Caveats for modeling disease free survival after radical prostatectomy.

I t has been more than a decade since Patrick Walsh described the “anatomic” radical prostatectomy for the treatment of localized prostate carcinoma. Since then, thousands of men with localized prostate carcinoma have been treated and we have come to recognize the tremendous heterogeneity in disease free survival. Early attempts at prognostication after surgery utilized the Whitmore A-D staging system, which has since undergone a number of revisions to become the American Joint Committee on Cancer/International Union Against Cancer TNM staging system. However, the evolution of staging classifications takes years to occur and is based primarily on existing models that do not account adequately for the myriad of other prognostic indicators such as prostate specific antigen (PSA) and tumor specific pathologic markers. More recently, investigators have incorporated staging into multivariate models. These models have elaborated on our understanding of the biology of prostate carcinoma by identifying the more important prognostic elements while simultaneously controlling for confounding variables. The Cox proportional hazards model specifically allows for the prediction of risk according to predetermined variables; however, the growing number of potential predictor variables makes the use of these traditional statistical methods cumbersome. To simplify the clinical application of these models, investigators often will construct classifying tables or nomograms. A major criticism of classifying modeled outcomes has been the arbitrary assignment of risk levels or cutoff values. In this sense, existing modeling techniques for disease free survival after radical prostatectomy have not been very satisfying. The article by Banerjee et al. in the current issue of Cancer addresses this issue by applying a nonparametric modeling technique known generally as recursive partitioning. This modeling technique requires a number of steps including 1) selection of the splits at each node, 2) determination of when to stop splitting, and 3) pruning the tree down to the right size. As a result, the models are only as good as the choices made by the investigator at each of these steps. In contrast to linear regression techniques, which require a predetermined linear model be fitted to the data with the assignment of various weights to each predictor variable, recursive partitioning is less restrictive and permits the data to dictate the structure of the model. 232