Ordinal Classification in Medical Prognosis

Summary

Objectives: Medical prognosis is commonly expressed in terms of ordered outcome categories. This paper provides simple statistical procedures to judge whether the predictor variables reflect this natural ordering.

Methods: The concept of stochastic ordering in logistic regression and discrimination models is applied to naturally ordered outcome scales in medical prognosis.

Results: The ordering stage is assessed by a data-generated choice between ordered, partially ordered, and unordered models. The ordinal structure of the outcome is particularly taken into consideration in the construction of allocation rules and in the assessment of their performance. The specialized models are compared to the unordered model with respect to the classification efficiency in a clinical prognostic study.

Conclusions: It is concluded that our approach offers more flexibility than the widely used cumulative-odds model and more stability than the multinomial logistic model. The procedure described in this paper is strongly recommended for practical applications to support medical decision-making.

Keywords

Bayes allocation - decision-making - medical prognosis - measures of ordinal separation - natural ordering

• References

• 1 Anderson JA. Regression and ordered categorical variables (with discussion). Journal of the Royal Statistical Society 1984; B; 45: 1-30.
• 2 Feldmann U, Steudel I. Methods of ordinal classification applied to medical scoring systems. Stat Med 2000; 19: 575-86.
• 3 Anderson JA. Separate sample logistic discrimination. Biometrika 1972; 59: 19-35.
• 4 Anderson JA. Logistic discrimination. Handbook of Statistics, North-Holland 1982; 2: 169-91.
• 5 Green MS. Evaluating the discriminatory power of a multiple logistic regression model. Stat Med 1988; 7: 519-24.
• 6 Panel on Discriminant Analysis, Classification and Clustering.. Discriminant analysis and clustering. Statistical Sciences 1989; 4: 34-69.
• 7 McCullagh P. Regression models for ordinal data (with discussion). Journal of the Royal Statistical Society 1980; B; 42: 109-42.
• 8 Agresti A. Categorical Data Analysis. New York: Wiley; 1990
• 9 Goodman LA. The analysis of dependence in cross-classifications having ordered categories, using log-linear models for frequencies and loglinear models for odds. Biometrics 1983; 39: 149-60.
• 10 Feldmann U, Osswald PM, Hartung HJ, Lutz H. Computer aided methods to predict perioperative risks. Computers in Critical Care and Pulmonary Medicine. Springer. 1985: 162-83.
• 11 König J. CANLOG: An IML-procedure in SAS to solve the canonical logistic model. 2001 http://www.uniklinik-saarland.de/imbei/canlog
• 12 Efron B. How biased is the apparent error rate of a prediction rule?. Journal of the American Statistical Association 1986; 81: 461-70.
• 13 Anderson JA, Philips PR. Regression, discrimination and measured models for ordered categorical variables. Applied Statistics 1981; 30: 22-31.
• 14 Agresti A. A survey of models for repeated ordered categorical response data. Stat Med 1989; 8: 1209-24.
• 15 Holtbrügge W, Schumacher M. A comparison of regression models for the analysis of ordered categorical data. Applied Statistics 1991; 40: 249-59.
• 16 Greenland S. Alternative models for ordinal logistic regression. Stat Med 1994; 13: 1665-77.
• 17 Campbell MK, Donner A, Webster KM. Are ordinal models useful for classification?. Stat Med 1991; 10: 383-94.
• 18 Bell R. Are ordinal models useful for classification? Comment. Stat Med 1992; 11: 133-4.
• 19 Rudolfer SM, Watson PC, Lesaffre E. Are ordinal models useful for classification? A revised analysis. Journal of Statistical Computation and Simulation 1995; 52: 105-32.