Constrained vs. unconstrained parameterization in Dynare’s MCMC

Estimation ML/Bayesian estimation
1

Hello,

Could you please clarify whether Dynare transforms parameters between constrained and unconstrained spaces during MCMC sampling?

For example, if a parameter has a Beta prior (support (0, 1)), does Dynare apply a logit transformation to map it to the real line before proposing draws, and then transform it back to the constrained space afterward? Or are the parameters sampled directly within their constrained domains without any such transformation?

Thank you

2

Dynare internally does not conduct such a parameter transformation to unbounded domains.

3

Thanks very much for your quick reply earlier.

Could I please confirm my understanding?

If a Metropolis–Hastings proposal draws parameter values that violate the prior bounds (i.e., fall outside the prior support), the draw is automatically rejected because the prior density is set to zero.

I assume that the same logic applies when the proposed parameters lead to model evaluation problems—specifically when the Blanchard–Kahn conditions are not satisfied —so that the likelihood is treated as zero and the proposal is also rejected.

Thank you

4

Yes, exactly. The latter is called implicit prior truncation and may cause issues in model comparison.