Maximum entropy estimation vs. multivariate logistic regression: which should be used for the analysis of small binary outcome data sets? [Breast cancer prognosis]

The principle of maximum entropy has been applied to problems with incomplete data but with well-defined hypothesis space. Applications include the spectral algorithm of Burg and the algorithm for image reconstruction of Gull and Skilling. In this paper, we explore the use of the entropy maximization network (EMN) in constructing multinomial distributions from small data sets for carrying out plausible reasoning. The EMN proves to be a better predictor than multivariate logistic regression for small binary outcome data and has consistent performance accuracy. Differences in performance are evaluated by comparing the areas under the receiver operating characteristic curve, A_x