Throw exception when <multiplier = 0>?

Developers
26

You have to be more specific. If we write

parameters {
  real<lower = 0> sigma;

then the underlying parameter over which Stan samples is log(sigma). But, every iteration, the constrained value sigma is calculated using exp(). Then sigma is the only variable exposed. To get back to the unconstrained variable, you’d have to do log(sigma), but it’s already too late.

Now if you want to do this yoursekf, it’s easy:

parameters {
  real log_sigma;
transformed parameters {
  real<lower = 0>  sigma = exp(log_sigma);

If you want to put a distribution on sigma, you also need the Jacobian adjustment

model {
  target += fabs(inv(sigma));  // Jacobian adjustment

But if you just wnat to put a prior on log_sigma, such as a normal to imply sigma is lognormal, that you can do. But you still can’t put the resulting sigma into a normal distribution if it underflows to zero.