The Effect of Spectral Line Components on Coefficients in Time Series Regression Models

Abstract

Inference on the fitted parameters from two time series regressions can be improved by considering their correlation structure. We investigate the performance of an estimator of covariance between two time series regressions in which the responses are correlated, which is based on the multitaper method (MTM) cross-spectral estimator for the response series. We compare the MTM-based covariance estimator to the "standard" one based on the Bartlett estimate of the cross-covariance function of the two response series. A simulation study is used to evaluate performance using realizations of bivariate autoregressive processes with different characterizations of their cross-covariance function, and the effect of embedding a common deterministic line component in the response and predictor is examined.

We find that the presence of a deterministic sinusoidal component has an effect on the estimated covariance between the two regressions, and greatly increases the bias of both estimators. When common line components are detected and removed using tools within the MTM framework, covariance estimates with lower estimated mean squared errors are produced. In all cases, the MTM-based covariance estimator is found to have greater efficiency than the Bartlett-based estimator.

In an application to hourly electricity demand and price data for the province of Ontario, Canada, a linear regression model is fit in overlapping time segments in which price is designated as the response variable, and a vector of regression coefficients is obtained. Using our MTM-based covariance estimator, a covariance matrix for the coefficient vector is estimated, and more informative confidence intervals for each coefficient are obtained.