This is the mail archive of the
gsl-discuss@sourceware.org
mailing list for the GSL project.
Re: [Bug-gsl] Covariance estimate in weighted regression
- From: Brian Gough <bjg at network-theory dot co dot uk>
- To: Alex Tartakovsky <tartak62 at yahoo dot com>, Giulio Bottazzi <giulio dot bottazzi at libero dot it>
- Cc: GSL Bug List <bug-gsl at gnu dot org>, gsl-discuss at sources dot redhat dot com
- Date: Fri, 17 Feb 2006 16:20:19 +0000
- Subject: Re: [Bug-gsl] Covariance estimate in weighted regression
- References: <20050821010321.56bc297e.giulio.bottazzi@libero.it><17163.13001.143071.36081@network-theory.co.uk><20050823192510.06e08ddc.giulio.bottazzi@libero.it><20051013074955.65975.qmail@web53805.mail.yahoo.com>
Alex Tartakovsky writes:
> >From GSL manual (pp. 361-362), standard texts, and just common
> >sense, one expects that the output produced by a weighted
> >regression with all the weights set to 1 should be the same as
> >from unweighted regression. This is not the case for the
> >covariance estimates produced by "fit" and "multifit"
> >least-squares GSL functions. The reason is that the cov estimates
> >in the straight versions of the functions include s2 (an estimate
> >of the error variance):
Giulio Bottazzi writes:
> Probably it would help to explicitly mention the formula used to
> compute variance-covariance matrix in ALL the routines, so that the
> average dumb user (like me), by comparing the different formulas,
> can immediatly understand were differences can possibly arise. What
> do you think?
Thanks for the comments, I have added some longer explanations in the
manual about how the covariance matrices are computed and their
definitions for the different cases. The new chapters are available
at http://www.network-theory.co.uk/download/gsl/newchaps.ps.gz
--
Brian Gough
Network Theory Ltd,
Publishing Free Software Manuals --- http://www.network-theory.co.uk/