SUHASINI SUBBA RAO
Department of Statistics
Texas A&M University
Linear Regression with Time Series Regressors
ABSTRACT
In several diverse applications, from the neurosciences to econometrics, it is of interest to model the influence observed regressors have on a response of interest. In many of these applications, the regressors have a meaningful ordering and are usually a long time series. The problem of linear regression, where the number of regressors n is of the same order or magnitudes larger than the number of responses p has received considerable attention. However, most of these approaches place a sparsity assumption on regressor coefficients. When the regressors are a time series, the sparse assumption can be unrealistic with no intuitive interpretation.
In this talk we consider the problem of linear regression with time series regressors, but work under the assumption that the regressor coefficients are absolutely summable. We propose a computationally efficient method for estimating the regression parameters, that avoids large matrix estimation and inversion. We show that the parameter estimators are consistent and derive a central limit theorem. The proposed estimation scheme, leads to a simple method for estimating the variance of the parameter estimators. Though consistent the parameter estimators are noisy, thus we describe a post-processing step to reduce the noise in the estimators.
Friday, 3/23/2018, 11:30 AM, BLOC 113
