Monday, March 12, 2001
Bayesian statistics have been around for more than 200 years.
But even though their use in clinical trials would offer advantages over the
more familiar frequentist statistics in certain situations, caution on the part
of regulators and the lack of experience with Bayesian methods among biostatisticians
have limited use of the technique. In fact, Bayesian statistics have never been
used for regulatory approval of therapeutics.
That is about to change. Pfizer Inc. is conducting the first
dose-finding Phase IIb trial of a therapeutic using an adaptive Bayesian design.
The pharma company is developing recombinant neutrophil inhibitory factor (NIF)
for acute stroke (UK-279,276). NIF was discovered by Corvas International Inc.
(CVAS, San Diego, Calif.) (see BioCentury, Dec. 4, 2000).
Adaptive trial design
PFE (New York, N.Y.) declined to comment on the trial. But Andy Grieve, a biostatistician in the biometric department of Pfizer Central Research who is involved in the NIF study, is an author of a book chapter in press called "Adaptive Bayesian Designs for Dose-Ranging Drug Trials." The chapter, written in collaboration with Donald Berry, chairman of the biostatistics department at M.D. Anderson Cancer Center in Houston, and colleagues details an adaptive Bayesian trial design for stroke drugs. It is to appear in "Case Studies in Bayesian Statistics".
The adaptive trial computer simulations described by the authors start with 16 possible doses or placebo. As data come in each week on patient responses, the next set of doses assigned are allowed to depend on information from patients treated previously. Thus as data accumulate that a certain dose is more effective, more patients will be assigned to that dose by the computer. The investigator remains blinded to the dose (see "Adaptive Dose-Ranging Trial Design", A2).
In contrast, a frequentist dose-finding trial would assign equal numbers of patients randomly to a set of pre-defined doses, typically three to five. The authors argue that such studies are stuck if the actual dose response curve turns out to be located at doses above or below those chosen for the trial; if the characteristics of the patient population turn out to be different than assumed; if the effect of treatment is different than assumed (if it is larger, the penalty is running a larger trial than necessary; if smaller, the penalty is missing the effect); or if the variability in the actual data is greater than assumed.
The simulation also was designed to roll over from a Bayesian dose-finding trial to a pivotal frequentist trial, with the duration of the dose-finding phase depending on the accumulating information about the dose-response curve. If the information is sufficient to show that there is a response to the drug and that the dose-response curve has been identified, the trial rolls over into a confirmatory, frequentist mode in which patients are assigned either the effective dose or placebo. If the compound doesn't work, the trial can be terminated early, saving money.
One advantage of such a trial may be to improve the design of confirmatory trials. "Since we know the true response at that dose we can easily calculate the power of the confirmatory study to show a significant difference from placebo," according to the authors.
They conclude that this approach uses "the flexibility and naturally adaptive nature of the Bayesian approach but with an overall goal that is consistent with frequentist tradition. It is a frequentist cake with Bayesian icing."