Message-ID: <Pine.LNX.4.61.0504181755560.26699@gannet.stats>
Date: 2005-04-18T16:59:05Z
From: Brian Ripley
Subject: refitting lm() with same x, different y
In-Reply-To: <Pine.LNX.4.62.0504181657180.7778@octopus.well.ox.ac.uk>
William,
As a first shot, use lm with a matrix response. That fits them all at
once with one QR-decomposition. No analogue for glm or lmer, though,
since for those the iterative fits run do depend on the response.
Brian
On Mon, 18 Apr 2005, William Valdar wrote:
> Dear All,
>
> Is there is a fast way of refitting lm() when the design matrix stays
> constant but the response is different? For example,
>
> y1 ~ X
> y2 ~ X
> y3 ~ X
> ...etc.
>
> where y1 is the 1st instance of the response vector. Calling lm() every time
> seems rather wasteful since the QR-decomposition of X needs to be calculated
> only once. It would be nice if qr() was called only once and then the same
> QR-factorization used in all subsequent fits. However, I can't see a way to
> do this easily. Can anybody else?
>
> Why do I want to do this? I'm fitting ~1000 different X's to a response
> vector (for biologists: 1000 genetic markers to a measured phenotype with
> 2000 cases) and wish to establish global significance thresholds for multiple
> testing. The fits have a complex dependency structure that makes the
> Bonferroni correction inappropriate. So I intend to refit all ~1000 X's with
> a shuffled response many times. However, this runs too slow for my needs.
>
> Of course, not having to redo QR will only help if QR is a rate limiting step
> in lm(), so if anybody can tell me it's not, then that would be very helpful
> too. I would also like to do this for glm() and lmer() fits. Ideally.
--
Brian D. Ripley, ripley at stats.ox.ac.uk
Professor of Applied Statistics, http://www.stats.ox.ac.uk/~ripley/
University of Oxford, Tel: +44 1865 272861 (self)
1 South Parks Road, +44 1865 272866 (PA)
Oxford OX1 3TG, UK Fax: +44 1865 272595