John Brinton, Abstract Co-Author: Nothing to Disclose

R. Edward Hendrick PhD, Abstract Co-Author: Consultant, General Electric Company Advisory Board, Bracco Group Medical Advisory Board, Koning Corporation

Deborah Glueck, Presenter: Stockholder, Bristol-Myers Squibb Company Stockholder, Life Technologies Corporation Stockholder, Johnson & Johnson Stockholder, Eli Lilly and Company Stockholder, Medco Health Solutions, Inc Stockholder, Merck & Co, Inc Stockholder, Pfizer Inc Stockholder, The Procter & Gamble Company Stockholder, Sigma Aldrich Corporation

The purpose of this research is to provide novel methods for data, power and sample size analysis for unpaired, partially-paired, and fully-paired screening trials which use the full area under the ROC curve as the outcome.

The partially-paired design is constructed by randomly assigning participants to one of three groups: participants who receive a single screening test, participants who receive the alternative screening test and a third arm of participants who receive both tests. In a fully paired design all participants receive both tests, whereas in an unpaired design all participants are randomized to one of the two tests. When running screening studies in a busy clinic, the fully-paired design may not be feasible. Conducting two examinations on all participants can slow clinic flow. A partially-paired design strikes a balance between efficient study design and clinical acceptability. We propose a novel F test that allows for the analysis of unpaired, partially-paired, and fully-paired designs. Power is calculated using an approximate F distribution. Sample size is estimated by numerically inverting the power function. We use a Monte Carlo simulation to evaluate the Type I error rate and power of the test statistic. We compare our proposed power and sample size estimation to that of a commonly used SAS macro, ROCPOWER. The novel method is applied to estimate sample size for a proposed partially paired breast cancer screening trial.  

The novel F statistic has accurate Type I error and power. The power approximations are within 0.045 of empirical estimates for all three trial designs (unpaired, partially-paired, and fully paired). Compared to results from ROCPOWER, the proposed F statistic resulted in smaller sample sizes. The sample sizes for the partially-paired designs fell between the sample sizes for the unpaired and fully-paired designs, with increasing efficiency as the ratio of participants with paired data to total participants increased.

Our approach provides an accurate approach for data, power and sample size analysis for partially-paired, fully-paired, and unpaired screening trials.

The method lets clinicians consider a range of partial-pairing ratios while designing screening trials, yielding studies that maintain power and minimize impact on clinic flow.

Brinton, J, Hendrick, R, Glueck, D, Hypothesis Testing and Power for Partially-Paired, Fully-Paired, and Unpaired Screening Trials  .  Radiological Society of North America 2014 Scientific Assembly and Annual Meeting, - ,Chicago IL. http://archive.rsna.org/2014/14016270.html