Variable Selection with Exclusion Frequency-Based Weights: Application to a Neuroimaging Study of Huntington's Disease

• Activity Number: 662
• Type: Invited
• Date/Time: Thursday, August 13, 2015 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Computing
• Author(s): Tanya P. Garcia* and Samuel Mueller and Karen Marder
• Companies: Texas A&M University and University of Sydney and Columbia University
• Keywords: Exclusion frequency ; Proportional hazards model ; Weighted Lasso ; Neurodegenerative disease

Abstract

Current clinical studies have evolved into collecting large amounts of data on a small subset of subjects, with the intent being to not overlook any important features. A key goal in such studies is parsing through the data to identify informative features and make inference about their effects on a response. The goal becomes challenging, however, when the features are naturally intercorrelated and the response is a possibly censored failure time. We propose to handle these challenges through a new variable selection technique which casts the problem into several smaller dimensional settings and extracts from this intermediary step the relative importance of each feature via so-called exclusion frequencies. The exclusion frequencies are used as weights in an adaptive Lasso and shown to yield low false discovery rates and high geometric mean of sensitivity and specificity. We show the method's advantages over existing ones in an extensive simulation study and apply the method to a neuroimaging study of Huntington's disease to identify those brain features most relevant to age of disease onset. The method is also shown to work in other settings, such as generalized linear models.