Abstract

Variable selection plays a major role in multivariate high-dimensional statistical modeling. Hence, we need to select a consistent model, which avoids overfitting in prediction, enhances model interpretability and identifies relevant variables. We explore various continuous, nearly unbiased, sparse and accurate technique of linear model using coefficients paths like penalized maximum likelihood and nonconvex penalties, and iterative Sure Independence Screening (SIS). The convex penalized (pseudo-) likelihood approach based on the elastic net uses a mixture of the ℓ1 (Lasso) and ℓ2 (ridge regression) simultaneously achieve automatic variable selection, continuous shrinkage, and selection of the groups of correlated variables. Variable selection using coefficients paths for minimax concave penalty (MCP), starts applying penalization at the same rate as Lasso, and then smoothly relaxes the rate down to zero as the absolute value of the coefficient increases. The sure screening method is based on correlation learning, which computes component wise estimators using AIC for tuning the regularization parameter of the penalized likelihood Lasso. To reflect the eternal nature of spectral data, we use the Functional Data approach by approximating the finite linear combination of basis functions using B-splines. MCP, SIS and Functional regression are based on the intuition that the predictors are independent. However, high-dimensional grapevine dataset suffers from ill-conditioning of the covariance matrix due to multicollinearity. Under collinearity, the Elastic-Net Regularization path via Coordinate Descent yields the best result to control the sparsity of the model and cross-validation to reduce bias in variable selection. Iterative stepwise multiple linear regression reduces complexity and enhances the predictability of the model by selecting only significant predictors.

Library of Congress Subject Headings

Grapes--Genetics--Data processing; Functional analysis; Multivariate analysis

Publication Date

5-19-2017

Document Type

Thesis

Student Type

Graduate

Degree Name

Applied Statistics (MS)

Department, Program, or Center

School of Mathematical Sciences (COS)

Advisor

Peter Bajorski

Advisor/Committee Member

Jan van Aardt

Advisor/Committee Member

Ernest Fokoue

Comments

Physical copy available from RIT's Wallace Library at QA320 .J43 2017

Campus

RIT – Main Campus

Plan Codes

APPSTAT-MS

Download

Share

COinS