www.cmm.bristol.ac.uk/forum

Hi,

I fitted a discrete-time multilevel event history model to estimate the probability of leaving social assistance.
Beneficiaries are cross-nested in welfare agencies.
I use MCMC estimation with IGLS estimation for the nested model as priors.

In the current model the random effects at individual level and at agency-level are uncorrelated.
I tried to re-estimate the model using an unstructured covariance matrix (option corresiduals(unstructured) for the mcmc estimation), but this model failed to generate an estimate of the covariance between the random effects.
I get exactly the same output as for the model with uncorrelated random effects. Does this mean that MLwin is not able to do the estimation?

Is there another way to proceed to allow the random effects to be correlated?

Best wishes,

Sarah

Hi Sarah,

In multilevel models, all random effects are assumed uncorrelated across levels.

The unstructured covariance matrix (option corresiduals(unstructured) is for something different. It is only relevant to multivariate continuous response models, typically where you fitting repeated measured data. The option allows use to specify different structures for the covariance matrix between the repeated measures, for example AR1.

http://www.bristol.ac.uk/cmm/media/runm ... siduals.do

Best wishes

George

Thank you for this information, George