Hi, Ben,
Thanks for your reply, that’s pretty useful!
I tried to experiment with a bunch of other priors, which helped a little with the divergences (but not too much) and also changed the link function, which decreased their number from around 2000 to around 100, so I guess that was a good move in this case.
I also had a look at Jonah’s paper on visualization and changed my priors a little, based on the pairs() plot, which helped some more.
Interestingly enough, regardless of the changes in the priors and the link I made, every time I run the whole model and just sample from the posterior, I don’t run into any divergences. What’s more, my posterior predictive checks seem very reasonable and the predictions I get are reasonable enough. I even kept a small test set, and my model performed reasonably well on the unseen values as well. So I guess in this case the emphasis falls on the data and not the weak priors?
I also tried running a simpler model (same formula but less random effects) and still ran into the same issue - prior divergences, but no problems when sampling from the posterior.
I am really confused as to what to do next. I ran out of ideas of things to try out. And also am still not completely sure where the problem is.
Do you have any suggestions of things I could potentially look into?
Thanks again!