A Study Of Computerized Adaptive Personality Test Based On Unfolding Response Mechanism

According to the individual reaction mechanism for the item, the IRT model can be divided into two categories: cumulative model and unfolding model, many studies have shown that the latter is more suitable for personality measurement, existing unfolding models are unidimensional model, the most representative model is GGUM. Nowadays, unidimensionality of unfolding data is assessed by basis of the ratio of the first to the second factor’s eigenvalue which meets Reckase’s criterion(more than 5) or Hambleton’s criterion(more than 3), or predecessors’(e.g., Davsion,1977;Andrich & Styles,1998) conclusions(two factors and figure of items arraying by loadings being a horseshoe). The former is directly cited from the field of ability, the latter is based on certain condition, it is unclear whether the two methods are feasible in different type of unfolding data. In this paper, the first study used factor analysis using component analysis(PCA) to six different type of unfolding data designed according to item location parameters’ distribution and subjects’ trait distribution. The study found that:(1) Precursors(e.g., van Schuur & Kiers,1994) considered that unfolding data didn’t satisfy the condition of linear factor analysis, the first study found this view is wrong according to the value of KMO and Bartlett sphere test of six types of unfolding data.(2) Ratio of the eigenvalue of first to the second was less than 3 when the hemogeneity of sample was strong and items’ locative parameters distributed uniformly, so unidimensionality of unfolding data couldn’t be assessed by the criterion of Reckase or Hambleton; predecessors(e.g.,Davsion,1977;Andrich & Styles,1998)thought that the perfect unfolding data had two factors, but the first study found unfolding data had three factors if the heterogeneity of sample was strong, the first study concluded the unidimensionality of unfolding data can’t be assessed by the number of factors.(3) The factor analysis using PCA analyzed a likert scale’s data conforming to the unfolding mechanism, the shape of items’ loading in two factors wasn’t a horseshoe, but a pair of parentheses. Extrem items mainly gathered at the upper end of parentheses, items closing to the middle position gathered at the bottomof parentheses, items’ position differed from horseshoe.(4) The ratio of the first to the second factor’s eigenvalue increased with the heterogeneity of sample becoming stronger, under the circumstance of item locating.(5) It is impossible that unidimentionality of unfolding data is assessed by factor analysis directly, it is advised that model-data fit should be made firstly, factor analysis can be made on the data of those items having good fit, if the results of factor analysis are suitable for the conclusions of the first study, which are choose according to distributions of items’ locative parameters and sample of subjects, this can be a proof of unidimentionality of data.CAT is the successful application of IRT, CAT based on the cumulative model has been used in many ability examinations, and there are many studies of relative technologies of CAT, however, no study of technologies of CAT based on the unfolding model appears. The second study made a simulating study about methods of trait estimation and strategies of item selection in the CAT based on GGUM, the second found:(1) If the test information was targeted for 25, extreme subjects need more items than middle subjects, the most efficient design of CAT was EAP-MFI, every subject needs six items on average, extreme subjects need about eight items.(2) For extreme subjects, CAT using EAP-KLI design was worst than other five, but in the period of middle subjects on the standard error of measurement and recovery of traits, but for the middle subjects, all design were almost same.(3) CAT using EAP-KLI was worst in aspects of mearsuring standard error and recovery of traits during the CAT with variable length; CAT using EAP-MFI was best in aspects of mearsuring standard error and recovery of traits during the CAT with fixed length.(4) It was advised that people should use EAP as trait estimating method and MFI as chosing item strategy when developing CAT based GGUMSo far, there is no real CAT based on unfolding model, To make up the blank of this field, the author developed UIE-CAT, the study through of real test found:(1) The undergraduates’ introversion-extroversion item bank based on GGUM could keep more neutral items, which helped improve measuring accuracy for subjects locating in the middle of traits.(2) The contrary items of likert scale method must score reversely, but contrary items based on GGUM had negative locative parameters when the IRT methodwas used to estimate the trait, items needn’t score reversely.(3) Validity related criterion showed that UIE-CAT was more effective than E scale of EPQ, Validity related criterion of CAT using different terminating rules testified efficiency of CAT.(4) The four-week retest reliability coefficient of UIE-CAT was 0.886 if the number of used items of CAT was targeted for 10, other UIE-CAT designs’ coefficients with different termination rules were more than 0.9, the measuring stability of UIE-CAT was perfect.