Computational cost of unused parameters?

When modeling, for example, random intercepts in a regression, I follow the standard of indexing a vector the length of the number of groups in the variable:

data {

  int<lower = 0> N;
  int<lower = 0> N_grps;

  int<lower = 1, upper = N_grps> grp [N];
parameters {
  vector [N_grps] mu_grp; 

Sometimes I test my code on subsets of the data without rows from every member of the group. I like to avoid re-indexing the group for simplicity, so instead of declaring a smaller N_grps, I keep it the same and end up having elements of the vector mu_grp that are never referred to in the likelihood and therefore (I think) never actually “fit,” beyond their prior.

This doesn’t seem to cause much of a computational difference for modest N and N_grp, but I’m wondering if I should expect worse performance for bigger datasets or, more broadly speaking, if I should avoid including unused parameters for other reasons.

1 Like

It’s OK if you have a proper prior (ideally with two finite moments) on the otherwised unused parameters.


Note that a related thread is here, though you’ve asked the question more succinctly!

1 Like