Model the precision parameter with predictors?

[JockLawrie]

Hi there,
Is there an intention to enable the precision parameter to be modelled with predictors?
Cheers

[ararslan]

Hi @JockLawrie! It's not implemented currently but it's generally on my radar. There are some minor changes on main that haven't yet made it into a release that begin to lay some of the groundwork for that. Would you consider variable precision models to be a critical feature?

[JockLawrie]

Thanks for the quick reply Alex, much appreciated.

Not a critical feature at this point. I'm working on a project where this feature may be required. If it turns out that way it sounds like implementing the feature in this package is a better option than starting a new package. I'll be in touch when I know more, which I anticipate will be some time in Sep or Oct.

[JockLawrie]

Hi again, a year later than expected. Turns out that the project I'm on indeed requires modelling the precision parameter with covariates.

I've recently updated MultinomialRegression.jl to use a cyclic block-wise coordinate descent alogrithm, together with a basic line search. The algorithm's implementation is about 100 lines and takes as arguments the objective function, a gradient function and a hessian function. For beta regression we could use this algorithm with 2 blocks of parameters, 1 each for the mean and precision. Note that the 2 sets of predictors needn't be the same, though they can be. Thoughts?

[JockLawrie]

And here's an implementation. Note that this repo can cover many model types, including those in GLM.jl and MultinomialRegression.jl. The tests include an example of the latter.