Improved Confidence Intervals for the Difference of Paired Binomials in Presence of Unpaired Data

• Activity Number: 648
• Type: Topic Contributed
• Date/Time: Thursday, August 2, 2012 : 10:30 AM to 12:20 PM
• Sponsor: Biopharmaceutical Section
• Author(s): Vivek Pradhan*+ and Yun Lu
• Companies: Boston Scientific Corporation and Boston Scientific Corporation
• Keywords: incomplete data ; paired binomial ; confidence interval ; coverage ; expected length

Abstract

Confidence interval (CI) construction of the difference of two paired binomials is required in many clinical or applied set up. In real world application, in constructing such CIs often incomplete/unpaired data are encountered; as standard practice these incomplete data are deleted and only completely observed paired data are utilized to compute CI. Tang, Ling and Tian (2009) proposed exact and approximate unconditional confidence intervals that incorporate incomplete data. It is well known that although the exact methods achieve the desired nominal coverage level, they are very conservative, and computationally costly; often for a slightly larger sample size confidence interval computation becomes very difficult. In this article, we propose a simple approach to compute CI of the difference of paired binomials in presence of incomplete data. The simulation study shows that the proposed method usually attains the nominal coverage level, and is better (in terms of coverage and expected lengths) than the existing asymptotic method.