Bayesian interim analysis of randomised trials

Abstract

Brophy and Joseph1Brophy JM Joseph L Bayesian interim statistical analysis of randomised trials.Lancet. 1997; 349: 1166-1168Summary Full Text Full Text PDF PubMed Scopus (9) Google Scholar describe one form of Bayesian interim analysis, in which information from other parallel or completed studies is combined with results of an ongoing trial. Such analysis is contentious, as was rightly pointed out in Køber and colleagues'2Køber L Torp-Pedersen C Cole D Hampton JR Camm AJ Bayesian interim statistical analysis of randomised trials: the case against.Lancet. 1997; 349: 1168-1169Summary Full Text Full Text PDF PubMed Scopus (4) Google Scholar accompanying case against. However, an alternative form of Bayesian interim analysis should be less controversial. This analysis being used for monitoring the UK Medical Research Council (MRC) Oesophageal Carcinoma Trial OE02, comparing survival for patients randomised to surgery with or without adjuvant chemotherapy.3Fayers PM, Ashby D, Parmar MKB. Tutorial in biostatistics: Bayesian data monitoring in clinical trials. Stat Med (in press).Google Scholar The method was tested on earlier MRC trials, and used in MRC trials of continuous hyperfractionated accelerated radiotherapy for head and neck cancer and bronchial cancer.4Spiegelhalter DJ Freedman LS Parmar MKB Bayesian approaches to randomised trials.J Roy Statist Soc. 1994; 157: 357-416Crossref Google Scholar It is simple to implement. Colleagues and I have described the method in detail, and recommend this procedure for monitoring trials.3Fayers PM, Ashby D, Parmar MKB. Tutorial in biostatistics: Bayesian data monitoring in clinical trials. Stat Med (in press).Google Scholar The aim of a clinical trial should be to influence medical practice. If a clinical trial detects a large treatment effect after entering half the patients, and consequently terminates early, the results may be received with considerable scepticism. Despite any significant p values, many clinicians may still remain unconvinced by the weight of evidence that has been produced. These clinicians are likely to continue treating new patients in the same way as in the past. The clinical trial, therefore, will have failed in its primary objective of altering the management of future patients. Thus, many trialists are cautious about stopping recruitment prematurely. The International Study of Infarct Survival (ISIS) trials, for example, state in protocols that the interim results will only be disclosed to the steering committee if there is “evidence that might reasonably be expected to influence materially the patient management of many clinicians”.5ICRF Clinical Trial Service UnitISIS 3 Protocol. CTSU, Radcliffe Infirmary, Oxford, UK1989Google Scholar Our Bayesian approach formalises the concept of prestudy beliefs being influenced by the results from an experiment such as a clinical trial, yielding revised beliefs. Clinicians' prestudy beliefs are evaluated and expressed as a prior distribution.4Spiegelhalter DJ Freedman LS Parmar MKB Bayesian approaches to randomised trials.J Roy Statist Soc. 1994; 157: 357-416Crossref Google Scholar The trial will only stop early if the results to date are deemed sufficiently conclusive to influence opinions. Unlike Brophy and Joseph, we do not attempt to combine information from other trials, but instead focus on the prevailing beliefs of clinicians, since that is what we seek to influence. Also, we use this Bayesian approach solely for trial monitoring and to decide whether to recommend early termination of the trial. On trial completion, we may additionally report traditional p values (adjusted for the number of interim looks at the data). Many clinicians find Bayesian concepts intuitively appealing. The idea of collecting sufficient data to convince not only enthusiasts but also those with open minds, and sceptics too, accords with most clinicians' experience of research and the introduction of new drugs. Statistical, including Bayesian, stopping rules should be used with circumspection. Trials should never be closed without full consideration of the impact of early termination. Nevertheless, we believe our Bayesian approach makes explicit many of the issues involved in monitoring of trials, and deserves to be more widely used.