Meaning of "max_treedepth exceeded"

Hi – a beta-binomial model (run using brms) just completed fitting (after 92 hours). At the end I get the warning “There were 81 transitions after warmup that exceeded the maximum treedepth. Increase max_treedepth above 12. See” (everything else about the fit is good: sensible estimates, no divergences, good ESS, Rhat, etc).

On the linked webpage it says “While divergent transitions are a validity concern, hitting the maximum treedepth is an efficiency concern.” I don’t understand what that means. Can someone help me to answer to this simple question:

Can I trust my model’s results or do I need to run it again with a higher max_treedepth?

Most likely they are fine, unless there is other evidence to the contrary. If you have time, maybe run it again with a higher max_treedepth, but usually something as simple as a beta-binomial would not encounter such problems. So, my guess is there is a lot of additional complication that you didn’t mention.

Thanks. I don’t think there are any additional complications. In summary, there are about 50000 observations, 25 predictors, the single most important of which varies (as a group-level effect) by participant (111 levels) and by “item” (91 levels). All other diagnostics look good. The requirement for a high max_treedepth seems to arise from using beta-binomial rather than binomial (LOO shows the former family is distinctly better).

Two levels of a non-nested hierarchy is a complication, but 5% hitting the max_treedepth is probably OK.