Clyde,
I can't provide an explanation for these convergence problems.
However, they can be avoided by refitting the binomial model as a
'modified' Poisson regression model, as recommended by Zou
(American Journal of Epidemiology 2004; 159: 702-6), i.e.:
poisson yvar xvar, robust irr
This produces the same risk ratio (and the same standard error
using the robust variance estimator) as the binomial model. I've not
encountered any convergence problems with this modified Poisson
model, even with covariates added to the model.
Roger Webb
University of Manchester (UK)
On 20 Jul 04, at 11:21, Clyde Schechter wrote:
I am running version 8.2 of Stata SE for Windows.
My situation is this: when I fit a logistic model to my data, everything
appears to run well, with convergence achieved after 3 iterations using
-logistic-, -binreg, or- or -glm, fam(binomial) link(logit)-. My client
would prefer to express the associations as relative risks, rather than
odds ratios. When I run the exact same data using -binreg, rr- or -glm,
fam(binomial) link(log)- the iteration log says that the likelihood
function is not concave, and after 50 iterations it gives up with
convergence not achieved.
Looking at the predicted values of p from the logistic regression, they are
all between 0.10 and 0.82, so neither the log nor logit functions exhibits
any particularly pathological behavior in the working range. In fact, over
this range, the relationship between log p and logit p is reasonably close
to linear. I am wondering why the two models would exhibit such different
behavior.
To make matters more interesting, if I precede the whole thing with
-version 6- (this problem actually came from a colleague who is running
version 6), the -glm, fam(binomial) link(log)- run converges after 6
iterations, but -binreg, rr- (which I thought is a wrapper for -glm-) still
fails.
To make matters still more interesting still, some of the independent
variables in the model are dummies created using xi. The results described
above are achieved using -xi i.depvars- as a separate command. If,
however, I run (after -version 6-) -xi: glm i.depvars, fam(binomial)
link(log)-, there is once more failure of convergence. (Without -version
6-, it doesn't matter how -xi- is used, convergence fails either way.)
What is going on here? Is there anything I can do to achieve convergence
under version 8.2 with a log link? And is the output of the convergent
-glm, fam(binomial) link(log)- run credible?
Any explanation is appreciated. Thanks in advance.
Clyde Schechter
Dept. of Family & Social Medicine
Albert Einstein College of Medicine
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
