Thanks for the insight. I’ll try thinning first, as you suggested.
As for marginalizing out K: do you think this is a problem with this particular model, or the implementation of MvHypergeometric in general? In other words, if thinning is unfeasible (due to the number of samples required), should I try to implement the distribution in such a way that it automatically marginalizes out K?
I’m asking this because I plan on submitting this distribution (as well as the bivariate Hypergeometric) as a PR on GitHub. If this issue with the high autocorrelation might show up on any model that uses this distribution, maybe it’s worth it to implement the distribution differently, yes?