High dimensional linear regression models under long memory dependence and measurement error
This dissertation consists of three chapters. The first chapter introduces the models under consideration and motivates problems of interest. A brief literature review is also provided in this chapter.The second chapter investigates the properties of Lasso under long range dependent model errors. Lasso is a computationally efficient approach to model selection and estimation, and its properties are well studied when the regression errors are independent and identically distributed. We study the case, where the regression errors form a long memory moving average process. We establish a finite sample oracle inequality for the Lasso solution. We then show the asymptotic sign consistency in this setup. These results are established in the high dimensional setup (p>n) where p can be increasing exponentially with n. Finally, we show the consistency, n^(1/2-d)-consistency of Lasso, along with the oracle property of adaptive Lasso, in the case where p is fixed. Here d is the memory parameter of the stationary error sequence. The performance of Lasso is also analysed in the present setup with a simulation study.The third chapter proposes and investigates the properties of a penalized quantile based estimator for measurement error models. Standard formulations of prediction problems in high dimension regression models assume the availability of fully observed covariates and sub-Gaussian and homogeneous model errors. This makes these methodsinapplicable to measurement errors models where covariates are unobservable and observations are possibly non sub-Gaussian and heterogeneous. We propose weighted penalized corrected quantile estimators for the regression parameter vector in linear regression models with additive measurement errors, where unobservable covariates are nonrandom. The proposed estimators forgo the need for the above mentioned model assumptions. We study these estimators in both the fixed dimension and high dimensional sparse setups, in the latter setup, the dimensionality can grow exponentially with the sample size. In the fixed dimensional setting we provide the oracle properties associated with the proposed estimators. In the high dimensional setting, we provide bounds for the statistical error associated with the estimation, that hold with asymptotic probability 1, thereby providing the l_1-consistency of the proposed estimator. We also establish the model selection consistency in terms of the correctly estimated zero components of the parameter vector. A simulation study that investigates the finite sample accuracy of the proposed estimator is also included in this chapter.
Read
- In Collections
-
Electronic Theses & Dissertations
- Copyright Status
- In Copyright
- Material Type
-
Theses
- Authors
-
Kaul, Abhishek
- Thesis Advisors
-
Koul, Hira L.
- Committee Members
-
Ramamoorthi, R.V
Zhong, Ping-Shou
Wang, Hao
- Date
- 2015
- Subjects
-
Error analysis (Mathematics)
Mathematical statistics
Regression analysis--Mathematical models
- Program of Study
-
Statistics - Doctor of Philosophy
- Degree Level
-
Doctoral
- Language
-
English
- Pages
- viii, 97 pages
- ISBN
-
9781321893854
132189385X
- Permalink
- https://doi.org/doi:10.25335/rc4m-zq68