[R-sig-ME] Multivariate Multilevel Mixed-Effects Model

[Malcolm Fairbrother]

Dear Eiko,

You can certainly conduct this analysis in R, though some aspects of what you want to do are not simple. You might do best to start with some general reading/self-study re. mixed/multilevel models (natural scientists and statisticians tend to refer to "mixed" models, social scientists to "multilevel" or "random effects" or "hierarchical linear" models). Alternatively, if your uncertainty is less about such models generally than about R as a specific tool/software environment, the suggestions below may get you started, and http://glmm.wikidot.com/ may be a good resource as well (even though it's aimed at biological scientists).

R packages that may be useful to you are lme4, MCMCglmm, and nlme--the latter because it would allow you to check for temporal autocorrelation, which may be a concern particularly if your measurement occasions are close together in time. There are helpful coursenotes and worked examples for the first two packages on the web, and for nlme the best resource to start with would probably be the book "Mixed Effects Models in S and S-Plus", by Pinheiro and Bates.

There are ways of modelling multiple outcomes simultaneously (see the MCMCglmm documentation on Multi-response models at http://cran.md.tsukuba.ac.jp/web/packages/MCMCglmm/vignettes/CourseNotes.pdf), but nine is a lot. Perhaps an exploratory factor analysis could help you reduce the nine items to a smaller and more manageable number? There are also techniques common in psychology that may be appropriate for what you want, though I am less familiar with those (see e.g., http://www.jstatsoft.org/v20/a02/paper).

Hope that's useful.
- Malcolm


> Date: Fri, 13 Apr 2012 11:32:14 +0200
> From: Eiko Fried <torvon at gmail.com>
> To: r-sig-mixed-models at r-project.org
> Subject: [R-sig-ME] Multivariate Multilevel Mixed-Effects Model
> 
> Hello.
> 
> I have a problem that I have not been able to solve within the last months.
> Maybe R offers options to tackle this.
> 
> My dataset:
> 
> * N = 1000, 5 measurement points
> * 9 categorical dependent variables (depressive symptoms, scored 0, 1, 2 or
> 3) that are intercorrelated
> * 7 time-varying covariates that are dichotomous life events (yes/no) at
> each measurement point (people can have multiple life events, therefor this
> is not just one categorical with 0=no life event, 1=life event1, 2=life
> event2 etc.)
> * A bunch of time-invariant baseline predictors (e.g. neuroticism, early
> family environment, genotype, gender, family history of depression ... ),
> measured only once at the first measurement point.
> 
> My core question is whether different life events lead to different
> depressive symptom profiles (controlling for baseline variables).
> 
> I can confirm this hypothesis in nine univariate repeated measurement
> mixed-models (some life events are predictors for some symptoms but not
> others, and vice versa), but that invites the problem of (1) not being able
> to control for multiple models, and (2) ignoring the correlated structure
> of my categorical response variables.
> 
> So I'm looking for multivariate multilevel models currently to do all of
> this in one analysis.
> Could this be achieved with R? The more palpable the recommendations the
> better, I have little experience with R.
> 
> Thank you
> Eiko