[R] Repeated measures using lme

[Douglas Bates]

On Thu, Mar 6, 2008 at 3:51 AM, Tom Cohen <tom.cohen78 at yahoo.se> wrote:
> Dear list,

>   I am trying to do a repeated analysis using lme in R and a little bit unsure if I have set up the right statement. The problem is the IL6 (interleukin 6) was measured 5 times on each individual in each of 6 companies. The hypotheses are to see whether there is a relationship between IL6 and the total dust in each of the companies and if there is any change in IL6 across time points. So the fixed effects are total dust, company and random effects is individual. The model would be like this

>   lme(IL6~ dust + time*company, random=~time|individual,
>                         correlation=corAR1(form=~time|individual), data=dat)

>    with time as a repeated measure.

>    the analysis in SAS would be

>   proc mixed data=dat;
>   class time individual company;
>   model IL6=dust company time company*time;
>   repeated time/Sub=individual(company) type=AR(1) r rcorr;
>   random individual;
>   lsmeans company time company*time/slice=time;
>   run;

I'm not sure that those are equivalent specifications.  If I read the
SAS code correctly (and I don't have a lot of experience with SAS) the
equivalent call to lme would have random = ~ 1 | individual

I would be more inclined to start with a model that does have a random
effect for time but does not have the additional correlation
structure.  In lme this would be as in your specification but omitting
the correlation argument.  In lmer it would be

lmer(IL6 ~ dust + time * company + (time | individual), data)

In general it is difficult to estimate both a random effect for the
slope w.r.t. time and an AR(1) correlation pattern over time within
individual.  The motivation for these two types of models is not
identical but the end results in terms of the marginal covariance
structure of the responses are similar.  Generally it is better to
stick with just one such correlation-inducing term, as you have done
in the SAS specification.

Before fitting models to such data I recommend extensive plotting to
see what types of models are indicated.  With longitudinal data
collected on individuals I would always begin with

library(lattice)
xyplot(IL6 ~ time | individual, data)

and probably enhance that by taking account of the grouping of
individuals within company by grouping the panels according to company
and ordering according to some characteristic of the individual within
company.

Let me take the opportunity to plug Deepayan Sarkar's book "Lattice:
Multivariate Data Visualization with R" due for release "real soon"
and the web site http://lmdvr.r-forge.r-project.org/ for the book.  It
is a great experience just to browse the figures for each chapter at
the web site.

I enclose a plot of the sort that I describe.  In this case there is a
single group of subjects.


>   Am I writing the right code in R that would give me the same results if doing the analysis in SAS. Also is there any command in R that does the same thing as SLICE in SAS does, to test when in time there is difference between companies?
>
>   Thanks for any help,
>   Tom
>
>
>  ---------------------------------
>  Sök efter kärleken!
>
>         [[alternative HTML version deleted]]
>
>
> ______________________________________________
>  R-help at r-project.org mailing list
>  https://stat.ethz.ch/mailman/listinfo/r-help
>  PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
>  and provide commented, minimal, self-contained, reproducible code.
>
>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: examp-sleepplot.pdf
Type: application/pdf
Size: 47127 bytes
Desc: not available
Url : https://stat.ethz.ch/pipermail/r-help/attachments/20080307/43ea3b6e/attachment.pdf