Variable Pruning and Colinearity and their Effect on Model Building in Pharmaceutical Sciences

• Activity Number: 18
• Type: Contributed
• Date/Time: Sunday, August 3, 2003 : 2:00 PM to 3:50 PM
• Sponsor: ENAR
• Author(s): Kimberly Sue Crimin*+ and Joseph W. McKean and Thomas J. Vidmar
• Companies: Pfizer, Inc. and Western Michigan University and Pharmacia
• Keywords: model building ; variable selection ; colinearity

Abstract

Recently in drug discovery, emphasis is on computational models to predict structure activity and property relationships. These data can often be characterized by a few observations and a large number of measured variables that are often highly correlated. Traditional approaches to modeling these data are principle component regression (PCR) and partial least squares (PLS). Often, the variables are pruned by looking at the loadings given to each variable and the criteria for removing variables is inconsistent between datasets. We present a variable selection method that is based on stabile numerical linear algebra techniques. Our technique ranks the predictor variables in terms of importance, taking into account the correlation between the response and the predictor variables. Our method deals with the colinearity problem by separating correlated variables during the ranking. We view our method as a variable pruning technique that can be used prior to multiple regression, PCR, or PLS. We also investigate the effect of colinearity on PCR and PLS.