Correlation in Polynomial Regression.
Abstract
It has been stated that centering of the independent variable in quadratic regression reduces the correlation between linear and quadratic terms. This in turn reduces or removes ill conditioning of the matrix to be fitted in the use of least squares. It is shown in this paper how the correlation between linear and quadratic terms depends on c, the departure from centering of the independent variable in terms of a measure of variation in the variable. It is shown that the correlation approaches unity in absolute value as the absolute value of c approaches infinity and that it is near unity in absolute value for even modest values of the absolute value of c. It is shown further that with appropriate centering, the correlation may be reduced to zero. The results developed for simple quadratic regression extend easily to polynomial regression with similar effects. (Author)
Document Details
- Document Type
- Technical Report
- Publication Date
- Mar 01, 1977
- Accession Number
- ADA039547
Entities
People
- Ralph A. Bradley
- Sushil S. Srivastava
Organizations
- Florida State University