Read on Twitter
Tricky Bayesian #rstats #brms methods question! I'm building a predictive model for a strangely-distributed outcome—it's a count with a bunch of 1s, a bunch of 32s (the max) and some numbers in between. How do you model this weird thing?! (reprex here: gist.github.com/andrewheiss/33…) 
- I can subtract 1 from the outcome and use zero-inflated poisson to pick up the excess 1s, but there's no such thing as "1-and-32-inflated poisson" like zero-one-inflated beta, and the pp_check is awful
- I can use a mixture of two poisson models, but pp_check is bad+ESS bad
- I can use a mixture of two poisson models, but pp_check is bad+ESS bad
- I can collapse all the 2–31 counts into a category and just do ordered logit with 1, 2–31, and 32 as categories, but collapsing like that feels icky
idk :(
Treating the outcome as a proportion (it's censored at 32 because it's actually a count of provinces each row works in and there are only 32 provinces in the country so 100% = 32) means I can use zero-one-inflated beta regression without any custom families and it works!
• • •
Missing some Tweet in this thread? You can try to
force a refresh
