Comparisons of Model Selection Strategies in Time-to-Event Models

• Activity Number: 531
• Type: Contributed
• Date/Time: Wednesday, August 5, 2009 : 2:00 PM to 3:30 PM
• Sponsor: Section on Statistical Computing
• Author(s): Lin Pan*+ and Zhiheng Xu and Haibin Wang and Mourad Tighiouart
• Companies: Emory University and Emory University and Emory University and Emory University
• Keywords: Automatic Model Selection ; Collett Manual Selection ; SAS macro ; rank criterion

Abstract

Automatic variable selection methods such as forward, backward, and stepwise are popular methods for determining prognostic models in linear and logistic regression models and time to event data such as overall survival and disease free survival. They are routinely used in the medical literature despite their poor statistical properties, especially in the presence of multicollinearity. In this talk, we compare different automatic model selection strategies with the manual selection procedure proposed by Collett (2003) by extensive simulations and real data sets. We present a SAS Macro for carrying out the 4-step procedure. Different model selection criteria, AIC, BIC, Cp, LRT and Bayesian criterion, are applied and evaluated when selecting a model. The advantages and disadvantages of each searching approach and ranking criterion are discussed.