Robust Estimation of Latent Ability in Item Response Models
Once responses involving aberrant behaviors, for instance cheating, guessing, or carelessness, we generally would eliminate aberrant responses which here is called as response disturbances from our computation of estimation, then follow person fit procedure to detect those who with aberrant responses. However, the present paper abandoned above method but strove for more robust estimation methods by adopting concept of weighting to adjust contribution of pure (free) and aberrant responses. Thus, the person estimates would obtain much more information from free responses more than aberrant ones. I think these two approaches aforementioned should be adopted on different situations, and that first one can be used in high stake assessments to avoid aberrant disturbances of responses, and more important we should take action for those who with cheating. In terms of low stake assessments, the methods proposed in the paper can be considered due to the fact that no one would argue that unfair but we also can get more accurate estimates. Back to the simulations of the present study, limited and surreal conditions were conducted, however it did offered exploratory investigation on robust estimation of simple context. It was found that new methods of Bisquare and Huber could have better performance on mean of person estimate with slight to moderate loss on RMSE for aberrant conditions. For null condition, all three methods have equally well performance. Future studies could consider purification procedure under existing study, and these two methods also can be applied to improve detect aberrant responses on fit studies. Moreover, this study is based on item parameters known, maybe we can consider condition of item parameters unknown because item and person estimates would probably be affected by response disturbances simultaneously in reality.