www.cmm.bristol.ac.uk/forum

I am conducting a study on binary and categorical repeated measures data. I tried fitting multilevel model for a smaller data set using MLwiN as a beginning. The example data set (Grizzle, Starmer and Koch (1969)) I tried is as follows,
Grizzle, Starmer and Koch (1969) analyze data where 46 experimental units treated with three different drugs, drug A, drug B and drug C. The response recorded for each drug was favorability or unfavorability of the drug.

H0 : Marginal probability of favorable response is same for all three drugs
H1 : Marginal probability of favorable response is not same for all three drugs

F- Favorable U - Unfavorable
A F F F F U U U U
B F F U U F F U U
C F U F U F U F U
6 16 2 4 2 4 6 6
(A, B, and C are the three drugs that need to be tested and there are 46 patients)
(I expand the data set before entering it to MLwiN)

I read the user manual for MLwiN and there, it is mentioned that, there are some issues that will be arise when dealing with discrete repeated responses. I am eager to know about those issues. Are there any problems arise when entering data?
I really want to know whether it is possible to fit a multilevel model using MLwiN for this data set and if so, what are the issues that can be faced when fitting a multilevel model.

Hopefully looking forward for a quick reply and answers
Thank you.

Would it be possible to clarify which section of the manual you are referring to (i.e. section and page numbers)?

Thank you for the reply.
The user guide for MLwiN users (version 2.16), 13th chapter, Fitting models for repeated measures data, page 193, 4th paragraph.
I need to analyze some categorical repeated measures data using multilevel approach in MLwiN, that I have simulated. I really want to know whether is it possible to carry out the analysis using MLwiN using for categorical repeated measures data or whether there are issues.
I would be much obliged if you could give me some solutions for this.
Thank you.

Hi bgawarammana,

It is possible to fit multilevel binary/categorical response models to data that have a structure of repeated measures within individuals. However just
like all repeated measures models the assumption is that the correlation between individual data points is equal and doesn't depend on the time between points.
There are other repeated measures models that for example allow correlation to depend on time between measurements etc. but these aren't implemented in
MLwiN aside from through macros.
I didn't write this chapter but the sentence you refer to may be in part due to the example in the chapter being a normal response model.
I'll see if I can find out who wrote the chapter and find what they meant.

Regards,
Bill.

Hi bgawarammana,
To add to the last reply. Having spoken to colleagues this comment was made in quite an old version of the manual and has persisted. The main issue with
repeated measures binary data is the assumption of normality for the individual random effects is often not satisfied by the data. Many individuals actually exhibit patterns of
all 0s or all 1s and far more than expected from a normal random effects distribution. There are a set of models called mover-stayer models that cater for such data but they are
not implemented in MLwiN.
Hope this helps,
Bill.

Mr. Bill,
Thank you very much for looking in to my problem and giving me an answer. It was really helpful.
Thank you again.