Da. Elston et Mf. Proe, SMOOTHING REGRESSION-COEFFICIENTS IN AN OVERSPECIFIED REGRESSION-MODEL WITH INTERRELATED EXPLANATORY VARIABLES, Applied Statistics, 44(3), 1995, pp. 395-406
Standard rank reducing methods such as principal components regression
and partial least squares take account of the x-variables being inter
related only through the observed correlation structure in the data. A
lternatively, such interrelationships can be taken into account by usi
ng generalized ridge regression, motivated as a form of penalized like
lihood estimation. When the form of the penalty function is chosen app
ropriately, it has the effect of smoothing the regression coefficients
corresponding to successive x-variables in a similar way to the use o
f spline functions for interpolation using a single x-variable. Such a
smoothing of the regression coefficients is demonstrated in an exampl
e in which the response variable is a single measure of the growth rat
e of Sitka spruce trees at each of 363 sites in Scotland and the 12 x-
variables available are 30-year mean temperatures for each calendar mo
nth at each site. The smoothing used penalizes the sum of squared thir
d differences of the regression coefficients and leads to a reduction
in the average variance of the fitted values of two-thirds when compar
ed with the unsmoothed regression. Furthermore, the smoothed regressio
n coefficients suggest a bimodal relationship between growth rate and
monthly mean temperature which would probably be missed by using stand
ard rank reducing techniques.