

Number of clusters for Bayesian estim... 

Message/Author 


Dear Linda and Bengt, I was wondering if there's any study (or rule of thumb) for the minimal number of clusters for a multilevel model with Bayesian estimator? I assume, the criterion here would be different from what is suggested for when the "frequentest" estimators are used (e.g. Muthe´n 1989; Meuleman and Billet 2009). Specifically, I am fitting a model using a longitudinal ANES datafile with 19 timepoints and also the same model restricted to 13 timepoints. The model seems to fit fine (although there's obviously no fit indices available) and confirms the hypotheses, but I would like to have some support for the # of clusters used. The same model estimated by WLMSV or ML methods doesn't converge. Thank you, Dmitriy 


A quick followup on my own post. When I check graphs for my model (plot2), I only get the mean, median and mode of (I assume) posterior distributions. These are given for each parameter in the model. Would there also be a way to get the posterior predictive pvalue for the model (observedreplicated) similar to what you describe in a "Short Courses Topic 9?" (e.g. slide 56). Thank you, Dmitriy 


See the multilevel section of Muthén, B. (2010). Bayesian analysis in Mplus: A brief introduction. Technical Report. This shows that you can get good results for only 10 clusters. There is no PPP approach available yet for multilevel. 


Thanks for the prompt reply, Dr. Muthen How would I then judge whether the model fits alright? Can I rely on the graphical output and conclude by eyeballing the distribution? 


You may want to work with competing models. So for instance, if you fit a linear model, also fit a quadratic to see if the quadratic parts are needed. 

Back to top 

