There’s a little bit of a funnel in the intercept, but nothing looks too crazy to me:
But re-parameterizing the intercept made a huge difference:
135.169487 seconds
NUTS slowdown factor: 189.685194039
effectively 10x faster. Increasing to 1000 observations shows a very nice scaling:
173.795094013 seconds
NUTS slowdown factor: 52.3025973161
subsequently increasing to 40 covariates (1000 x 40 data matrix) looks even better
174.969336987 seconds
NUTS slowdown factor: 33.5733716568
so it looks like it’s scaling really well. Thank you.