The problem of performing adequately sized randomized trials to demonstrate small survival benefits.

Two decades of clinical research have resulted in survival improvements for advanced non-small cell lung cancer. These results have been established by multiple randomized trials completed by meta-analyses such as by the review conducted by the Non-Small Cell Lung Cancer Collaborative Group [1]. Chemotherapy, combined with radiotherapy in locally advanced disease, has indeed been shown to reduce the risk of death by 10% (reduction corresponding to a 3% absolute survival benefit at 2 yrs). Compared to best supportive care alone, mainly in the metastatic stage of the disease, chemotherapy also improves survival with a reduction in mortality of ~25%, corresponding to an increase in the median survival of 1± 2 months and a 10% absolute benefit at 1 yr. To demonstrate such a small survival benefit, the researchers should ideally conduct large size randomized trials, reaching sufficient power to detect a realistic therapeutic improvement, that will provide reliable conclusions. One possibility is to conduct a trial with a fixed, a priori planned, sample size. This means, in the case of survival as the primary end-point, that the number of events required at the time of the analysis determining the study's power, will be calculated prior to the trial activation and specified in the protocol. This approach will lead to a convincing conclusion with a narrow confidence interval for the true treatment effect. The approach will, however, cause the clinical research group to be confronted by the difficulty of successfully recruiting the required number of patients in a reasonable time period, in order to maintain enthusiasm for the trial among the participating investigators and relevance for the scientific question addressed by the trial. An attractive alternative, for the researchers and for the patients, is to choose a design allowing for interim analyses with the hope of detecting an earlier difference, if there is one, between treatment arms and to stop the trial's accrual prematurely. However, multiple examinations of the data inflate the probability of falsely detecting a difference (the real p-value is greater than each nominal p-value calculated at each analysis) and may bias the treatment effect estimate. Stopping boundaries, like those proposed by POCOCK [2] (using a constant nominal level at each interim analysis) or by O'BRIEN and FLEMING [3] (using nominal p-values increasing with the amount of information already collected in the trial and therefore very conservative at the beginning of a trial) are constructed to maintain the overall …