lol (version 1.20.0)

lasso.stability: Stability and randomised lasso

Description

point-wise controled lasso stability selection

Usage

lasso.stability(y, x=NULL, alpha=.5, subsampling=.5, nSubsampling=200, model='linear', pi_th=.6,  alpha.fwer=1, lambda1=NULL, steps=10, track=FALSE,  standardize=FALSE,  ...)

Arguments

y

A vector of gene expression of a probe, or a list object if x is NULL. In the latter case y should a list of two components y and x, y is a vector of expression and x is a matrix containing copy number variables

x

Either a matrix containing CN variables or NULL

alpha

weakness parameter: control the shrinkage of regulators, if alpha = 1 then no randomisation, if NULL then a randomly generated vector is used

subsampling

fraction of samples to use in the sampling process, default to 0.5

nSubsampling

The number of subsampling to do, default to 200

model

which model to use, one of "cox", "logistic", "linear", or "poisson". Default to 'linear'

pi_th

The threshold of the stability probablity for selecting a regulator. It is to determine whether a coefficient is non-zero based on the frequency it is subsampled to be non-zero, default to 0.6

alpha.fwer

Parameter to control for the FWER, choosing alpha.fwer and alpha control the E(V), V being the number of noise variables, eg. when alpha=0.9, alpha.fwer = 1 control the E(V)

lambda1

minimum lambda to use

steps

parameter to be passed on to penalized

track

track the progress, 0 none tracking, 1 minimum amount of information and 2 full information

standardize

standardize the data or not?

...

Value

A list object of class 'lol', consisting of:

beta

coefficients

beta.bin

binary beta vector as thresholded by pi_th

mat

the sampling matrix, each column is the result of one sampling

residuals

residuals of regression model

Details

The function first selects lambda that approximately give maximum sqrt(.8*p) predictors, while p is the number of total predictors. Then it runs lasso a number of times keeping lambda fixed. These runs are randomised with scaled predictors and subsamples. At the end, the non-zero coefficients are determined by their frequencies of selections.

References

N. Meinshausen and P. Buehlmann (2010), Stability Selection (with discussion), Journal of the Royal Statistical Society, Series B, 72, 417-473.

See Also

lasso

Examples

data(chin07)
data <- list(y=chin07$ge[1,], x=t(chin07$cn))
res <- lasso.stability(data, nSubsampling=50)
res