Three Approaches to Modeling Binary Outcomes in Multicenter Clinical Trials with Clustering Due to Clinical Sites

• Activity Number: 230
• Type: Contributed
• Date/Time: Tuesday, August 9, 2005 : 8:30 AM to 10:20 AM
• Sponsor: Biopharmaceutical Section
• Author(s): Douglas Thompson*+ and William McCarthy and Bruce Barton
• Companies: Maryland Medical Research Institute and Maryland Medical Research Institute and Maryland Medical Research Institute
• Keywords: multicenter clinical trials ; clustered binary data ; random effects models ; logistic regression ; SUDAAN ; SAS

Abstract

In clinical trials, it is commonplace to model a binary outcome (e.g., recovery from an illness) as a function of a set of predictors (e.g., treatment, age, gender) using logistic regression. Ordinary logistic regression assumes independence among the observations (e.g., patients). This is problematic in multicenter trials because clustering within clinical sites may create nonindependence. If the clustering is ignored, incorrect variance estimates may result. The purpose of this paper is to describe three approaches to modeling binary outcomes in multicenter trials where there is clustering due to clinical sites: conditional logistic regression, Taylor series variance estimation techniques, and random effects modeling. First, the three methods are described and contrasted. Second, implementation of the techniques using SUDAAN, Mplus, and SAS (procs PHREG, NLMIXED, and SURVEYLOGISTIC) is discussed. Third, an example of each approach is provided using data from the Neurologic Outcomes and Preemptive Analgesics in Neonates Multicenter Trial. Finally, the effects of different degrees of clustering on the estimates of each approach are illustrated using Monte Carlo simulations.