www.cmm.bristol.ac.uk/forum

Hi,
I'm very new to MLwiN and multilevel modelling, and am trying to replicate an analysis of some data that was completed using MLwiN. They have used a two level binomial model, and compared models with and without interaction terms using the Likelihood Ratio Statistic. These are my problems:
1) When you use the 'like' command in the command interface does it give you -2*log(likelihood) or 2*log (likelihood)? There is some descrepancy between teaching materials and the help pages
2) If it does give you -2*log(likelihood), why does the null model (no interaction terms etc) have a lower value of -2*log(likelihood) than the model with the interaction terms included?

I fitted the models using MQL1 then PQL2 (and dabbled in MCMC but the ESS was tiny and the correlation very high)
Thanks

Discrete response models within MLwiN are estimated using quasi-likelihood rather than maximum likelihood, which means that it isn't possible to calculate a likelihood statistic for these models and this is why it isn't displayed in the equation window. In order to compare these models it is recommended that you use a Wald test instead.