A Comparison of Model-Based Clustering Approaches for Skewed Data

• Activity Number: 653
• Type: Contributed
• Date/Time: Thursday, August 13, 2015 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section
• Author(s): Meredith Wallace* and Daniel Buysse and Martica H. Hall and David Kupfer and Satish Iyengar
• Companies: University of Pittsburgh and University of Pittsburgh and University of Pittsburgh and University of Pittsburgh and University of Pittsburgh
• Keywords: clustering ; mixture model ; sleep ; polysomnography ; skewed data ; medicine

Abstract

Clustering with finite mixture models (FMMs) can reveal subgroups with similar characteristics. However, skewed data are common in medicine, and different FMM approaches can produce substantially different findings for such data. The Gaussian FMM is a commonly used method that assumes a skewed distribution is a mixture of normal distributions. Researchers might also consider transforming the data prior to fitting a Gaussian FMM in an attempt to better meet the model's assumptions. Alternatively, the recently proposed skew-t FMM allows clusters to follow skewed distributions. The aim of this work is to provide an empirical assessment and comparison of these recent and conventional FMM methods in order to guide researchers in selecting the most appropriate approach for their data. To accomplish this aim, these approaches are reviewed with respect to the way they manage skewed data, empirically evaluated and compared in a simulation study, and demonstrated using sleep diary and inflammatory marker data.