The Cytel blog keeps you up to speed with the latest developments in biostatistics and clinical biometrics.
Imagine if we were to count the number of possible reasons that investigators might have for monitoring a biomarker during a clinical trial, and multiply that number by the number of possible adaptive designs available for such investigation. We would naturally assume that whatever the number, it would be rather large. This poses an interesting question for a sponsor of an adaptive clinical trial. Are there any general principles for trial design that may be gleaned from these various possible scenarios?
Cytel statisticians are looking foward to attending the Conference of the International Society for Clinical Biostatistics, which will be held in Vienna during the week of August 24th. Members of Cytel will be contributing to four sessions at this conference, including an invited talk on innovation entitled 'Beyond Wild Horses: Developing Innovation at Cytel.' They will also be contributing to a session called Adaptive Designs II, in which they will discuss Backward Image Confidence Intervals, a solution to the problem of parameter estimation at the end of an adaptive trial.
Complexities with identifying suitable test populations in oncology studies contribute significantly to the 60% attrition rate in Phase III trials. Cyrus Mehta, (President of Cytel) has recently authored a paper on ‘Biomarker Driven Population Enrichment for Adaptive Oncology Trials,’ (forthcoming in Statistics in Medicine) which provides an innovative method for using two-stage adaptive designs for population enrichment.
Mehta, et al., are sensitive to the dilemma faced by Phase III trial designers choosing between open and restricted enrollment. Open enrollment allows for a large number of patients, and ensures that all patients who may benefit from a therapy have an opportunity to be involved. By contrast, limiting enrollment is a superior practice for revealing the efficacy of a trial for a targeted population. The proposed method allows for biomarker driven enrichment at interim analysis, meaning that only those subgroups that appear to benefit from therapy need to progress to the second stage of the trial.
The core methodological problem that would eventually spur the development of Cytel’s StatXact software was first posed by Harvard’s Marvin Zelen at a computational seminar in the late 1970s. Zelen, a distinguished professor of statistical sciences and head of the Department of Biostatistics at Harvard University, was also serving as the Director of the Dana Farber Cancer Institute.
The analysis of serious adverse events from cytotoxic agents in oncology trials were heavily dependent on an imprecise Cochran rule to measure the signifincance of small sample categorical data. The crude calculation meant that estimations of p-values were wide off the mark. Zelen challenged his students to find ways to expand Fisher’s exact test to r x c contingency tables, and by doing so to seal the promise of more effective development and delivery of urgent cancer treatments.
Cyrus Mehta and Nitin Patel took up Zelen’s challenge, publishing a series of papers on exact significance testing throughout the 1980s. Despite offering novel statistical solutions to persisting problems, the implementation of such solutions clearly required assistance from software. Unfortunately, few venture capitalists were willing to invest in a package of arcane statistical tests that were largely still in development.
Cytel was created with a grant from the National Cancer Institute, with a view to developing software that would make newer exact tests widely available for clinical studies. Its first software package, StatXact, is now used for exact testing in oncology, as well as environmental studies, public health, demography, law, and several areas of medicine and clinical development. The widespread use of exact tests has led to an array of intriguing research questions involving the power of various exact tests. Below we present a favorite finding, on the power of conditional versus unconditional exact tests: