Problems with Reparameterisation

[GDistiller]

A while back I was trying to work out what the best way was to setup and run a model using MCMC and I managed to get some very useful advice in response to a posting I made here. I was uncertain about whether the intercept parameter needed to be included in the hyperdbn when there was dummy coding used ie when the reference category is parameterised with a row of zeros. The resolution was to use effects coding with a row of -1's thereby making the intercept equate the overall mean, and the random effects are then deviations around this grand mean with a mean of zero.

Since then I have got an extra year of data that I have added on and unfortunately have been battling a bit to get my models to run (as fixed effects). I have now realised that somehow the -1 effects coding seems to be causing problems. So for example I try and fit a particular model which takes quite long to converge and then has a whole bunch of estimates with std errors of zero. I go into the DM and change the ref cat to have a row of zeros rather than a row of -1's and then the model runs much much quicker, and most importantly produces a set of estimates with std errors (for almost all parameters). The other thing that has been worrying me is that the model Deviance changes quite a bit where I would have expected a reparameterisation to produce exactly the same Deviance?

Has anyone experienced similar problems after a simple reparameterisation? I'm trying to understand what is going on here and any advice would be much appreciated.

Thanks

Greg