Comparison of Statistical Classification Methods Based on a Prostate Cancer Study

Abstract

Prostate cancer is one of the most common cancers in American men. Management of prostate cancer depends on its stage, because only cancers that are confined to the organ of origin are potentially curable by radical prostatectomy. In this article we have considered different statistical methods to predict the probabilities of non­organ confined prostate cancer based on its clinical stage. Modern computer intensive methods such as bagging, neural networks and support vector machines are compared to more classical methods such as linear, quadratic and logistic discrimination and less computer intensive nonparametric methods such as smoothing splines and classification trees . All these methods are allied to a dataset from a recent prostate cancer study. We have presented sensitivity, specificity, positive predictive value, negative predictive value, and overall accuracy for each of the methods. The study shows linear discriminant analysis and support vector machine perform better than other methods for this data set, at least from a predictive view point .