R devel - Sep 2004 - Covergence FLAG in glm (PR#7235)

Full_Name: Daniel R Jeske
Version: 1.8.1
OS: Windows 2000
Submission from: (NULL) (138.23.228.79)


We have just noticed that when you use glm() it seems the logical
output 'converged' is always TRUE.  The same data set that shows
FALSE in version 1.7.1 shows TRUE in 1.8.1.  And I know that FALSE
is the correct answer...so it seems like we cannot trust the 'converged'
flag for glm() in version 1.8.1 ?????

daniel.jeske@ucr.edu wrote:
> Full_Name: Daniel R Jeske
> Version: 1.8.1
> OS: Windows 2000
> Submission from: (NULL) (138.23.228.79)
> 
> 
> We have just noticed that when you use glm() it seems the logical
> output 'converged' is always TRUE.  The same data set that shows
> FALSE in version 1.7.1 shows TRUE in 1.8.1.  And I know that FALSE
> is the correct answer...so it seems like we cannot trust the
'converged'
> flag for glm() in version 1.8.1 ?????
Please try again with a recent alpha version of R-2.0.0 (R-1.8.1 is 
quite outdated).
If it still happens, please specify a reproducible example and explain 
why converged=FALSE is expected.

Uwe Ligges

> ______________________________________________
> R-devel@stat.math.ethz.ch mailing list
> https://stat.ethz.ch/mailman/listinfo/r-devel

>>>>> "daniel" == daniel jeske
<daniel.jeske@ucr.edu>
>>>>>     on Sat, 18 Sep 2004 04:10:40 +0200 (CEST) writes:
    daniel> Full_Name: Daniel R Jeske Version: 1.8.1 OS: Windows
    daniel> 2000 Submission from: (NULL) (138.23.228.79)


    daniel> We have just noticed that when you use glm() it
    daniel> seems the logical output 'converged' is always TRUE.
    daniel> The same data set that shows FALSE in version 1.7.1
    daniel> shows TRUE in 1.8.1.  And I know that FALSE is the
    daniel> correct answer...so it seems like we cannot trust
    daniel> the 'converged' flag for glm() in version 1.8.1
    daniel> ?????

I'm pretty sure you're wrong:

The  NEWS  file for 1.8.0  contains
>>    o	   The defaults for glm.control(epsilon=1e-8, maxit=25) have been
>>         tightened: this will produce more accurate results, slightly
slower.
Hence, compared to 1.7.1, glm() in your (still very outdated!!)
version 1.8.1 will by default use more iterations and may well
converge in cases it didn't in 1.7.1 

If there weren't such cases, the change wouldn't have been
worth!

---

Note that even if you were right, your bug report would be
basically unusable as bug report and we ask to   
PLEASE only send bug reports (to R-bugs@... that is) if you know
how to do it -- i.e. if you have read that section in the R FAQ.

Regards,
Martin Maechler