PerturbationDSGEPosterior.marginal_likelihood#

PerturbationDSGEPosterior.marginal_likelihood(*, tau=_MHM_TAU)[source]#

Geweke’s modified harmonic mean of the log marginal likelihood.

Computed in the chain’s unconstrained space, where the Gaussian envelope is least wrong; the evidence is invariant to the reparameterization. The likelihood entering the kernel is the particle-filter estimate carried by each draw, so the number is the Herbst-Schorfheide practice rather than an exact ordinate: the estimate noise biases it, and the record says so.

Parameters:

tau (float) – Probability mass of the envelope retained.

Return type:

_MarginalLikelihoodSelection

Returns:

The record, with the Monte Carlo error and a coverage note.