| Abstract: | Optimal classification formulation is adapted to the context of discrimination when the response is ordinal. The resulting method, optimal discriminant analysis for ordinal responses (ODAO), is presented and compared with two reference discrimination techniques used in this context (proportional-odds ordinal logistic regression and normal discrimination) using a study of prognosis following burn injuries and simulated data. The ODAO method clearly outperforms both reference methods in terms of classification accuracy (in training and validation samples), robustness to outliers, simplicity of use and applicability in clinical settings. ODAO is a promising method for improving classification performance in discrimination with ordinal responses and merits further investigation. © 1997 by John Wiley & Sons, Ltd. |
