www.cmm.bristol.ac.uk/forum

Hi everyone, I have a problem in running MLwiN for analysing a 2-level modelling. I have data of 229 individual nested in 20 groups, and am going to estabilsh a multilevel modelling using MCMC bayesian estimation. When I run the program, it always comes to a error message of either 'SSP matrix for random part has gone negative definite' or 'SSP matrix for fix part has gone negative definite'. Both the dependent and independent variables are continues, however, for some independent variables, there are only 5 or 6 values due to the calculation methods. Does it influence the estimation?

I thought about the reason that the lack of my data especially in the second level may be problematic, but there are some studies shown that less than 20 groups could do multilevel analysis. So I don't know how does the problem happen.


Could anyone please help and explain what reasons causing such problems? and what should I do to overcome them? Thanks a lot!

These error messages often do indicate a problem with the data, however as you will be fitting your models with MCMC the (R)IGLS estimates will just be used as starting values (and in the case of the random part co-variance matrix, priors). You could therefore manually put in starting values and see how well the model is estimated with MCMC. I would suggest using a range of these to test the sensitivity of the model to the initial values. Chapter 5 of the MCMC guide (see: http://www.bristol.ac.uk/cmm/media/soft ... mc-web.pdf) covers using your own starting values, although if you have a recent version of MLwiN you can right-click the parameters in the equation window and enter them directly.