MCMC tuning samples

[polmvand]

Dear all,

I have a question about how to determine an appropriate number of tuning samples in MCMC variance components models.

Currently I am running a multistate live-dead model on 25 years of data (effective sample size ~8,000), in which I have modeled several hyperdistributions on s and psi beta-parameters to estimate variance components. Using values from a classical non-MCMC analysis (with annealing) as initial estimates, I ran some MCMC models using the program GIBBSIT feature to determine appropriate burn-in, storage and thinning parameters.

Ideally, I would like to also optimize the sample size of the tune-in sample (both from an estimation-precision point of view as well as runtime point of view -a single model takes days to run....).

However, how do I determine what is an appropriate sample size to tune the size of the proposal distribution SD? The default value in MARK is set to 4,000. What is the rationale behind this 4,000 value, and on what factors should the tune-in sample size depend?

Anyone got some ideas about this?
thanks in advance and best wishes,
Martijn