Dear forum
I am using cross-classified multilevel models in MLwiN to investigate effects of primary and secondary schools. In a cross-classification model, you first need to specify a ‘naïve’ hierarchical multilevel model and in a second step you need to specify your cross-classification. My question is: does it matter in which order you specify your classifications in the first, ‘naïve’ multilevel step?
I ask this question because I found different results according to the order of classifications I use. In the first case, I situate students within primary schools within secondary schools and I don’t find significant cross-classified results. In the second case, I situate students within secondary schools within primary schools and I do find significant cross-classified results. It seems intuitive to use the ‘student-primary school-secondary school’ order because of the chronological order, but in other cross-classification models there is no chronological order (f.e. with students in neighbourhoods and schools). On the other hand, in the first step, adequate starting values are estimated using the ‘naïve’ hierarchical model, so you would argue the order of the classifications does matter.
Is there a good reason to use one order instead of the other? Or doesn’t it matter?
Best regards
Griet
Classification order in cross-classified multilevel models
Re: Classification order in cross-classified multilevel models
Hi Griet,
It shouldn't matter but
(1) you should make sure you tick the cross-classified button so that it is fitting the XC model!
(2) you should ensure you run the MCMC estimation for sufficiently long that the chains converge to similar answers - they will not be exactly the same due to the Monte Carlo nature of the methods.
Best wishes,
Bill.
It shouldn't matter but
(1) you should make sure you tick the cross-classified button so that it is fitting the XC model!
(2) you should ensure you run the MCMC estimation for sufficiently long that the chains converge to similar answers - they will not be exactly the same due to the Monte Carlo nature of the methods.
Best wishes,
Bill.