How to stop Stan from sampling a transformed parameter

General
1

Hey folks,
A model I’m working on uses a Dirichlet-multinomial conjugate pair, and runs accurately. However, the simplex that I feed to my Dirichlet prior (“alpha_vec”) is being needlessly sampled, slowing things down. How can I tell Stan that I don’t need this to be sampled (after all, it’s the same every time)?
Here’s the code

data {
	// Prior Parameters
   int<lower = 1> K; 		// COLS. for Dirichlet & txn mat
   int<lower = 1> N; 		// ROWS. for txn mat
   real<lower = 0> alpha; 	// for Dirichlet

        // Experimental Parameters
   int T[N,K]; 		// input RAW txn matrix, the data
 }

parameters {
  simplex[N] P[K];       // parameter of interest
 }

transformed parameters {
    vector[K] alpha_vec;
    for(i in 1:K){
    alpha_vec[i] = alpha;
   }
}
  

model {
    for(i in 1:N) {
    P[i] ~ dirichlet(alpha_vec);  // prior
    T[i] ~ multinomial(P[i]);     // LLK
    }
 }

The output of the fit (in Rstan) is:
Screen Shot 2020-07-13 at 11.23.43 AM

I’ve tried relabelling the “transformed parameters” block to “transformed data” and “generated quantities,” but both of these prevent the model from compiling.
Much appreciated!

2

Could you make it in transformed data?

2 Likes
3

Thanks for the reply! I tried that, but it doesn’t seem to like it:
Screen Shot 2020-07-13 at 11.36.21 AM

4

Is it after data like this? transformed data needs to be before the parameters block.

data {
	// Prior Parameters
   int<lower = 1> K; 		// COLS. for Dirichlet & txn mat
   int<lower = 1> N; 		// ROWS. for txn mat
   real<lower = 0> alpha; 	// for Dirichlet

        // Experimental Parameters
   int T[N,K]; 		// input RAW txn matrix, the data
}

transformed data {
  vector[K] alpha_vec;
  for(i in 1:K) {
    alpha_vec[i] = alpha;
  }
}

parameters {
  simplex[N] P[K];       // parameter of interest
}

If the above works then we should really improve that error message because it’s not terribly informative

2 Likes
5

Yeah good call. It should say something about the required order of the blocks.

2 Likes
6

Woops — completely forgot about that rule. Worked like a charm. Thanks so much, Steve!

1 Like
7

The new stanc3 compiler has a different message:


Syntax error in 'model.stan', line 11, column 0 to line 13, column 1, parsing error:
   -------------------------------------------------
    11:  parameters {
    12:    simplex[N] P[K];       // parameter of interest
    13:  }
          ^
    14:  
    15:  transformed data {
   -------------------------------------------------

"transformed parameters {", "model {" or "generated quantities {" expected after end of parameters block.

Still not ideal; the caret points to the wrong line it didn’t recognize transformed data as a valid (if misplaced) block name.

3 Likes
8

Yeah still not ideal, but that’s definitely an improvement!

3 Likes