Research On Online Calibration Method Based On Graded Response Model Under Random Design

Online calibration,which carried by computerized adaptive testing(CAT),is a technology-enhanced architecture for item calibration.Computerized adaptive testing,with the flexibility of testing time and variety of items,is expected to get the same accuracy of paper and pencil test as long as half the length test and so on,so that it can be applied in many large evaluation pro jects.At present,the item bank of computerized adaptive test is facing the problems of high construction cost,updating and complicated expansion technology.The maintenance and management of item bank is essential for ensure the validity of CAT.Moreove r online calibrations of new items under dichotomously scored models have achieved good results,but under polytomously scored model is reported rarely.Hence it is necessary to explore online calibration of polytomously scored items,because polytomously scored items are widely used in practice more than dichotomously scored items due to it can provide more information.Therefore,it is helpful to enrich the structure of the question bank by exploring the online calibration of the multilevel score project.Online calibration usually adopts the iterative method,and the selection of the initial value of the iteration is very important.To explore the performance of online calibration for polytomously scored items,a method to calculate the initial values of the multiple EM cycle method(MEM)is proposed based on graded response model(GRM),which is focus on the extended squeezing average method and the multiple-sequence correlation coefficient method to calculate as the initial parameters of the new item,then use the multiple EM cycle method to estimate parameters.Results of Monte Carlo simulation show that the parameter estimation of new items can get acceptable estimation accuracy and the parameters of new items are more accurate with a small increase of the sample size for calibration.When the sample size for calibration reaches 500 per item,the estimation accuracy of new items is better,this sample size is relatively small compare with the MMLE / EM method.Hence it is enough to justify online calibration of the new items under GRM model can maintain the advantage of parameter estimation under dichotomously scored model.