www.cmm.bristol.ac.uk/forum

Hello

I am struggling to get my head around a multi-level model for my data - and whether I should even be using one. (I could use linear regression but it would mean key insights I want to look for would be lost.)

I have panel wave data (two time points of c.500 individuals). My research question relates to behavioural 'spillover'. I want to know whether a change in one (any) behaviour between the two time points is related to change in other behaviour(s). I also have individual-level predictors in the form of people's attitudes and values.

Though the data is not relating to health, it is more intuitive to explain in these terms (I substitute health behaviours below for those I'm actually measuring):

I have measures for frequency of drinking, smoking, running, sleeping, meat consumption (ordinal data, ten levels) at the two time points. I am interested to know whether:

(i) change in a specific behaviour over time is related to change in any or all the other behaviours; e.g. to what extent is change in smoking behaviour associated with change in drinking behaviour?
(ii) the more general extent to which behaviours change *independently* or *together*; e.g. does change in frequency of health behaviours happen in isolation or do they tend to move 'in tandem'.
(iii) to what extent do individual, time, and 'behavioural' factors determine how a behaviour or behaviours change?

I have spent a lot of time trying to work out how this would make sense, and feel that a multi-level model might be best, but have hit a brick wall and I can't quite figure it out.

I am thinking that I could use a three-level model, with behaviours (c.7 of these) at level 1, time at level 2 (2 time points), individuals at level 3 (n=500). That way it would be possible to say something about the extent to which individual measures (e.g. attitudes) predict behaviour change, the extent to which change in time does, and the extent to which an individual/time-independent association between behaviours does.

I am attempting to use MLwiN to do this, but when it comes to specifying the model in terms of x's and y's I am flummoxed. I saw a paper that did something (kind of) similar, but there were not details enough to see how they worked this out. The paper is doi: 10.1136/jech.2004.025742

Any tips on the conceptual structure that could be used to model this data and/or how to do this in MLwiN would be massively appreciated!

Thanks, Stuart

Hi Stuart,
Some thoughts. As you are looking at change and you only have 2 time points then I would have thought you would be considering the baseline as a predictor variable and the second time point as the response thus removing your time point level. As for your behaviours I would suggest that here a multivariate modelling approach would be more appropriate so that you can capture the differing variability and correlations between individuals. This can be done using multilevel level modelling which in essence would as you suggest have behaviours nested within individuals by using clever ways of parameterising the model to capture correlations etc. Take a look in the user grade at the multivariate modelling chapter - that looks at reading and written scores and instead think of these as behaviours.
Hope this helps,
Best wishes,
Bill.

Hi Bill

Thanks for your response to my question (some months ago now, I realise!)

A multivariate model, as you suggest, seems to be the best approach. The worked example in the MLWiN manual provides a useful analogy.

I'll proceed with this and hopefully will be able to work out an appropriate multi-level model.

Best wishes

Stuart