I’m fitting a variant of the multivariate probit model (as in the docs) and have noticed that the elements of the diagonal metric are all unity. Because the large majority of parameters are the latent variables z, which follow a multivariate normal distribution with unit variances, that is of course not extremely surprising.
Because fitting the model takes several days however, it would be expedient to skip tuning the metric matrix M, but not the stepsize parameter. Is it correct that the best way to go about this is to set window=0 in sample()? Is the stepsize parameter still sufficiently tuned by the fast timestep adaption intervals? Would it be wise to also drop the second fast timestep interval?
For completeness: The sampling efficiency does benefit notably from an unrestricted M (ie metric="dense_e", but the warm-up time this takes is way greater than sampling more at a lower efficiency base on a diagonal M.