ABSTRACT
I. INTRODUCTION Since Sir Austin Bradford Hill in 1946 introduced the controlled clinical trial for the Streptomycin in Tuberculosis Trials Committee of the Medical Research Council (1948), the use of double-blind, randomized, and controlled clinical trials has emerged as a principal method for the evaluation of new drugs or therapeutic procedures in medicine. Ethical and economic considerations are prominent in all clinical trials in new drug research and development. With the exception of phase I and early phase II studies, most clinical trials are multicenter studies, in which patient entry is usually sequential and staggered. There is a strong ethical or economic obligation for the sponsor or research group to review or analyze the interim data periodically for evidence of efficacy and safety over the course of the trial. As compelling evidence emerges, either favoring or disfavoring the new therapy, it may become ethically or economically necessary to terminate the trial before schedule. Although periodic evaluation of interim data is a frequent and necessary practice in drug development, particular statistical problems of multiple testing and bias do appear. Classical clinical trial designs do not formally provide the option for early termination. Rather, classical designs consider only fixed-
this approach to sequential comparison of two exponential distributions with censored data. The optimal criterion is to minimize the posterior expected regret of stopping. Cornfield (1966a, b) proposed another approach, termed relative betting odds (RBOs). Although the RBO approach was used by Cornfield and later explored by Lachin (1981), it has not been used extensively. Recently, Berry (1985, 1987, 1989) criticized the arbitrariness of a and {3 in the hypothesis testing (Neyman-Pearson) formulation and failure of frequentists to exploit a priori information. He repeatedly argued that the virtue of the Bayesian approach lies in the fact that subsequent Bayesian inferences are not affected by frequent or even continuous evaluations. However, the need to prespecify and adhere to a prior distribution on the parameters remains one of the major barriers for the adaptation of a Bayesian approach to clinical trials. Using previous trial information to develop a priori is not as straightforward as it may seem; not to mention the ever-changing patient population characteristics and medical practices. In advocating the use of the fixed-sample p-value without adjusting for sequential tests as a pragmatic compromise, Dupont (1983) argued that the primary use of p-values, which are most familiar and almost universally accepted in medicine, is as measures of inferential strength. However, Brown (1983) called it a deliberate misusage for not providing any necessary adjustments.
