Proper Vectorization of Model

Modeling
1

I am working on a model like the one below and I am trying to come up with the right way to vectorize the V[k] inference in the model.

V[k] ~ binomial_logit(I[,kk[k]] , logit_p);

where logit_p is a vector of length N and I[,kk[k]] is array of length N.

Stan Code:

data {
  int<lower=0 > N;
  int<lower=0 > K;
  int<lower=0 > J;
  int<lower=0> V[K,N];
  int<lower=1> I[2,N];
  int<lower=1, upper=J> jj[K];
  int<lower=1, upper=2> kk[K];
}

parameters {
  vector[N] theta;
  vector<lower=0>[K] alpha;
  vector[N] gamma[J];
  vector[K] beta;
}

model {
  beta~ normal(0,1);
  theta~ normal(0, 1);
  alpha ~ lognormal(0, 1);
  for (i in 1: J){
    gamma[i] ~ normal(theta,1);
  }

  {
    vector[N] logit_p;
    for (k in 1:K) {
       logit_p = alpha[k]*gamma[jj[k]]  + beta[k]

       V[k] ~ binomial_logit(I[,kk[k]] , logit_p);
    }
  }
}
2

Add a semicolon at the end of the logit_p line so it looks like:

logit_p = alpha[k]*gamma[jj[k]]  + beta[k];

and swap the indices on I (so that the kk term goes first).

3

Oops thanks.

Ok. That makes perfect sense.
Would there be any measurable improvement for eliminating the for loop?

4

The gamma for loop? Probably not. These loops translate directly to C++. It’s should be fast.

But usually with these things just try it and see how it goes. If it helps, great, if it doesn’t do anything whatevs.