ECE: Electrical & Computer Engineering
ECE News

CBIL Seminar: Functional Group Lasso Using Adaptive Weights Based on Extremum

12:00 AM - 10:30 AM on Wednesday, October 16, 2013
Location: VTRC - Arlington; 4-024; 900, N Glebe Rd, Arlington, VA 22203

Invited Speaker: Dr. Qing Pan, George Washington University, Dept. of Statistics

Many clinical trials and epidemiology studies produce
longitudinal clinical outcome together with high-dimensional genomic
markers, where screening methods accounting for temporal dynamic
genetic effects are desirable. We propose a novel group sparse
variable selection method with adaptive weights based on the maximum
effects of each marker over time. The effects of candidate markers are
modeled as nonparametric smoothing functions of time. Splines with
large peaks are selected with adaptive group lasso, whose weights are
negative functions of the largest absolute value of the spline
coefficients. The proposed adaptive procedure shows improved
specificity, larger area under the receiver operating curve (ROC),
smaller bias in the coefficient estimates as well as lower prediction
errors in extensive simulation studies. We apply our method to the
Genome Wide Association Study (GWAS) data from the Epidemiology and
Intervention of Diabetes Complication (EDIC) trial where Type 1
Diabetes patients are followed for up to 26 years. Fourteen SNPs
possibly associated with glomerular filtration rates (GFR) are