項目間分散を考慮した項目反応モデルの提案

―小論文評定への応用―

抄録

This study proposes a graded response model including a new parameter, which represents rater effects affecting inter-item variance. The validity of this model was shown by parameter restoration simulation, in which artificial item response data were generated and calibrated under 12 conditions each. Another simulation study, which compares estimation accuracy of θ in the presented model and original graded response model, was conducted. Results indicated the proposed model can remove the rater effect. Using a rating experiment that induced the halo effect, it was shown that the new parameter w_ir can be regarded as a halo effect.