Handling Missing Observations in the Estimation of Differences in Proportions from Clustered Data in a Matched-pair Design

• Activity Number: 383
• Type: Contributed
• Date/Time: Wednesday, August 11, 2004 : 2:00 PM to 3:50 PM
• Sponsor: ENAR
• Author(s): Carsten Schwenke*+ and Jaakko Nevalainen
• Companies: Schering AG and Schering AG
• Keywords: missing values ; matched-pair ; clustered data ; binary outcomes

Abstract

In diagnostic studies, often the difference in the sensitivity or specificity of two diagnostic procedures is to be estimated. We consider studies with a matched-pair design, where all procedures are applied to each patient. With more than one observational unit per patient, e.g., several tumors in the same patient, statistical methods are to be used, which account for the clustering effect. Obuchowski (1998) developed such methods for complete datasets. In terms of clinical studies, this approach is often equivalent to the per-protocol-analysis, where incomplete cases are excluded. For an intent-to-treat analysis, these methods are insufficient. We propose a simple extension, where incomplete datasets can be analyzed without excluding incomplete cases. We divide the full sample into a set of data with complete cases and another set of data with incomplete cases. The differences in proportions are estimated for each subset considering the correlation between patients for the complete data and the clustering effect within the patient for both datasets. Then the two estimated differences are combined by appropriate weights.