www.cmm.bristol.ac.uk/forum

For cross-classified models the residual number should correspond to the sorted ID numbers for that classification, i.e. if your classification is in column c1 and you wanted a short version in c2 to match the residual ordering then the command:

UNIQ c1 c2

should give you a short ID column in ...

Unfortunately this currently isn't available and we don't have access to a Linux version of Matlab to build it. The Realcom-impute application isn't directly called from Stata it just outputs files that the -realcomImpute- Stata command can use. This means that you can impute the data using a ...

1) There isn't an option to save as one output dataset, however as you are using Stata you should be able to use Stata's http://www.stata.com/help.cgi?mi commands to convert the data to the format that you want. To do this you would use:

mi import flongsep ...

to load all the data files and then ...

Manually setting the starting values is covered in chapter 5 of the MCMC manual (see http://www.bris.ac.uk/cmm/software/mlwi ... ection.5.5).

If you're getting all zeros for the level 2 covariance matrix then it's worth checking the hierarchy viewer to make sure that it matches what you expect. If not then your data may not be sorted by the model hierarchy. Assuming that it does look as expected you can choose alternative starting values ...

That's a bit of a mystery then. I would suggest that you update to the most recent version of MLwiN ( http://www.bris.ac.uk/cmm/software/mlwin/download/upgrades.html ) in case this was caused by something that has been fixed since version 2.21. If the error message is highlighting the correct ...

Have you checked whether the version of MLwiN is the same on both machines? If it isn't then what are the version numbers of the one that works and the one that doesn't?

Without seeing your model/data I can't give a definite reason for this error message. What linearisation method are you using? You will sometimes see messages like this when it is having trouble fitting using second order PQL. If this is the case you can sometimes get the model to converge by first ...

Realcom only understands numeric variables, so you will need to generate the categorical dummies before importing the data.

If you still find the problem after running again and can replicate it on simulated data then we could look at that instead.