I really like the http://www3.stat.sinica.edu.tw/statistica/oldpdf/A10n416.pdf>Barnard, McCulloch and Meng (Statistica Sinica 10(2000), 1281-1311) way of formulating priors for covariance matrices. That is they place log-normal priors on the standard deviations, and let the elements of the correlation matrix be sampled from (essentially) rescaled Betas (so that they are on [-1,1] instead of [0,1]). If S is diagonal matrix contain the standard deviations and R is the correlation matrix then we can reconstruct the covariance from the product S x R x S.

I'd very much like to do this in JAGS, but I cannot see a way around the issues of assigning constants to stochastic nodes. Is it possible?

Last edit: James 2013-01-22