cox-regression SPSS 20 "no-events-problem"

[arvid]

Hi,

I'm new the forum and hoping that I'm not being offensive with a stupid question!

I'm trying to do a cox-regression on a data set with patients treated for meningioma (a kind of brain tumor). I have varying follow-up times after surgery (time variable), whether the tumor recurred (status variable) and a number of predictors. My sample size is about 40 patients, nicely distributed over categorical variables (treatment, gender)

When I run the analysis I get the error message "since coefficients did not converge, no further model will be fitted."

I think the problem is that the most important predictor (treatment strategy) has no events in the new treatment strategy. As a clinician it's a huge success because the recurrence rate in the patients receiving standard treatment is quite high, so we're obviously doing the right thing, but I can't get the cox regression to work..

If I enter one event in the treatment group that has no events, the analysis works fine..

I'm suspecting that this outcome is inherently logical from the way the cox regression works, but still curious if there's anything I could do - I mean, there's an obvious difference between the groups! Wait for an event in the treatment group? :)

Any suggestions greatly appreciated!!

Arvid

[Andy W]

Hopefully you get more actionable advice than this, but I know this is referred to as perfect seperation when it occurs in logistic regression. Some quick googling brings up this citation for Cox regression, http://www.jstor.org/stable/2676848, and the same author has SAS and SPlus functions to estimate the models, http://cemsiis.meduniwien.ac.at/fileadmin/msi_akim/CeMSIIS/KB/volltexte/Heinze_Ploner_2002_CompMethProgBiom.pdf.

Congrats on the successful experiment, and hopefully your reviewers are not so curmudgeonly to accept such obvious results at face value. I was thinking you could just plot the confidence interval for the predicted hazard function for the non-treated group to show the variability in the estimates, but others on this list-serve can probably give you better advice than that.

[David Marso]

Perusing the paper Andy cites leads to initial optimism.

"The core routines of both programs reside in
a dynamic link library (DLL) and can thus also
be accessed from other software packages. "

YAY!  Bring on the snakeheads !

This hope soon dispelled by the apparent unavailability of the URL cited in the paper ;-(
Might just be my system so give it a whirl.
--
Technical Report [4] which is available along with the programs at the WWW location
http://www.akh-wien.ac.at/imc/biometrie/fc

Blah blah blah...
I hate it when that happens ;-((
---

Andy W wrote

Hopefully you get more actionable advice than this, but I know this is referred to as perfect seperation when it occurs in logistic regression. Some quick googling brings up this citation for Cox regression, http://www.jstor.org/stable/2676848, and the same author has SAS and SPlus functions to estimate the models, http://cemsiis.meduniwien.ac.at/fileadmin/msi_akim/CeMSIIS/KB/volltexte/Heinze_Ploner_2002_CompMethProgBiom.pdf.

Congrats on the successful experiment, and hopefully your reviewers are not so curmudgeonly to accept such obvious results at face value. I was thinking you could just plot the confidence interval for the predicted hazard function for the non-treated group to show the variability in the estimates, but others on this list-serve can probably give you better advice than that.

[arvid]

Thanks for quick and informative responses!
And thanks Andy for pointing out the similarity to separation as in logistic regression, though I'm thinking that this is a case of quasi-complete separation rather than complete separation since there are non-treated patients with and without recurrence, but no treated cases with recurrence. So the only cell with zero-frequency would be the "treated-with-recurrence" - am I right?

I'll try and email the corresponding author of the paper and see if I can get the functions.. Wish me luck :)