Would shorter clinical trials have HIGHER statistical power? A supercomputer simulation study
Abstract number :
1.279
Submission category :
7. Antiepileptic Drugs / 7B. Clinical Trials
Year :
2017
Submission ID :
336375
Source :
www.aesnet.org
Presentation date :
12/2/2017 5:02:24 PM
Published date :
Nov 20, 2017, 11:02 AM
Authors :
Daniel M. Goldenholz, National Institutes of Health; Joseph Tharayil, Duke University Medical Center; Robert Moss, SeizureTracker.com; Evan Myers, Duke University Medical Center; and William H. Theodore, National Institutes of Health
Rationale: Recent evidence that natural variability in seizure frequency may be predictable suggests that placebo response rates may be predictable based on parameters of clinical trials. Therefore, it may be possible to predict clinical trial parameters a priori that would achieve higher statistical power, and possibly lower cost. Methods: : Using one of the world’s largest patient reported seizure diary databases, SeizureTracker.com, we derived virtual patients for simulated RCTs. We ran 1000 repetitions of simulated trials using random sampling with replacement for each combination of trial parameters, including baseline and test period duration, number of patients, eligibility criteria, and drug effect size. Drug effect was simulated as a randomly chosen number, per patient, from a Gaussian distribution centered around 10,20,30 or 40% (depending on the trial parameters). Seizures were deleted (or added) probabilistically using the per patient drug effect size. Placebo was simulated by leaving the natural variability of the seizure diary unchanged, in line with recent studies. We simulated 6,732,000 trials using the NIH supercomputer. We studied the resulting trial statistical power of the Fisher exact test to distinguish simulated drug from placebo. To estimate power of a given combination of parameters, the fraction of the 1000 repetitions that achieve p Results: Simulating all trials derived from 5097 original patients, we obtained maps of statistical power as a function of drug strength, trial duration and eligibility criteria. Plotted in the figure is a heat map of statistical power across 1000 repetitions of a given parameter set. The maps show combinations of trial durations, number of patients, and average drug strength. One striking feature of these maps is that using smaller trial durations, such as changing the “test” period from 12 weeks to 4, resulted in a dramatic decrease in trial duration, required number of patients, and large increases in statistical efficiency. Conclusions: Under the assumption that natural variability drives placebo response, higher statistical power can be achieved using smaller trial durations, even with fewer patients. In these simulations, there may be a benefit for trial size, power and ultimately cost by shortening standard trials. Funding: National Institute of Neurological Disease and Stroke Intramural Research Program
Antiepileptic Drugs