are also designed. One is proportion of biased items (e.g. proportion of item scores that are converted from normal to aberrant score pattern). This condition contains three levels, e.g. 10%, 20% and 40%. The other is test length (e.g. the whole test contains 30, 50 or 80 items). Values of the three PFS index mentioned above are calculated using real abilityθand estimated ability ^θ. Examinees'real abilityθare generated by normal radom function of SAS ---- rannor(0), and their estimated ability ^θare estimated by the IRT software MULTILOG7.0 using the item score pattern that contain 10%, 20% and 40% aberrant items.SAS 11.0 is used to conduct the data simulation. For each study condition 25 samples are generated, each sample contains 1000 examinees. Result: Findings indicate that the three polytomous PFSs with GRM can in some extent discern examinees who have aberrant item scores from the examinees who don't. Index ECI4 can detect all the three kinds of bias rather correctly, index LnU and LnW showed higher power to detect spuriously low and spuriously high response, their power decrease when detecting indifference response bias.The detection power of three polytomous PFSs based on GRM all increase when test length become longer, and their power all decrease when the percentage of bias items elevated. The detection power become slightly lower in estimated ability ^θcondition than in real abilityθcondition, but the difference is not large.Conclusion: The three polytomous PFSs based on GRM presented in this study exhibit similar quality as dichotomous PFS studied by previous researchers, indicate that their formula deductions are successful. These PFSs perform nearly well under estimated ability ^θcondition than under real abilityθcondition, indicate that they can be used in real psychological tests. These conclusions are tentative, further researches are needed to confirm them.