Dear Marco, 

First you have to test for significant differences between the two time periods. 

Dear Beatrice,

thanks a lot! Indeed there are significant difference. This is an additional test. I would like to go beyond the general effect and some interactions.

I would like to know how many changed their behavior, e.g. how many display a higher/lower value than predicted according to past coefficients.

marco

Jochen here at RG is good at this, ask what he thinks.

This is a trick that you can do in MLwin

Take the later data and stack it under the earlier data and put in a new column 0 for the earlier data and a 1 for the new data and called this dummy variable Exclude

In Model - Hierarchy Viewer - Option tick on conditional exclusion using the Exclude dummy variable

Estimate the model in the normal way you will see that a lot of data are missing (the left out later data)

Return to the Hierarchy viewer and tick off exclusion - do not re-estimate the model but

Model - Predictions in the normal way to get the complete set of predictions based on all the data and the estimated coefficients for the earlier model

You can use this procedure to perform a stratified model.

You could also take the full data set and  include the Exclude variable and form an interaction between it and each and every predictor to test for change  in an elaboration of a Chow test https://en.wikipedia.org/wiki/Chow_test

Dear Kelvyn,

thank you so much.

Illuminating as usual!

best wishes

Marco