Dear all,

I hope this message finds you all well. I was wondering what would be an appropriate number of draws to run when using the ‘slice’ posterior sampling method for estimation in Dynare. I am planning on sending my paper to a journal, so I was curious to know what is the academic stance on this. Many thanks in advance!

Whereas 1 Metropolis iteration requires 1 likelihood evaluation, 1 slice iteration requires N likelihood evaluations. A rule of thumb value for N is

N = npar*7, where npar is the number of estimated parameters.

So, if you have a budget of 100000 likelihood evaluations, you should scale this down by dividing by N, i.e. 100000/N, in order to avoid never ending chains.

The chain you will get is much less autocorrelated wrt the MH one, and the number of ‘effective draws’ is a function of the the inefficiency factor E.

Dear Professor Ratto,

Many thanks for your reply. I have in total 21 parameters to estimate, which implies that 1 slice iteration requires 147 likelihood evaluations (N = 21x7). Does this mean that I should select the number of draws such that number of draws = desired number of draws / (21x7)? For instance, if I wanted to run the model (without the slice sampling) of 2 million draws, then I should now consider running 2 million / 147 draws if I am not mistaken. Thank you very much for your time and consideration!

yes, you should take 2 million / 147 draws

Dear Professor Ratto,

Thank you very much for your help and time, it is very much appreciated!

Best wishes,

Oscar