TaRB-MH algorithm, `mh_jscale`

Dear all,

Do we need mh_jscale when using TaRB-MH algorithm as MCMC sampler?

Dynare manual says that it should be specified. Even, the source codes use bayestopt_.jscale, for example in posterior_sampler_iteration.m, line 116.

But the paper, Chib and Ramamurthy (2009), says that it is a black-box algorithm and scaling with mh_jscale is not mentioned in the paper (to the best of my knowledge).

Moreover, specifying it seems not very meaningful. Because the Hessian changes for each block and at each iteration. It might get substantially different from the initial one as sampler progresses. So, the initial mh_jscale will not work anyway after a while.

Chib and Ramamurthy (2010) simply use the Hessian, i.e. an mh_jscale of exactly 1. Dynare allows you to set another value if you want to.

1 Like