Monitor subset of parameters in cmdstanr?

General
21

I really hope this comes across as helpful and not argumentative or anything like that:

FWIW, I you can now pass vectors of bounds where the bounds are themselves parameter vectors :)

For example:

parameters {
  vector[2] bounds;
  vector<lower=bounds>[2] x;
}

model {
  bounds ~ std_normal(); // note that the interaction of this prior and the 
  // constraint means that this is not the realized prior margin

  x ~ std_normal(); // same comment applies here
}

This is new-ish functionality, and maybe not as widely known as it should be.

2 Likes
22

This is great! I definitely was not aware of this. I guess the documentation about varying upper/lower bounds can be updated.

2 Likes
23

Sorry if iā€™m being dense but does this solve the issue i discussed in my original thread (which is specifically regarding the nuisance parameters in the GHK parameterisaton of the multivariate probit model, (see e.g., post from @bgoodri here)? iā€™m pretty sure thereā€™s no way to declare those ā€œuā€™sā€ there as local variables or to exclude them from the cmdstanr/rstan output (which, can sometimes be a really big deal, since these models can easily take up >100GB of disk space for every run on large datasets with many chains).

Anyway, for anyone else interested in MVP models, iā€™ve since implemented my own HMC-based sampler (using algorithms based on the very recently proposed adaptive MALT and RHMC algorithms - which are both much more efficient** than Stanā€™s current HMC-NUTS based algorithm - I wouldnā€™t be surprised if variations of these algorithms will be implemented in Stan within a few years!) which solves the issue which this thread is about. The sampler uses manual gradients (which is much faster than Stanā€™s C++ autodiff library, at least for latent class MVP models using the GHK parameterisation). It is a custom-coded algorithm (mostly in C++ and some R) which currently only works for 2-class latent class MVP models, so it is not a replacement to Stan at all, but specifically for MVP models for large-ish datasets (e.g., N = 8000 and 6 outcomes, giving 48,000 total observations) it is over 15-times more efficient than Stan - I am yet to test bigger datasets but the relative efficiency will likely be much higher than 15 for larger N. It also doesnā€™t monitor or save the nuisance parameters, so hardly takes up any disk space whereas Stan quickly becomes impractical for large N. Plan on releasing an R package for it soon.

**edit: at least for some models, such as those discussed in the 2 papers I linked and MVP-based models. There may well be some models for which the adaptive MALT and/or RHMC-based algorithms do worse than Stanā€™s NUTS-HMC sampler.

1 Like