bigRfit_biglm_xc: Big data rank-based estimation based on centered design...

In kloke/bigRfit: Big data version of Rfit

Big data rank-based estimation based on centered design matrix.

Description

Fits the regression coeficients for the linear regression model with a centered design matrix. Algorithm designed for big data sets which uses a partial ranking and step scores.

Usage

bigRfit_biglm_xc(formula, data, intercept = FALSE, yhat0 = NULL, ehat0 = NULL, B = 1000, scores = wscores, max.iter = 50, eps = sqrt(.Machine$double.eps), TAU = "DT", ...)

Arguments

formula	an object of class formula
data	data frame
intercept	indicator to request estimate of alpha (FALSE by default)
yhat0	optional initial estimate of responses
ehat0	optional initial estimate of residuals
B	number of bins to use (default of 1000)
scores	an object of class 'scores' (wscores by default)
max.iter	maximum number of iterations
eps	specify tol
TAU	version of estimation routine for scale parameter. DT for approximate binned estimate (using data.table), F0 for Fortran using full set of residuals, N for none
...	Additional arguments.

Details

Given a formula y ~ x fits the model y ~ xc where xc is the centered x. Estimation of an intercept is not needed. However, an estimate of alpha in the model y = alpha 1 + xc beta + e may be requested using the option intercept=TRUE.

Rank regression estimates beta by minimizing the dispersion functions: D(beta) = a(R(y-xc beta))^T (y-xc beta) where R denotes rank and a is a non-decreasing score function.

In this implementation, instead of using a full ranking, the observations are binned into bins of approximate equal size. The parameter B specifies the number of bins to use. B must be <= the number of observations. Step scores are used so that observations within the same bin are assigned the same score.

This function is (mostly) intended to be an internal function. See bigRreg for general purpose regression.

Author(s)

John Kloke <[e-mail]>

See Also

bigRreg, ~~~

kloke/bigRfit documentation built on April 20, 2023, 4:33 p.m.