강한 베이지안 사전분포를 사용한 가교문항-고정 IRT 추정 방법의 원리와 기능

초록

When test data are obtained from the nonequivalent-groups anchor test design and are analyzed by using item response theory (IRT), fixed-anchor IRT calibration is used for estimating the parameters of the non-anchor items, while estimating the underlying ability distribution, on the same scale of the anchor items' parameters to be fixed. To successfully conduct fixed-anchor IRT calibration, previous research has used a specialized marginal maximum likelihood (MML)-EM algorithm method in which the anchor items' parameters are explicitly excluded from the estimation. The present study presented an alternative to the specialized MML-EM method for fixed-anchor IRT calibration, which is a generalized Bayesian modal (BM)-EM approach that allows one to impose very strong Bayesian priors for the anchor items' parameters to be fixed, and examined through computer simulations how the strong-priors based BM-EM method might perform depending on the degree of strength of the priors. With the strong-priors based BM-EM method, one can conduct fixed IRT calibration for an anchor item's parameter by setting the mode of the chosen prior to the value of the parameter and setting the standard deviation (SD) of the prior to very small values such as 0.000001~0.01. The results from simulations involving various testing conditions showed that the strong-priors based BM-EM method performed successfully even when the values of SD of priors were assigned up to 0.01.

키워드