Different ESS in MLwiN and R2MLwiN

[billb]

Morning Andy and Chris,
When I wrote the ESS calculation in MLwiN 20 odd years ago as you'll see I decided to limit the calculation to the first 1000 elements of the autocorrelation function probably partly because of time/space considerations (computers were slower at the time) and partly because if correlations were so high that even after a lag 1000 there would be big correlations then I would assume MCMC was mixing incredibly badly.
Looking at your chains Andy it looks like even after lag 100 the correlation hasn't shifted much from 1 so I suspect it is still big at 1000 and hence the difference.
I would check also Andy that your 2 models seem to possibly have hierarchical centering switched on at different levels as you have ordered birth_year and year the other way around. This might explain the difference in variance ESS.
Best wishes,
Bill.

[andrewjdbell]

Thanks Bill, makes sense (and yes I checked with and without hierarchical centering at both levels - didn't make any meaningful difference).

These are age-period-cohort models so perhaps unsurprising they aren't mixing well - we're interested in the deviations so in a sense it doesn't matter so much if the linear components aren't reliable, would just have been nice to have the whole model functioning well...

Thanks both