R help - Mar 2012 - glm.fit: fitted probabilities numerically 0 or 1 occurred?

Hi all,

I am doing bootstrap with logistic regression by using glm function, and I
get the errors;

glm.fit: fitted probabilities numerically 0 or 1 occurred and
glm.fit: algorithm did not converge

I have read some things about this issue in the mailing list. I can guess
what was the problem. My data contains one or may be two outliers. Does the
error occur due to these extreme values or something else such as MLE? Is
there any way to to fix this problem? 

Regards,

Ufuk

-----
Best regards

Ufuk
--
View this message in context:
http://r.789695.n4.nabble.com/glm-fit-fitted-probabilities-numerically-0-or-1-occurred-tp4490722p4490722.html
Sent from the R help mailing list archive at Nabble.com.

> I get the errors;
> glm.fit: fitted probabilities numerically 0 or 1 occurred and
> glm.fit: algorithm did not converge
> .....
> Is there any way to to 
> fix this problem? 
There are two separate issues.
One is the appearance of fitted values at 0 or 1. 
The other is the lack of convergence.

The first is usually not fatal; it means that the probabilities are so close to
0 or 1 that a double precision value can't distinguish them from 0 or 1.
Often that's a transient condition during iteration and the final fitted
values are inside (0,1), but final fitted values can also be out there if you
have continuous predictor values a long way out;  by itself, that usually
won't stop a glm.

The second is a bit more problematic. Sometimes it's just that you need to
increase the maximum number of iterations (see the control= argument and
?glm.control). That is always worth a try - use some absurdly high number like
1000 instead of the default 25 and go find some coffee while it thinks about it.
If that solves your problem then you're OK, or at least as OK as you can be
with a data set that hard to fit.
But if you're bootstrapping with some anomalous values it is also possible
that some of your bootstrapped sets have too high a proportion of anomalies, and
under those conditions it's possible that there could be no sensible optimum
within reach.

One way of dealing with that in a boostrap or other simulation context is to
check the 'converged' value and if it's FALSE, return an NA for your
statistic. But of course that is a form of censoring; if you have a high
proportion of such instances you'd be on very thin ice drawing conclusions.

S Ellison

*******************************************************************
This email and any attachments are confidential. Any use...{{dropped:8}}

Dear Ellison, 
Many thanks for your reply.

The information you typed is clear and now I know what to do.  Your
suggestion about finding some coffee while running simulation is so good =)

Regards

-----
Best regards

Ufuk
--
View this message in context:
http://r.789695.n4.nabble.com/glm-fit-fitted-probabilities-numerically-0-or-1-occurred-tp4490722p4492436.html
Sent from the R help mailing list archive at Nabble.com.