Canonical K-Means Clustering for Constructing Moderator Importance Plots

• Activity Number: 340
• Type: Topic Contributed
• Date/Time: Tuesday, August 6, 2013 : 10:30 AM to 12:20 PM
• Sponsor: Mental Health Statistics Section
• Author(s): Thaddeus Tarpey*+ and Eva Petkova
• Companies: Wright State University and New York University
• Keywords: functional data analysis ; moderators ; placebo effects

Abstract

An approach of growing popularity to finding interesting patterns in functional data is cluster analysis. The k-means algorithm is often employed for clustering data. A shortcoming of k-means clustering is that it places cluster means in the direction of primary variability which may not necessarily correspond to clinically interesting sources of heterogeneity in treatment response. In this talk, an iterative clustering algorithm is proposed based on canonical linear transformations that minimizes within cluster variability relative to between cluster variability in order to steer the clustering algorithm in the direction of finding partitions that explain interesting heterogeneity in the data. In particular, a partitioning of the data is sought that identifies subjects that respond due primarily to specific and non-specific (i.e. placebo) effects of treatment. From this partitioning, potential moderators of treatment can be accessed for their importance in predicting whether subjects will be specific drug responders, placebo responders or non-responders to treatment.