Hugo,

I’m sorry you’re seeing this error. I’m not sure what to make of it, as it comes from mgcv in R. My guess is that the “Non-finite coefficients at iteration 1” is what ultimately causes the “NAs in d(mu)/d(eta)”. I would start by looking at why “Non-finite coefficients at iteration 1” might be reported. Perhaps some predictor variables have extreme values, or the model formula includes transformations that cause them to take on extreme values.

If you privately send me the complete output of the tool and a copy of your input table, I will try it myself and see if I can develop a better guess.

There is some possibility that we could eventually implement negbin for GLM. The reason we have not already is that the standard R glm package does not support negbin. My understanding is that the traditional way to fit GLMs with negbin is to use the glm.nb function from the R MASS package. Thus we’d have to code MGET to use MASS for negbin distributions, and the glm package for all other distributions. That is probably doable but we’d have to be careful that the MASS package was interchangeable with the glm package in the ways that MGET currently uses the glm package. That requires some work. Because we’ve never had a need for negbin GLMs, we’ve never bothered to do it. If this is a critical need for you, we might be able to do it sometime in the next few months.

Best regards,

Jason