Dear all, I’ve been trying to track down the following error message: “Exception: bernoulli_logit_lpmf: Logit transformed probability parameter[2815] is nan, but must not be nan! for a model whose code is shown in the end of the post (because maybe not relevant to the question).” To do this, I p…

So the parameters by default are initialized -2 to 2 on the untransformed space. The initialization really doesn’t take into account the priors, so the very first numbers are just totally random (priors aren’t special in any way in Stan – just more terms for the log density). Since you’ve got the > …

Presumably the problem is with initialization as Ben’s suggesting. All these powers can be problematic with random inits. You can either try to rescale the parameters, or you can try initializing in a smaller range than (-2, 2) on the unconstrained scale.

So if my choice to get the model to sample ok is to either use inits that are close to the true values or use hard constraints on the parameters, which one is less likely to bias my results? With hard constraints it now works well and parameters are not estimated too close to the hard boundary. And …

[image] Jacquie: after ~ 4 days it only has <1000 samples per chain Yikes! Is it possible that ntr is much greater than nsub*maxSess? (i.e. replicated observations of the outcome for each combination of subject & session) If so, a trick to reduce computation is to compute only the unique rows…

I think there’s a problem with: for (is in 1:nsub) { for (isess in 1:maxSess) { noiseP_sess_pr[is] ~ normal(0,1); // to give: normal(noiseP_subj_pr[is], p_subj_sd[is,1]); } } I think either the inner index needs to be [is, isess] or you can get rid of the loops entirely and do: to_array_1…

[image] bbbales2: I think there’s a problem with: thank you so much for spotting this!! i’ve looked so many times through this code and didn’t see it, so thank you very very much! let’s see whether this goes any better now…

[image] mike-lawrence: it possible that ntr is much greater than nsub*maxSess? (i.e. replicated observations of the outcome for each combination of subject & session) If so, a trick to reduce computation is to compute only the unique rows of the design matrix then indexing into this when express…

ok, this now actually works :) but it still does not converge when I add in more parameters. I am wondering whether the following features of my model may be causing problems: per session (lowest level in hierarchy), I only have 20 measurements, and I am trying to fit 4 parameters. Maybe this is …

per session (lowest level in hierarchy), I only have 20 measurements, and I am trying to fit 4 parameters. Maybe this is not a good idea? 20 isn’t a lot for sure haha. Since it’s 4 parameters, start by looking at the pairplots of all these and see if you find correlations n’ such. Do you have pos…

[image] bbbales2: 20 isn’t a lot for sure haha. Since it’s 4 parameters, start by looking at the pairplots of all these and see if you find correlations n’ such. Do you have posterior predictives for this model? Do those remain good? thanks for the quick reply. just to check, to look at pair …

Parameters during warmup far outside prior range

Modeling

mike-lawrence February 26, 2018, 9:03pm 17

A demo of what I mean here: Tip for speeding up models with repeated observations

Topic		Replies	Views
Divergent transitions & BFMI low in a state-space model Modeling	6	1125	August 17, 2017
NLME model in STAN for missing + unbalanced longitudinal data Modeling rstan , mixed-model , missing-data	20	752	October 21, 2024
Issues with running Bayesian hierarchical reinforcement learning model in Stan Modeling rstan , fitting-issues , specification , performance	12	1012	June 28, 2021
Partial Pooling Scale Parameter is Inf Modeling	9	673	March 9, 2021
Divergences in a non-centered computational model Modeling fitting-issues	21	1721	October 30, 2019

Parameters during warmup far outside prior range

Related topics