How to choose a distribution for the likelihood step?

I have been looking over quite a few examples.
I am just wondering how to choose a proper distribution for the likelihood step?

For instance, I am looking at at hierarchical switchpoint analysis, with continuous measurement instead of count data?
Should I be using a pm.Poisson for the likelihood? or should I use a pm.Normal?

If your measurement is continuous on the real domain, pm.Normal usually is a good choice as a first start.

It seems you are confusing a decision fundamental to statistics (“which distribution do I choose for my model?”) with PyMC3’s inner workings somehow, because using hierarchical analysis or not has no bearing on the distribution decision.

I suggest you work through McElreath’s book, which, among other modelling decisions, addresses this question of when to use which distributions.

1 Like