Estimates diverging (missing predictions) error in GEE with log link

[Anseh Danesharasteh]

Dear Stata Members,

I am running a code using GEE with log link option and Gaussian family. I also have several conditions in my code. With some conditions, I receive the error "Estimates diverging (missing predictions)" and when I change the link to "identity", the code runs with no error with the exactly same covariates! I was able solve the error by excluding some of the covariates with log link, however, I would highly prefer to keep them in my model (Journal editors asked us to use a log link model).

Could anyone please help me on this?

[Leonardo Guizzetti]

Sometimes this happens with log-links. You don't say if you're using the Poisson family or Gaussian family, so you could try the canonical pair (Poisson family with log link). Sometimes the choice of working correlation matrix can cause issues (the default being exchangeable). You could try to change it to independent.

[Anseh Danesharasteh]

Thank you for your response, Leonardo.

I was using Gaussian family. I tried with Poisson family, again for some conditions it worked but for some of them I still receiving the same error.

Any other solutions?

[Leonardo Guizzetti]

It's difficult to say with any certainty. The conditions you apply may simply reduce the sample size too much to make the model identifiable

On the one hand, if the model converges, then you can reasonably trust those results. However, if you need to go hunting for the right combination of family, link, correlation matrix, then I would be hesitant to accept that those results are valid. You might just consider reporting to the reviewer that you have tried to fit the model using a theoretically justified model, but it failed to converge and so you sought to simplify the model in some reasonable way, then detail the results and what the model was that produced those results. The key is to be honest and transparent about what you did, while keeping the spirit of the request in mind, but not trying to mislead anyone with those results.

[Anseh Danesharasteh]

Thank you for your response, Leonardo. I also think the same way as you. Originally I ran the model with Gaussian family and identity link options and the results totally made sense. So I think I can explain the issue to them since it is not reasonable to use a different link/family/corr on each condition.