www.cmm.bristol.ac.uk/forum

Hello

I have some cross-sectional count data I wish to model using MLwin. It is two-level data, patients within GP practices, 1 observation per patient. From my work in STATA, I know there is overdispersion that needs to be taken account of. With STATA, I can add another random effect, so that observations (=1 per patient) are clustered within patients within practices e.g.

xtmepoisson (outcome predictors) || practiceid: || patientid: , irr

Could someone advise how to do this in MLWin? When I set my outcome to a 3 level variable, with the highest level as practice, I am unsure how to correctly specify the other two levels.

Many thanks in advance

Ron

As you are a Stata user I will assume that you are using -runmlwin-, for which George has previously provided the following example:

https://www.cmm.bristol.ac.uk/forum/vie ... t=453#p985

MLwiN uses quasi-likelihood when fitting these models with (R)IGLS so it is recommended that you fit at least your final models using the MCMC estimation method.

Other ways that you can model overdispersion via -runmlwin- are with the extra option or fitting the model as negative-binomial, however these options are not currently allowed by -runmlwin- when fitting the model with MCMC.

Thanks for that. It has been a while since I used mlwin and had forgotten about the runmlwin option.