Background. Traditional statistical analysis of 2 surgeons' experiences with resectable malignant melanoma during a 30-year period (November 1970-July 2000) was compared with new tree-structured recursive partitioning regression analysis. Methods. A total of 1018 consecutive patients were registered and 983 patients were evaluable. Disease-free survival (DFS) and melanoma survival (MS) were calculated by Kaplan-Meier method for stage, thickness, ulceration, site, lymph node involvement, age, sex, and type; and compared with log-rank tests. Cox proportional hazards model was used for multivariate analysis. Multivariate predictors were used to analyze DFS and MS with a classification and regression tree model that partitioned patients into progressively more homogenous prognostic groups with significantly different Kaplan-Meier curves. Results. Multivariate correlations were with thickness (millimeters), ulceration, age (per year), type, and sex in predicting DFS (relative risk = 1.18, 2.10, 1.05, 1.71, and 1.71, respectively). Thickness, ulceration, age, and type remained significant predictors of MS (relative risk = 1.14, 3.02, 1.02, and 2.30, respectively). Classification and regression tree analysis showed thickness, age, ulceration, and sex affected DFS. Only thickness and ulceration were significant in predicting MS. Conclusion. The Cox model is an important tool for analysis of clinical data but has flaws. New statistical technology to predict outcome should be considered. Classification and regression tree analysis of larger published series may reveal new predictors useful for staging, prognosis, and guiding clinical decisions.
ASJC Scopus subject areas