Research On Item Response Process And Data Analysis Methods Of Personality Tests

The Life Orientation Test(LOT) and its revision(LOT-R) were selected for exploring the item response process and data analysis of personality test. LOT and LOT-R were representative of personality tests, which consisted of a few balanced positive and negative wording items. The dimensionality of LOT and LOT-R was debated among most western researchers. Most of researchers found that when factor analysis method was used to explore the dimensionality of LOT and LOT-R, unidimensionality of these two tests were often violated. So some researchers favored that LOT and LOT-R should include two dimensions, but others believed that the violation of unidimensionality was only the methodological effects of item wording. The author of this research confirmed the dimensionality of these two tests in the culture of Chinese by factor analysis and ideal point methods, and found that when factor analysis was used, two factors were retained, which was consistent with the former researches, alternatively unidimensionality was favored under ideal point methods. As a result the author concluded that there was only one dimension underlying LOT-R, and the item response of this test conformed to ideal point process.Since the item responses of LOT-R conformed to ideal point process, the author supposed that unfolding models can fit the response data of LOT-R better than dominant models. After comparing generalized graded unfolding model(GGUM) to graded response model(GRM), and to generalized partial credit model(GPCM), evidences were provided to indicated that GGUM fit to the data of LOT-R better indeed than GRM and GPCM. When five categories were recoded to two ones, model-data fits of GGUM were improved, and when GGUM was extended to generalized partial unfolding model(GPUM), which was more complex than the former, model-data fits were further improved too. Together with all results, the author concluded that more complex models can provide better fit to personality tests.After investigating model-data fits, the methods of estimating parameters of unfolding model were also examined. The results indicated that when marginal maximum likelihood(MML) method was used, the item parameter and standard errors estimates of GGUM were incorrect. After recoding response categories and throwing some constraint to threshold parameters, estimates of parameters were improved. But estimates of item locations and standard error were overestimated. Markov Chain Monte Carlo(MCMC) method in place of MML method was used to estimate parameters in GGUM, and more accurate estimates of item parameters were obtained. MCMC can also used to more complex model(for example, GPUM).A program was developed to test model-data fit by the author and his collaborator in this dissertation. The results obtained by this program were almost consistent with those by MODFIT program. Furthermore, this program did not restrict sample size and item numbers, and can be used to test model-data fit of new model.But this program can not be used to computer adjustedχ~2/df ratios.