www.cmm.bristol.ac.uk/forum

I have been using runmlwin to do multilevel Poisson modelling (repeated measures data) using MCMC. I understand that MQL (and PQL?) as used in MLwin can produce estimates that are biased towards the null. However, I am wondering how these compare to estimates produced by xtpoisson in Stata. Also does anyone have a reference illustrating how the methods compare (or something that will give me a bit more information about this)?

Thanks,
Rosie.

Hi Rosie,

Yes MQL and PQL typically give downwards biased results. In some situations (when you have a high degree of clustering and small clusters) these biases can be large. In other situations (when you have lower clustering and larger clusters) these biases can be very small. PQL2 is the least biased, MQL1 is the most biased.

In MLwiN we recommend you use MCMC which will give very similar results to adaptive quadrature in Stata (e.g. xtmelogit and xtmepoisson and gllamm). It will also give similar results to xtpoisson. You can use the MQL1 estimates as starting values for fitting the model by MCMC.

Note that if you are a Stata user then I strongly recommend you use the runmlwin command to call MLwiN from within Stata.

http://www.bristol.ac.uk/cmm/software/runmlwin/

For literature, see

http://link.springer.com/chapter/10.100 ... -73186-5_9

and also backwards and forwards references for further relevant literature

I hope that helps

George

Thanks George - that's great.

Hi all,
On the topic of xtpoisson, is there an easy way to calculate the VPC? The reference I've been reading is here http://people.upei.ca/hstryhn/iccpoisson.ppt. But I'm having trouble making sense of it all. This is how far I've got: to calculate the VPC you use the upper/upper+lower formula which is the same as exp(VAR)/exp(yhat +(VAR/2)).
Is this correct does anyone know? Or is there an easier way to get stata to spit out the VPC?

Thanks in advance!