[moving to list]
Hi, see below.
AThe problem with documenting all this is that it changes
the posterior.It doesn’t change the posterior for any predictions, it just changes the labeling of the modes.
That’s not what we concluded last time. You’re the one who
convinced me the posteriors wouldn’t be the same even without
label switching.
So I’d like to run some simulations to see if the results really
are the same. Then I can talk about both models. I find it’s
easier that way to see what’s going on than just asking people
to do what we do without understanding why. Yes, I can lead
with what you consider the “good” model — I know that’s important
to you.
I really think we shouldn’t have the mixture model as it is in the manual right now, because fitting this model gives nonsensical inferences for the individual parameters. I’m worried that people will fit this model as written and get bad results. But, again, ti’s not urgent in that if it’s a problem, it’s been a problem for several months…
It’s been a problem for years and it keeps popping up on the list.
I added a new discussion about inferences that were supported even
under label switching. But you lose R-hat as a convergence diagnostic.
- Bob