s-news
[Top] [All Lists]

Re: [S] model selection in glm - logit link vs identity

To: John Maindonald <john.maindonald@anu.edu.au>
Subject: Re: [S] model selection in glm - logit link vs identity
From: Prof Brian D Ripley <ripley@stats.ox.ac.uk>
Date: Wed, 27 Jan 1999 22:19:45 +0000 (GMT)
Cc: s-news@wubios.wustl.edu
In-reply-to: <199901272154.IAA01409@leonard.anu.edu.au>
Sender: owner-s-news@wubios.wustl.edu
On Thu, 28 Jan 1999, John Maindonald wrote:

> 
> On Wed, 27 Jan 1999, Brian Cade wrote:
> >      (2)
> >      An interesting issue occurred with estimating the logit link model in 
> >      3 different packages, S-Plus 4.5, SAS, SYSTAT 7; different parameter 
> >      estimates or standard errors.  Results below.  Note that deviances 
> >      (RSS = 16.5365) and estimates of dispersion (MSE = 0.1323) were the 
> >      same for all packages.  Anyone have any comment on what S-Plus is 
> >      doing differently than SAS and SYSTAT?
> >      
> >      S-Plus                      SAS Insight                   SYSTAT 7 
> >      glm( family=quasi(link         glm (quasi,link=logit,      Nonlinear
> >       =logit,variance=const)            (variance=normal,        model     
> >                                          deviance)            (loss = RSS,
> >                                                               Gauss-Newton)
> >      
> >      b0(se)  -0.098 (6.707)        -0.383 (8.295)           -0.383 (6.767)  
> >      
> >      b1(se)   0.002 (0.003)         0.002 (0.003)            0.002 (0.003)  
> >      
> >      b2(se)  -0.003 (0.002)        -0.003 (0.002)           -0.003 (0.002)  
> >      
> >      b3(se)   0.557 (0.208)         0.567 (0.262)            0.567 (0.211)
> >      
> >      b4(se)  -0.033 (0.009)        -0.034 (0.011)           -0.034 (0.009)
> >      
> >      SYSTAT nonlinear least squares parameter estimates are similar to 
> >      those from SAS glm but SYSTAT standard errors are more similar to 
> >      S-plus glm standard errors.  Great discrepancy between SAS/SYSTAT b0 
> >      estimate and that of S-Plus is disturbing.
> 
> Brian Ripley responded
> 
> >Have these converged? Try changing the control tolerances to be
> >sure. Often the likelihood surface is very flat if you have highly
> >related or ineffective carriers, and the convergence criterion in
> >glm() is rather lax.  Actually, I don't see that a difference of less
> >than 5% of an s.e. should disturb you.
> 
> One does have to worry about the difference in b0.  How do the fitted
> values for the three fits compare?  I doubt that this is a convergence
> issue, because it ought to affect other parameters also.  What are the

Why, please? The likelihood is clearly a great deal flatter in that one
direction: just look at the size of the standard errors. I just cannot
understand how you can know that the other coefficients `ought' to be
affected, so could you please point us to the relative theory that ensures
this.

I _have_ seen precisely this in a logistic regression due to lack of
convergence, and changing the convergence criteria did give much
closer agreement.

Please explain how I should act differently if two fits with a very 
similar RSS have fitted coefficients varying by a few percent of a
standard error. What is the practical significance?

-- 
Brian D. Ripley,                  ripley@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 272860 (secr)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595

-----------------------------------------------------------------------
This message was distributed by s-news@wubios.wustl.edu.  To unsubscribe
send e-mail to s-news-request@wubios.wustl.edu with the BODY of the
message:  unsubscribe s-news

<Prev in Thread] Current Thread [Next in Thread>