Sequential statistical methods applied to a randomised controlled study with survival data: the MRC superficial bladder cancer trial of Mytomicin-C.
by
Nora Donaldson
King's College London, Department of Oral Health Reseach
Coauthors: Andres Michael Donaldson, King's College London Dental Institute, UK, Juan G Gonzalez, King's College London Dental Institute, UK, Mahesh K.B. Parmar MRC Cancer Trials Office, Cambridge, UK. Mahesh K.B. Parmar MRC Cancer Trials Office, Cambridge, UK.

Background

Reduction in the duration of a clinical trial and in the number of patients required, compared to a conventional design can be obtained when we adopt a sequential design. In this paper we use data from a trial completed by the British Medical Research Council on the effects of Mitomycin-C on newly diagnosed superficial bladder cancer to examine the savings in sample size and to highlight some potential problems when the assumptions of the model are not satisfied. These problems are not exclusive to the sequential design, but are also present when a predetermined fixed sample size is used.

Methods

The stopping rule adopted was a double triangular test with 90% power to detect a difference in the 2-year survival rate of 0.20, both in favour of, or against, mitomycin-c. This symmetric requirement allowed us to demonstrate inferiority if it existed; to discourage use of a widely available and popular therapy.

Results

The first interim look was performed 15 months after the first recruitment. Patients with no recurrence at the time of analysis were considered censored at that time. The most important variable in predicting recurrence time was the maximum diameter of the largest tumour, fitted with two levels: 0-1 cm and  1 cm (p=0.03). Grade of tumour assessed by the reference pathologist, fitted with three levels: 1-2, 3-4 and 5, (p=0.18); and number of tumours, fitted with two levels: 0-2 and 2, (p=0.20), were also found to be of some prognostic importance. Consequently, the first and all subsequent interim analyses were stratified for these prognostic factors, forming twelve strata in all.

The sequential version of the trial was stopped at the eighth interim analysis, 27 months after the first recruitment. There were 305 patients recruited and 121 recurrences accumulated. The null hypothesis was rejected, suggesting there is a significant advantage of Mitomycin-C on the recurrence rate of bladder cancer. The p–value at the time of stopping, allowing for interim looks, was p=0.04. The 95% confidence interval for the log hazard ratio was (0.02, 0.81), with median unbiased estimate 0.42. All values were adjusted for previous interim looks.

Conclusions

These results were consistent with a final fixed sample analysis at 72 months. In the sequential re-analysis recruitment was stopped 27 months from entry of the first patient, and an analysis with the intended power was available. This effectively reduced the duration of the trial by 45 months. The number of patients recruited was reduced by 17 and number of recurrences accumulated at the time of analysis by 52.

Stratifying by important prognostic factors had the advantage over covariate adjustment in a Cox model of not assuming proportional hazards between strata and is seen to provide one way of accounting for departures from a proportional hazards model. The use of survival data in grouped form is another.

Date received: February 15, 2007