A cross-classified data structure?

[erol1945]

Dear MlwiN user,
It might be a very basic question but I would like to check with you how I should consider the following data structure:
At first glance this looks like a "basic" four levels" structure with repeated surgical procedures nested within patient nested within surgeon nested within hospital. After each procedure a binary outcome is measured (0 vs 1).

However, the surgical procedures of a patient can be performed either
by the same surgeon in the same hospital,
by the same surgeon in different hospitals,
by different surgeons in the the same hospital
by different surgeons in different hospitals
or a mix of those scenarios...

Obviously a surgeon can operate in different hospital; and some patients will only have one procedure.



An example of my dataset could be:
Surgery Patient Surgeon Hospital Outcome
1 1 1 1 1
2 1 1 1 0
3 1 1 1 0
4 2 2 2 1
5 3 3 3 1
6 3 3 4 1
7 4 4 2 0
8 4 5 2 1
9 5 5 4 0
10 5 6 5 0
11 5 7 6 1
12 6 5 7 0
13 6 5 7 0
14 6 8 8 0
15 6 8 9 1
etc.

To me it looks like a longitudinal study with a cross-classified four level structure with
repeated measures of outcomes assessed at each procedure (level 1), ie procedure nested within patient (level2) but patient nested within different surgeons (level3, the first cross-classified level) and surgeons nested in different hospitals (level 4, the second cross-classified level).

Is it correct?If yes, would it be easier and computationally quicker to model a logistic three levels cross-classified model within a MCMC framework rather than a RIGLS ones?
(I intend to use runmlwin for this purpose)

I would be grateful if you could advise me on the best way to conceptualise and model this structure.

Thank you very much for your help
Erol

[GeorgeLeckie]

Dear Erol,

Yes you have a three-way cross-classified data structure. You have procedures/surgeries nested within the cells formed by the three-way classification of patients, surgeons and hospitals.

The binary nature of your response, the complexity of the data structure, and the likely size of your data each make this problem very hard to carry out in a (quasi) maximum likelihood estimation framework. So, yes fit this model using MCMC methods.

You need to provide starting values for all parameters in the model. I would recommend supplying starting values for the fixed-part regression coefficients from a single-level logistic regression, and then simply manually specifying starting values of, say, 1 for the patient, surgeon, and hospital variance-components.

Do read through Module 12 of our free multilevel modelling online course which gives 50 pages of theory and 50 pages of MLwiN instructions for fitting cross-classified models, albeit to continuous responses.

http://www.bristol.ac.uk/cmm/learning/o ... index.html

Let us know how you get on

Best wishes

George