glmnet

Hi Andra.

I wonder how you come about trying to use LASSO without knowing what lambda
is. I'd advise you to read up on it. In the help (?glmnet) you can find
several paper references, but for a more gentle introduction, you can read
http://www-stat.stanford.edu/~tibs/ElemStatLearn/

In a nutshell, though: lambda is the parameter that balances the weight
given to the penalty. The bigger this one is, the more 'pressure' there is
on the coefficients to be small (or better yet: disappear).
The way you use LASSO is: you look at a reasonable set of lambda values
(this is e.g. done by glmnet), calculate some measure of success with each
lambda value (e.g.: misclassification, AUC,...), generally by using
crossvalidation (as is provided by cv.glmnet: read its help).

Having this measure of success (say the AUC) for each lambda in your
reasonable set allows you to pick the most optimal (lambda.min) or, to avoid
happenstance peaks, a more conservative and parsimonious one (lambda.1se),
after which you can rerun your lasso with this selected lambda on the full
dataset, to find the variables in your model.

Finally, to avoid downward bias, you could run a normal glm with only the
variables selected in the previous step.

Good luck!


Nick Sabbe
--
ping: nick.sabbe at ugent.be
link: http://biomath.ugent.be
wink: A1.056, Coupure Links 653, 9000 Gent
ring: 09/264.59.36

-- Do Not Disapprove

-----Original Message-----
From: r-help-bounces at r-project.org [mailto:r-help-bounces at r-
project.org] On Behalf Of Andra Isan
Sent: woensdag 10 augustus 2011 5:59
To: r-help at r-project.org
Subject: [R] glmnet

Hi All,
I have been trying to use glmnet package to do LASSO linear regression.
my x data is a matrix n_row by n_col and y is a vector of size n_row
corresponding to the vector data. The number of n_col is much more
larger than the number of n_row. I do the following:
fits = glmnet(x, y, family="multinomial")I have been following this
article: http://cran.r-project.org/web/packages/glmnet/glmnet.pdfpage
8, but there are some unclear parts that I dont understand. The lambda
variable only returns 100 and I exactly dont know what lambda
represents. So, basically I would like to know how to get the
coefficients weights and what exactly lambda is? how I can see the
difference between predicted values and observed values?
If there is a sample code that helps me to understand how to use these,
that would be great.
Thanks a lot,Andra

	[[alternative HTML version deleted]]

Finally, to avoid downward bias, you could run a normal glm with only the
variables selected in the previous step.

At the cost, of course, of introducing upward bias....

Patrick Breheny
Assistant Professor
Department of Biostatistics
Department of Statistics
University of Kentucky