Please see if you know the answers to four questions concerning the problems described below. I followed @bgoodri ’s suggestion in this post to turn centered parametrization to non-centered parametrization of correlated parameters. This significantly improve the accuracy of the parameter estimates b…

[image] tlyim: And if yes, how to deal with it? How many parameters do you have total?

Thanks for asking. Is it the exact number of the original parameters that you want to know? Perhaps, I should outline the structure of the model components. Those shown in the estimation table in my post are all the original parameters: (i) for the component of the state-space model of a latent AR…

[image] bbbales2: You can probably make that lower and it’ll still work. 0.95 is sufficient to avoid any divergence and get reasonable estimates of the parameters with the following setup but the downside is that it took over 7 hours (for a fast machine on a cluster) to complete fitting the m…

So it looks like this model has 20 some parameters? Maybe try adding control(metric = “dense_e”) and see if that helps? Also, are your lowest Neffs 2000~? You might consider just running your calculations a shorter amount of time if so. Can you run get_num_leapfrog_per_iteration(fit) on your fit? …

[image] bbbales2: So it looks like this model has 20 some parameters? Maybe try adding control(metric = “dense_e”) and see if that helps? Thanks for suggesting the metric = "dense_e" option. Tried but it doesn’t seem to work in this case: still at the very early stage of warmup after 15 hour…

[image] bbbales2: Also, are your lowest Neffs 2000~? You might consider just running your calculations a shorter amount of time if so. Can you run get_num_leapfrog_per_iteration(fit) on your fit? Thanks for the suggestions. Tried 1500 iterations (with 1000 warmups for 4 chains for simulated …

is a fitting time of 5.5 hours considered uncommonly long? It’s more about the number of gradient evaluations than total walltime. It looks like your chains are taking 255+ leapfrog steps before hitting a U-turn. That is a lot. If you have a relatively modest number of parameters (is it actually…

[image] bbbales2: If you have a relatively modest number of parameters (is it actually just 20?) originally 20 but after non-centered reparametrization of the correlated parameters, the total number of raw parameters to sample becomes 78. [image] bbbales2: to dump data constructed for…

Oh whoops, my bad, that last argument should be named: stan_rdump(names(data), file = "mydata.dat", envir = list2env(data)) And this is assume that data is a list like: data = list(N = ..., y = ..., ...) (The thing you’d pass into a Stan sampling statement). 78 parameter…

Problems remain after non-centered parametrization of correlated parameters

Modeling

Charles_Driver May 29, 2019, 4:14pm 11

I don’t think the fix for the dense metric has made it into the rstan on CRAN yet…

Linear Regression with Highly Correlated Covariates

Topic		Replies	Views
Adaptation and non-centered parameterisations Developers	10	1349	August 2, 2021
Divergences in a non-centered computational model Modeling fitting-issues	21	1706	October 30, 2019
How to speed up sampling in rstan? Modeling rstan , performance , hierarchical-model	6	6316	July 24, 2020
How to determine the combination of adapt_delta / max_treedepth / metric=dense_e vs. diag_e / term_buffer / window / warmup / iter? Modeling	2	1757	May 8, 2019
Non-centered parameterization for non-hierarchical parameters? Modeling reparametrization	15	1658	May 6, 2021

Problems remain after non-centered parametrization of correlated parameters

Related topics