genridge (version 0.8.0)

ridge: Ridge Regression Estimates

Description

The function ridge fits linear models by ridge regression, returning an object of class ridge designed to be used with the plotting methods in this package.

It is also designed to facilitate an alternative representation of the effects of shrinkage in the space of uncorrelated (PCA/SVD) components of the predictors.

The standard formulation of ridge regression is that it regularizes the estimates of coefficients by adding small positive constants \(\lambda\) to the diagonal elements of \(\mathbf{X}^\top\mathbf{X}\) in the least squares solution to achieve a more favorable tradeoff between bias and variance (inverse of precision) of the coefficients.

$$\widehat{\mathbf{\beta}}^{\text{RR}}_k = (\mathbf{X}^\top \mathbf{X} + \lambda \mathbf{I})^{-1} \mathbf{X}^\top \mathbf{y} $$

Ridge regression shrinkage can be parameterized in several ways.

• If a vector of lambda values is supplied, these are used directly in the ridge regression computations.

• Otherwise, if a vector df can be supplied the equivalent values for effective degrees of freedom corresponding to shrinkage, going down from the number of predictors in the model.

In either case, both lambda and df are returned in the ridge object, but the rownames of the coefficients are given in terms of lambda.

coef extracts the estimated coefficients for each value of the shrinkage factor

vcov extracts the estimated \(p \times p\) covariance matrices of the coefficients for each value of the shrinkage factor.

best extracts the optimal shrinkage values according to several criteria: HKB: Hoerl et al. (1975); LW: Lawless & Wang (1976); GCV: Golub et al. (1975)

Usage

ridge(y, ...)
# S3 method for formula
ridge(formula, data, lambda = 0, df, svd = TRUE, contrasts = NULL, ...)
# S3 method for default
ridge(y, X, lambda = 0, df, svd = TRUE, ...)
# S3 method for ridge
coef(object, ...)
# S3 method for ridge
print(x, digits = max(5, getOption("digits") - 5), ...)
# S3 method for ridge
vcov(object, ...)
best(object, ...)
# S3 method for ridge
best(object, ...)

Value

A list with the following components:

lambda

The vector of ridge constants

df

The vector of effective degrees of freedom corresponding to lambda

coef

The matrix of estimated ridge regression coefficients

scales

scalings used on the X matrix

kHKB

HKB estimate of the ridge constant

kLW

L-W estimate of the ridge constant

GCV

vector of GCV values

kGCV

value of lambda with the minimum GCV

criteria

Collects the criteria kHKB, kLW, and kGCV in a named vector

If svd==TRUE (the default), the following are also included:

svd.D

Singular values of the svd of the scaled X matrix

svd.U

Left singular vectors of the svd of the scaled X matrix. Rows correspond to observations and columns to dimensions.

svd.V

Right singular vectors of the svd of the scaled X matrix. Rows correspond to variables and columns to dimensions.

A data.frame with one row for each of the HKB, LW, and GCV criteria

Arguments

y

A numeric vector containing the response variable. NAs not allowed.

...

Other arguments, passed down to methods

formula

For the formula method, a two-sided formula.

data

For the formula method, data frame within which to evaluate the formula.

lambda

A scalar or vector of ridge constants. A value of 0 corresponds to ordinary least squares.

df

A scalar or vector of effective degrees of freedom corresponding to lambda

svd

If TRUE the SVD of the centered and scaled X matrix is returned in the ridge object.

contrasts

a list of contrasts to be used for some or all of factor terms in the formula. See the contrasts.arg of model.matrix.default.

X

A matrix of predictor variables. NA's not allowed. Should not include a column of 1's for the intercept.

x, object

An object of class ridge

digits

For the print method, the number of digits to print.

Author

Michael Friendly

Details

If an intercept is present in the model, its coefficient is not penalized. (If you want to penalize an intercept, put in your own constant term and remove the intercept.)

The predictors are centered, but not (yet) scaled in this implementation.

A number of the methods in the package assume that lambda is a vector of shrinkage constants increasing from lambda[1] = 0, or equivalently, a vector of df decreasing from \(p\).

References

Hoerl, A. E., Kennard, R. W., and Baldwin, K. F. (1975), "Ridge Regression: Some Simulations," Communications in Statistics, 4, 105-123.

Lawless, J.F., and Wang, P. (1976), "A Simulation Study of Ridge and Other Regression Estimators," Communications in Statistics, 5, 307-323.

Golub G.H., Heath M., Wahba G. (1979) Generalized cross-validation as a method for choosing a good ridge parameter. Technometrics, 21:215–223. tools:::Rd_expr_doi("10.2307/1268518")

See Also

lm.ridge for other implementations of ridge regression

traceplot, plot.ridge, pairs.ridge, plot3d.ridge, for 1D, 2D, 3D plotting methods

pca.ridge, biplot.ridge, biplot.pcaridge for views in PCA/SVD space

precision.ridge for measures of shrinkage and precision

Examples

#\donttest{
# Longley data, using number Employed as response
longley.y <- longley[, "Employed"]
longley.X <- data.matrix(longley[, c(2:6,1)])

lambda <- c(0, 0.005, 0.01, 0.02, 0.04, 0.08)
lridge <- ridge(longley.y, longley.X, lambda=lambda)

# same, using formula interface
lridge <- ridge(Employed ~ GNP + Unemployed + Armed.Forces + Population + Year + GNP.deflator, 
		data=longley, lambda=lambda)


coef(lridge)

# standard trace plot
traceplot(lridge)
# plot vs. equivalent df
traceplot(lridge, X="df")
pairs(lridge, radius=0.5)
#}

# \donttest{
data(prostate)
py <- prostate[, "lpsa"]
pX <- data.matrix(prostate[, 1:8])
pridge <- ridge(py, pX, df=8:1)
pridge

plot(pridge)
pairs(pridge)
traceplot(pridge)
traceplot(pridge, X="df")
# }

# Hospital manpower data from Table 3.8 of Myers (1990) 
data(Manpower)
str(Manpower)

mmod <- lm(Hours ~ ., data=Manpower)
vif(mmod)
# ridge regression models, specified in terms of equivalent df
mridge <- ridge(Hours ~ ., data=Manpower, df=seq(5, 3.75, -.25))
vif(mridge)

# univariate ridge trace plots
traceplot(mridge)
traceplot(mridge, X="df")

# \donttest{
# bivariate ridge trace plots
plot(mridge, radius=0.25, labels=mridge$df)
pairs(mridge, radius=0.25)

# 3D views
# ellipsoids for Load, Xray & BedDays are nearly 2D
plot3d(mridge, radius=0.2, labels=mridge$df)
# variables in model selected by AIC & BIC
plot3d(mridge, variables=c(2,3,5), radius=0.2, labels=mridge$df)

# plots in PCA/SVD space
mpridge <- pca(mridge)
traceplot(mpridge, X="df")
biplot(mpridge, radius=0.25)
# }