Unpacking info in c1090 etc for multiple response models

[socgdj]

I am developing a two-level multiple response model with three responses and a set of random and effects for each response. I wanted to make predictions for individual cases, but I am not able to unpack the information in c1090 etc. and the corresponding help page (Reserved groups, columns and boxes) does not really help me. There is a cryptic reference on the help page referred to: For the multivariate setup window C1093 holds values in table. In my case however c1093 is empty!

I don't know how to proceed. Can anyone help me?

Göran

[ChrisCharlton]

Can you be more specific as to which values you are trying to extract to perform your predictions? Are you not able to do what you want with the the predictions/customised predictions windows?

The reference to c1093 is left over from an older version of MLwiN, currently this column is not used to store any model related information.

[socgdj]

Thanks for the clarifying question, Chris!

I wanted to weight my data and make predictions from a multiple response model, but then I encountered this message in the help text of the weighting menu: "Note that weights cannot be used with multivariate models." So I tried to locate the parameter values for the units (at different levels), in columns 1090 etc., which I could not.

After your reply my question boils down to one on weighting: How should I proceed if I want to do weighted predictions from my multiple response model?

Göran

[ChrisCharlton]

The point estimates corresponding to the equations window can be found in c1096-c1099, and are also available through the "store model" functionality and the FIXE/RAND commands. The residuals at each level are not stored by default so you would need to calculate these with the residual window/command. As you have seen MLwiN does not support weighting in multivariate models so if you wanted to use them here you would have to follow the procedures outlined in the "weighting in MLwiN" section at http://www.bristol.ac.uk/cmm/learning/topics.html and write a macro. A colleague has tried this with simulated data however found that the answers for the random part were somewhat different to those expected, so you would need to do extensive checking if you wanted to follow this approach.

[socgdj]

Thanks Chris for a good answer. This is all I need to know (I hope)!

Göran