Hi all, I’m fitting a big hierarchical nonlinear model, on ~18k points and a bunch of groups: about 12k parameters, and p_loo = 1879. It seems to be doing a great job: my fits look excellent, and the parameters seem reasonable. I’m busy working on simulated data to test it now, but based on the fit…

I’ve seen this behaviour in models where I’ve forgotten to put priors on some parameters, and the default uniform priors give some tricky posterior geometry that the sampler struggles with. However this can come from a lot of different areas. If you can post your model code we’ll be able to provide…

Sorry for taking so long to get back to you both! Haven’t had a second free the last couple of days. Thank you so much for the suggestions! @spinkney : Looking into the QR decomposition, it sounds like this could be a really amazing quick fix! I’m using brms actually, and I see that there’s an opt…

Update on the QR decomposition: I started it running it a hurry as I didn’t have much time last week. I just used the decomp = "QR" input argument in brms, and set it into motion. Seeing that it was not going any quicker than usual during the weekend, I looked at the STAN code, and I see that it w…

I tried a simpler model (too simple for eventual purposes, but hopefully helpful for debugging), and I’m now only saturating the treedepth in 16.5% of the iterations after removing the random effects for (1|WAYID:Region), which is about 3600 (highly-constrained) parameters. So removing this reduces…

Have you checked Rhats and effective sample sizes for hints along which parameters the sampler has problems moving fast enough (mixing)? It might be helpful to check that also in the unconstrained space.

Thank you so much @avehtari : this is extremely helpful advice! I hadn’t thought that looking into the Rhats and ESS might also reveal information about which parameters to look into for treedepth issues, but I guess that with not being able to effectively explore the posterior, it’s more likely th…

Exceeding treedepth doesn’t make the sample invalid, but indicates inefficiency and often also problems in the model. Your model is complex and certainly has difficult posterior and also looking at the posterior intervals some parameters are not well informed by data. [image] mathesong: All my…

Thanks so much for the suggestions, and the followup! I can now update (I accidentally ran the first model with all the iterations again, and noticed too late to stop it). Tightened up priors over intercept terms Firstly, I tried tightening up my priors over the intercept terms. This didn’t seem …

Maybe @paul.buerkner could comment whether he knows that this brms would be prone to problems? Family: gaussian Links: mu = identity; sigma = log Formula: TAC ~ twotcm_log_stan(logK1, logVnd, logBPnd, logk4, logvB, t_tac, t0, slope, alpha, beta, gamma, A, B, C, peaktime, meanblood) si…

Max_treedepth saturated, but increasing it slows sampling to a standstill

Modeling

spinkney November 18, 2020, 10:34am 2

Check out (Another) Slow Hierarchical Model in Stan

Topic		Replies	Views
Setting Max Treedepth in difficult high-dimensional models Modeling	13	4356	August 6, 2017
How to speed up sampling in rstan? Modeling rstan , performance , hierarchical-model	6	6251	July 24, 2020
100% exceeding max treedepth Modeling fitting-issues , specification	6	568	August 22, 2022
Max treedepth warnings in hierarchical multivariate model Modeling	10	2007	November 17, 2018
Suggestions for improving model of bird monitoring data Modeling performance	11	1063	November 17, 2020

Max_treedepth saturated, but increasing it slows sampling to a standstill

Related topics