www.cmm.bristol.ac.uk/forum

Hi,

I'm relatively new to MLwiN. Recently I tried to fit a discrete time event history model with cross-classified structure in MLwiN. The problem I had was that the ESS kept being very small when I did the variance component model. The data I have is more than 8 million observations with 30000+ units and 250+ units for the two higher levels that are crossed. I have tried a 5000 burn-in and 50000 iterations with both orthogonal parameterisation and hierarchical centering (it took more than two days to run the model). Yet, the ESS for two of the parameters are still below 30. (a longer burn-in and way longer iteration with thinning might work but could take weeks to run I suppose?)

Any suggestions and insights on the potential cause and solution to it would be really appreciated.

Thanks!

Dear Xinyuzou,
There are many reasons why mixing might be poor and different methods e.g. orthogonal parameterisation and hierarchical centering will fix different issues. Other things to consider are centering predictor variables. It may end up that you have to run for longer. The paper by Browne et al. (2009) in JRSS Series A might be helpful as that looks at discrete time event history.
Best wishes,
Bill.

Thanks Bill !