MCMC variance overestimation

[amjeannet]

Hello,

I am new to MCMC techniques so please bear with me if this question is very basic.

I am using the European Social Survey so I am working with individuals nested within countries (15 countries) and several years have been pooled together. (Pooled Time Series Cross Sectional).

I am struggling with strange variance estimates when I run MCMC estimates of my model using MLwiN. When I use RIGLS, the variance components are country level (0.4) and individual level (5.4) but when I use MCMC the country variation goes way up to to 5.2 and the individual variance stays the same (5.4). Plus, it is strange that when I added country-level predictors to the model when using MCMC the country variance increases rather than decreases.
This does not happen when I use RIGLS and there is a strong theoretical basis for including these predictors and they are significant under the maximum liklihood framwork..

Any help would be much appreciated.

Anne-Marie

[ChrisCharlton]

There is no particular reason why this should happen. Have you checked the convergence diagnostics to ensure that all the chains are stable, as it may just take a while for the model to get to the correct distributions, meaning that you would need to change the burnin and chain length from the defaults?

[amjeannet]

Thanks for getting back to me with some thoughts.

I have increased the burn-in/iterations (even up to a million iterations) and I still have the same problem.
The coefficients converge (the intercept with greater difficulty but eventually it converges).

The strange thing is that the country-level variance is 5 times bigger than in RIGLS when I run the null model in MCMC.

Any help on this would extremely appreciated!

[billb]

Dear Anne-Marie,
If you were able to send me the worksheet I'd happily take a look and check that there wasn't anything silly going on.
Best wishes,
Bill.