STATISTICS IN MEDICINE, VOL. 13, 1471-1472 (1994)

DISCUSSION OF ‘EARLY STOPPING RULES - CLINICAL PERSPECTIVES AND ETHICAL CONSIDERATIONS’

G. BURKE Dirision of‘ O n c o l o g ~ ~ , Food und Drug Arlniinis!rurrori. Rockoilli~. M D 20857. U . S . A

Dr. Baum has described data monitoring behaviours for four clinical trials supporting the need for flexible approaches to trial curtailment based on information available to a data monitoring committee from various sources. He has discussed these trials in the context of ethical consider- ations and the pitfalls of premature stopping. That the same stopping rule for every clinical trial is not appropriate is clear. The appropriate choice of stopping rule should depend upon the design and the specific clinical questions that are addressed in each trial.

I will briefly comment upon three of the trials discussed by Dr. Baum in an attempt to further explore the consequences of monitoring procedures - the Neutron Trial, the Carotid Endarterectomy Trial, and the Anal Cancer Trial - to highlight particular issues from a regulatory point of view.

The neutron trial as initially planned did not include a prospective plan for interim analysis. In settings where it may be anticipated that ethical concerns could arise that would lead to a desire not to continue entry of patients into an inferior arm, a pre-specified stopping rule may provide a reassuring structure for subsequent statistical analysis. Most trials we now see include some such plan. The early termination of the Neutron trial, while not done according to a plan, appears reasonable, and there would, of course, be no regulatory impact of the decision as the experi- mental treatment was inferior. Consider, however, what would have occurred had the monitoring committee maintained a blinded status as to treatment arm or if neutron therapy had been known to be superior. It is doubtful that the trial would have been stopped in either case, but if it had, such early stopping would have introduced considerable uncertainty and probably disbelief. Repetition might well have been necessary.

In contrast, the endarterectomy trial apparently incorporated a pre-specified plan based on three defined strata. A n early termination in this setting would have supplied strong evidence for benefit in the high risk group, but without the plan, without specification of strata, the results would have been far less convincing because of the post-hoc nature of the analysis. An interesting question is whether the early morbidity from surgery was uniform across all strata and if so the demand (results that are worth the cost) for the continued middle stratum could have been readjusted for an early stopping rule based on the requirement for the effect to be greater than a certain cost of the now better estimated early surgical morbidity. An additional ethical issue is whether newly entered patients in the middle stratum should be informed of the early results now available from the interim analysis.

The anal cancer study presents again the issue of the sensitivity of stopping behaviour to blinding effects. If the study had been blinded at the time of the analysis, the decision to proceed may not have been made. When additional patients are added to a study in this setting the results

0 1994 by John Wiley & Sons, Ltd.

1472 DISCUSSION

become conditional upon the data already analysed. The true (or generalizable) effect can not be accurately estimated.

There are many statistical, logistic and ethical questions that may arise in the setting of potential early closure (or reporting) of a clinical trial that need careful consideration. These issues can arise from data internal or external to the study. It is very important to minimize the possibility that early closure would ultimately raise doubts as to the validity of the clinical result. The greatest potential for harm occurs when changes are made in trial conduct or design as a result of accumulating unblinded data.

I will discuss briefly some of the issues that should be considered by investigators, patients, regulators, and data monitoring committees in the context of potential early closure or reporting.

If early closure for ethical reasons is at all possible in a clinical trial, a pre-specified interim analysis plan should be included in the initial protocol. The precise selection of the plan may depend on many factors such as the relative importance ofefficacy versus toxicity issues, timing of ascertainment of different endpoints, possible asymmetric stopping boundaries, and whether the monitoring committee will be blinded as to primary endpoints.

An important issue for consideration is the need in clinical trials for reliable estimates of effects which may be overestimated by stopping on a random high. Likewise, subsequent estimates of effect after longer term follow-up of cohorts from trials that were stopped early become condi- tional on the early effect and may be biased as well.

There are many additional questions. How should early results coming from other trials be used? Can there be a valid ‘meta’ interim analysis across different studies? Should analysis plans be adjusted for significant covariates? What effect will this have on the estimation problem? Is there a role for interim analyses in equivalency trials?

An important logistic question for data monitoring committees is their confidence in the quality control for the early data. Uneven censoring and other data quality issues must be addressed. Early misclassification of endpoints may lead to premature termination and com- pletely invalidate a trial. Therefore, only the firmest endpoints should be used to stop a trial.

An interesting issue is presented for ethicists, patients and investigators alike by interim analyses of accumulating data. When does one’s prior expectations regarding efficacy or toxicity sufficiently change to require reporting to I R B s and possible changes in consent forms?

The main issues that trialists must consider in achieving appropriate ethical, statistical and logistic standards are illustrated by Dr. Baum’s examples. I believe future open discussions of data monitoring behaviour such as these will be very valuable as we will sort out the complexities by understanding each others’ experience.