|
Abstract:
|
The rater-item binary data commonly occur in many medical decisions. For example, a group of radiologists evaluate a set of mammograms and provide their diagnostic results about whether the patients have breast cancer or not. For such a type of data, additive models with different link functions are usually applied to estimate item and rater effects. However, in a real data analysis of a mammogram data (Beam et al., 1996) using the probit link shows a big discrepancy between the predicted ratings and the data. This inspires us to investigate the effect of a more flexible link. Skewed and nonparametric links (Kim et al., 2008; Chen et al., 1999; Mallick and Gelfand, 1994), are proposed and investigated for binary response data when covariates are available. However, their effect for rater-item binary data, where the covariates are not available, is lack of investigation. In this paper, skewed and nonparametric link functions are investigated for both additive and interactive models to accommodate different structures of this type of rater-item data, including balanced or unbalanced designs, small- or large-size experiments, skewed or non-skewed data structure.
|
