Conditional Power—How Likely Is Trial Success?

INTRODUCTION

This JAMA Guide to Statistics and Methods article examines conditional power, calculated while a trial is ongoing and based on both the currently observed data and an assumed treatment effect for future patients.

Because of the costs and potential risks to participants, clinical trials should be initiated or continued only when there is a reasonable chance the results will improve clinical care or represent important new knowledge. McNamee et al¹ reported the results of a randomized clinical trial evaluating the therapeutic benefit of lower tidal volume ventilation coupled with extracorporeal carbon dioxide removal compared with standard ventilatory support in patients with acute hypoxemic respiratory failure. During a pause in recruitment for investigation of a severe adverse event, the independent data monitoring and ethics committee (DMC) made a recommendation to stop the trial early for futility, based on the feasibility of future recruitment and the low likelihood of eventual trial success. The DMC used a conditional power calculation to show that even under the most optimistic assumptions of treatment benefit, the trial had only a 44% chance of demonstrating a statistically significant benefit from the intervention on 90-day all-cause mortality.¹

WHAT IS CONDITIONAL POWER?

The power of a clinical trial is the probability of obtaining a positive result (eg, a statistically significant P value or a bayesian posterior probability above a threshold), assuming there is a real beneficial treatment effect of a specified magnitude. Before initiation of a trial, the trial’s power is estimated according to assumptions about the magnitude of the treatment effect, the variability of patient responses, and other parameters.²

In contrast, a conditional power calculation is conducted while a trial is ongoing and is based on both the currently observed data for some (but not all) patients and an assumed treatment effect for future patients. Data that have already been collected are considered known and assumed not to change. Future data are assumed to reflect a single treatment effect, or 1 of a set of possible treatment effects. Given both these observed data and assumptions regarding future data, one can compute the probability that a positive trial result will be obtained at the planned final sample size. This probability is known as conditional power because it is conditional on the data obtained to that point.^3-6 Common choices for defining the assumed future treatment effect are to project the currently observed treatment effect or to assume the treatment effect used in the original trial planning.³

Conditional power is sensitive to the treatment effect assumed for future patients.³ Because of the uncertainty in this assumption, conditional power is often computed for each of multiple possible values for the assumed treatment effect, resulting in a set of conditional power values.⁴ One can evaluate the likelihood of trial success using this set of conditional power calculations, combined with a ...