No correspoding probability function for bernoulli_logit_glm?

Interfaces RStan
#1

Operating System: Clear Linux (28100) and Antergos (Linux)
Interface Version: rstan (Version 2.18.2, GitRev: 2e1f913d3ca3)
Compiler/Toolkit: gcc 8.3.1 and gcc 8.2.1

Hi!
Here is documentation for bernoulli_logit_glm, yet when I try it in a simple logistic regression model:

data {
  int<lower=0> N;
  int y[N];
  matrix[N,4] X;
  real mu0;
  real mu1;
  real sigma0;
  real sigma1;
}
parameters {
  real beta0;
  vector[4] beta1;
}
model {
  beta0 ~ normal(mu0, sigma0);
  beta1 ~ normal(mu1, sigma1);
  y ~ bernoulli_logit_glm(X, beta0, beta1);
  // y ~ bernoulli_logit(beta0 + X * beta1);
}

I get the error:

Probability function must end in _lpdf or _lpmf. Found distribution family = bernoulli_logit_glm with no corresponding probability function bernoulli_logit_glm_lpdf, bernoulli_logit_glm_lpmf, or bernoulli_logit_glm_log

This comes up both in RStudio (preview) and when trying to compile the model.

I was intending to benchmark this model, and given Bob Carpenterā€™s statements in this post from October 2017:

Matthijsā€™s first major commit is a set of GLM functions for negative binomial with log link (2ā€“6 times speedup), normal linear regression with identity link (4ā€“5 times), Poisson with log link (factor of 7) and bernoulli with logit link (9 times). Wow! And he didnā€™t just implement the straight-line caseā€”this is a fully vectorized implementation as a densityā€¦

Iā€™d expect a hefty speed up.

#2

Unlikely, but does calling rstan::stan_version() in R return anything other than "2.18.0"?

#3

It returns ā€œ2.18.1ā€.

#4

More basic assumption tests: If you switch to:

target += bernoulli_logit_glm_lpmf(y | X, beta0, beta1);

in the model block, do you still get the same error?

#5

Still not found:

No matches for:
Function bernoulli_logit_glm_lpmf not found.

#6

Yeah weird. There is some version mismatch somewhere. My only suspicion is that this:

is too high a compiler version and something has broken somewhere. Maybe check which versions of StanHeaders and BH you have installed?

#7

So, can you confirm you have these functions?

I could also try clang 7.0.1.
I also have a roughly 1-week old build of gcc-trunk (ie, the development branch).

Even my Ubuntu laptop has gcc 8.2.0 as the default compiler, but Iā€™d (not knowing the Stan code base) be surprised if itā€™s the compilerā€¦
I can try on the laptop later.

> packageVersion("BH")
[1] ā€˜1.69.0.1ā€™
> packageVersion("StanHeaders")
[1] ā€˜2.18.1ā€™
#8

Probably should have started by checking this myself - I also get the same compilation error.

#9

Itā€™s not in RStan yet.

1 Like
#10

Do you know if there is an easy way to build and install the develop branch?
If I were to install_github, with ref=ā€œdevelopā€, it would use the development branch of rstan, which still points at Stan 2.18.

#11

You might want to check the bottom of this if you havenā€™t already.

remotes::install_github("stan-dev/rstan", ref = "develop", subdir = "rstan/rstan", build_opts = "")

(I never did this. If you still have questions you might want to ping Ben Goodrich.)

#12

The problem is that the development version of RStan (the interface) looks like it does not use the development version of Stan. FWIW, I did just run that, restarted R, and see the same behavior. stan_version() still returns 2.18.

@bgoodri, any suggestions, or time-frame until a release where these glm functions are supported?

#13

It depends on when Stan 2.19 gets released.

#14

Iā€™d just like to pop in and mention I am also having these difficulties, or a related one at least. I was going to start a new thread but found this via the searchā€¦

data {
  int<lower=1> N;  // total number of observations
  int Y[N];  // response variable
  int<lower=1> K;  // number of population-level effects
  matrix[N, K] X;  // population-level design matrix
}
parameters {
  real Intercept;  // temporary intercept
  vector[K] beta;  // population-level effects
}
model {
  target += normal_lpdf(Intercept | 0, 1);
  target += normal_lpdf(beta | 0, 1);
  for (i in 1:N){
    target += bernoulli_logit_glm_lpmf(Y[i] | X, Intercept, beta);
  }
}

Nor does this work

data {
  int<lower=1> N;  // total number of observations
  int Y[N];  // response variable
  int<lower=1> K;  // number of population-level effects
  matrix[N, K] X;  // population-level design matrix
}
parameters {
  real Intercept;  // temporary intercept
  vector[K] beta;  // population-level effects
}
model {
  target += normal_lpdf(Intercept | 0, 1);
  target += normal_lpdf(beta | 0, 1);
  target += bernoulli_logit_glm_lpmf(Y | X, Intercept, beta);
}

They wonā€™t compile since there is no function found. The latest version of rstan, etc are installed (and just reinstalled to be sure).

#15

You need rstan 2.19, which hasnā€™t been released yet.

CmdStan 2.19 was released 10 days ago, so I would guess rstan 2.19 is coming soon.