Checking for Convergence / Handling Divergences

[PK1706]

Hello,
since PyMC-Marketing is my first step into bayesian statistics, deciding whether my MMM is trustworthy / converged is quite hard to do.
While the model results look quite promising and not to unrealistic I almost always get around 2 to 20 divergences (target accept already being set to 0.99). In what way should I respond to this? Use the model regardless or try to improve the model?
I have read Bayesian Analysis with Python by Osvaldo Martin as well as countless other articles, but am still a bit lost.
What is often mentioned is that a reparametrization of the model can help with convergence, this is however 1) not trivial, and 2) not really possible when using PyMC-Marketing, where the model itself is wrapped and not accessible by the user.

I have tried changing priors (although I have read this isn't really best practice) to combat this, but didn't get satisfactory results.
The only thing that did reduce and in some runs eliminate divergences was using a student-t likelihood. I am however unsure about how to choose the right prior for the normality parameter.

Are there any solutions to reducing divergences that I am missing or are divergences not as big of a deal as some make it seem?
I would highly appreciate some advice on this and possibly a rule of thumb step-by-step approach.
Thanks a lot!

[AlfredoJF]

Hi @PK1706

Have a look at these Prior Choice Recommendations and see if you can decrease the number of divergences by selecting better priors for your use case.

Also, try to increase the default params chains, draws and tune in the fit method.

mmm.fit(X=X_train,
        y=y_train,
        target_accept=0.95,
        chains=6,
        random_seed=123,
        draws=3_000, 
        tune=2_000, 
        )

[PK1706]

Thanks, tuning prior distributions a bit already reduced divergences in most cases!