684 Abstracts

P44 CLOSED TESTING PROCEDURES FOR MULTIPLE END-POINT ANALYSIS

L. Hothorn German Cancer Research Center

Heidelberg, Germany

Multiple end-points may occur in clinical tdals. A powerful way of analysis consists in the use of the so- called multiple end-point analysis according to O'Brien [1]. This method enables the assertion of a difference between treatment and placebo group for the global vector of all end-points only. In some situations the effect of single end-points or interaction between selected end-points may also be of interest in the multivariate context. Based on the closed testing methodology according to Marcus et al. [2], a related modification of the multiple end-point analysis will be presented.

This procedure generates a large set of hypotheses (global, partial, elementary), which will be tested stepwise, each at the level alpha.

A PC-program which allows up to eight end-points will be demonstrated using real data sets.

1. O'Brien PC: Biometrics 40:1073-1087, 1984

2. Marcus R, et al.: Biometrika 63:655-660, 1976.

P45 GLOBAL TESTS OF EFFICACY FOR MULTIPLE END-POINTS WHEN NOT ALL ENDPOINTS APPLY

TO EACH STUDY PATIENT

Jose Ma. J. Alvlr Hillside Hospital

Glen Oaks, New York

Some diseases consist of several symptoms, not all of which need be present for the disease to be diagnosed. A global test which would provide an overall P value is needed to summarize the comparative efficacy of different therapies given multiple end-points in which the number of endpoints varies across the study sample. For situations in which all endpoints are available for all subjects, O'Bden and Shampo (1988) proposed a simple test which ranks patients on each separate end-point, sums ranks across patients, and compares these sums between therapies, using standard statistical procedures such as t-tests. Simple procedures can also be done when the number of end-points varies across subjects. T-tests or similar statistical tests can be performed on the proportion of improved over affected symptoms. Similarly, regression analyses can compare the number of improved end-point parameters between therapies, with the number of affected parameters used as a covariate. Problems with the use of ratio variables, power, and ease of interpretation of the proposed procedures will be discussed.

P46 APPLICATION OF REGRESSION TREE METHOD AS A TOOL TO ANALYZE A CLINICAL TRIAL

K. Ulna and C, Schmoor Institute for Medical Statistics and Epidemiology

Technical University Munich, Germany

In the analysis of a clinical trial comparing different treatments the problem arises that the patients may be rather heterogeneous with regard to their natural prognosis. Therefore, a simple overall comparison of treatment groups can lead to a biased estimate of the treatment effect even in a well-balanced randomized study, at least when the survival time of the patients is the criterion. Adequate analysis of the treatment effect is only feasible in a multivariate framework where the important prognostic factors are accounted for and, additionally, treatment-covadate interactions may be evaluated.

As an alternative to the traditional Cox-model in a more explorative way the Classification and Regression Tree (CART) technique can be applied. Using CART homogeneous subgroups of patients can be identified. Within these subgroups a treatment effect can be investigated. CART and Cox-model have been applied to the data of randomized clinical tdal on patients with brain tumor comparing two chemotherapies.

Within this tdal overall there was no treatment difference. In beth methods however subgroups of patients could be worked out who get benefit from one or the other treatment.

The differences between these approaches are outlined and discussed.