Is it ok to declare intermediate variables which are used repeatedly in for-loops?

Modeling
1

Eg, I have a transformed parameters block that looks like

transformed parameters {
real a;
real b;

for (i in 1:N) {
a <- some_expression_that_depends_on_i
b <- some_expression_that_depends_on_i
# compute things using a and b.
}
}

this is okay, right? I don’t have to re-declare a and b in every iteration of the for-loop.

2

Yes, but you probably should declare a and b inside the for loop. Otherwise, you will get back posterior margins for them that pertain only to the i = N case.

3

Ah, perfect, thanks so much for the quick reply. I don’t need the posteriors of a and b (ie, I never access them in the fitted stan model). I just wanted to make sure that computations that used them subsequently would be correct…eg,

transformed parameters {
real a;
real b;
vector[N] phi

for (i in 1:N) {
a <- i
b <- i
phi[i] <- a + b
}
}

will fill in the correct values for phi (phi[1] = 2, phi[2] = 4, etc).

4

Yes, but don’t do it that way. Do it like this:

transformed parameters {
  vector[N] phi;
  for (i in 1:N) {
    real a = i;
    real b = i;
    phi[i] = a + b;
  }
}
5

Thanks! I understand that’s the preferred way if I wish to recover the posterior marginals of a and b (and I’ll do it that way in future). But either

transformed parameters {
 vector[N] phi;
 for (i in 1:N) {
   real a = i;
   real b = i;
   phi[i] = a + b;
 }
}

or

transformed parameters {
  real a; 
  real b;
  vector[N] phi;
  for (i in 1:N) {
    a = i;
    b = i;
    phi[i] = a + b;
  }
}

should yield the same inferences for phi, correct?

6

They both yield the same distribution for phi but the first version does not even let a and b exist outside that loop, so not only do you not waste space storing them, you cannot make mistakes accidentally referencing them later.

1 Like
7

Yep, makes sense! Thanks so much for taking the time; I understand how things work better now.