biglars (version 1.0.2)

biglars.fit: Least-Angle, Lasso and Stepwise Regression

Description

Least-angle, lasso and stepwise methods for linear regression.

Usage

biglars.fit(x, y, type = "lasso", removeColumns = TRUE, eps = sqrt(.Machine$double.eps), blockSize = NULL, maxStages = NULL)

Arguments

x

A matrix or ff of numeric predictors. The number of columns should not be larger than the number of rows.

y

A numeric response vector or 1-column ff.

type

The type of regression to be performed. The usual choices are

lasso

lasso method (L1 penalty)

lar

least-angle regression

stepwise

forward stepwise regression

but see the details section below for other possibilities. Default is "lasso".

removeColumns

A logical scalar indicating whether columns with small variance should be removed from consideration as predictors; default `TRUE'.

eps

Numerical tolerance used for assessment of sign, equality, rank determination, column removal, etc. The default is the square root of .Machine\$double.eps, the relative machine precision.

blockSize

If NULL, the block size is determined using methods of class ff. Otherwise the passed value is used.

maxStages

The maximum number of stages allowed in the algorithm. This argument applies only to the lasso option. The defaults is 2p for lasso, where p if the number of predictors.

Value

A list with the following elements:

coef

An array of regression coefficients for each stage.

moves

Any array describing variables added or removed at each stage.

RSS

Residual sum of squares.

Details

An intercept is always included in the regression. This function calls other routines to do the core calculations, one of biglars.fit.lasso, biglars.fit.lar, or biglars.fit.stepwise. These functions are associated with Fraley et~al. (2007) and will not be undergoing further development except for things like bug fixes. For ongoing development of least-angle regression, see the glars library.

References

B. Efron, T. Hastie, I. Johnstone and R. Tibshirani (2004), "Least Angle Regression" (with discussion), Annals of Statistics 32, 407-499.

C. Fraley and T. Hesterberg (2007), " Least-Angle Regression for Large Datasets", Technical Report, Insightful Corporation.

See Also

qrBlockApply

Examples

data(diabetes)

larFit <- biglars.fit(diabetes$x, diabetes$y, type = "lar")
larFitBlocked <- biglars.fit(diabetes$x, diabetes$y, type = "lar",
                           blockSize = 50)

lassoFit <- biglars.fit(diabetes$x, diabetes$y, type = "lasso")
lassoFitBlocked <- biglars.fit(diabetes$x, diabetes$y, type = "lasso",
                              blockSize = 34)