Methods for Matching Subjects on Baseline Outcome Measures Prior to Randomization

• Activity Number: 219
• Type: Contributed
• Date/Time: Tuesday, August 10, 2004 : 10:30 AM to 12:20 PM
• Sponsor: WNAR
• Author(s): Misook Park*+ and Robert E. Johnson
• Companies: Virginia Commonwealth University and Virginia Commonwealth University
• Keywords: clustering ; matching ; baseline measures

Abstract

Matching subjects prior randomization may increase the power to detect differences between subgroups by reducing the variance of the differences. Common methods include pairing where two subjects are matched on one or more characteristics. Matched groups of size greater than two may be formed in a similar way. One approach to the latter is to use an algorithm for identifying clusters of observations. Once identified, subjects within each cluster are randomized to the study arms. It is necessary to place constraints on the clustering algorithm, namely, each cluster must have at least the same number of subjects as arms of the study. For example, a hierarchical method for identifying clusters may result in one or more clusters having less than the required number of subjects. These clusters must then be fused with a neighboring cluster. This fusion, though necessary, may result in a nonoptimal result. For example, variation may be improved by breaking apart the newly formed cluster. We will explore adaptations of known methods in the presence of such constraints. Hierarchical methods such as the centroid and average linkage will be considered in addition to an all-clusters approach.