www.cmm.bristol.ac.uk/forum

Hi,

I am doing some binomial logistic regression modelling. In the past I have always estimated my models using both IGLS and MCMC, but this was because my models were cross-classified and as such required MCMC estimation.

I am now running binomial logistic regression models, but they are either single-level or are hierarchical models, and so I'm not sure if they require the additional MCMC estimation?

I have done both (with and without MCMC) so far and they produce very different estimates, with this in mind I'm not sure which results are correct?

Many thanks.

MLwiN uses quasi-likelihood when fitting discrete models with (R)IGLS, which can give biased results (see http://www.bristol.ac.uk/cmm/software/s ... entresults). If you see differences betwen the (R)IGLS and MCMC estimates then the recommendation would be to use the MCMC results, although obviously you would want to check the chains to make sure that they have converged to a stable distribution.

Hi Chris,

Thanks for this. I have checked the trajectories for the MCMC estimations and they look fine.

I just need to double-check that it is correct to use MCMC estimation, even for single-level models?

That's my understanding, as MLwiN will use the same quasi-likelihood methods for single-level models. If you have other software available you could of course compare the results you get with their estimates.