|
Activity Number:
|
61
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Sunday, August 3, 2003 : 4:00 PM to 5:50 PM
|
|
Sponsor:
|
Biopharmaceutical Section
|
| Abstract - #301988 |
|
Title:
|
Using Random Forests to Detect Covariate Interaction Effects in Case-Control Studies: Applications to Screening for Disease Genes
|
|
Author(s):
|
Steve Horvath*+ and Peter Kraft
|
|
Companies:
|
University of California, Los Angeles and University of California, Los Angeles
|
|
Address:
|
Dept. of Human Genetics, Gonda Building, School of Medicine, Los Angeles, CA, 90095-7088,
|
|
Keywords:
|
random forest ; variable selection ; SNPs ; allelic association test ; screening
|
|
Abstract:
|
Many studies in data-rich fields such as bioinformatics require screening a large number of covariates for a potentially small number associated with an outcome. This entails a number of statistical problems, including multiple comparisons, low power to detect covariates with small marginal effects but large interaction effects, and a potentially large number of parameters to be estimated relative to the number of observations. Random forest predictors [Breiman 2002] are a state-of-the-art supervised learning method well-suited for data with many covariates but relatively few observations. One of the attractive features of random forest predictors is that they produce model-free measures of covariate importance. We present a novel permutation procedure which uses these importance measures to select a set of covariates plausibly associated with outcome and limits the overall false-positive rate. We contrast this procedure to standard covariate selection procedures such as forward stepwise selection by applying it to simulated and real data. In particular, we discuss how to use the procedure in allelic association studies.
|
