[R] "repeated"repeated measures in ANOVA or mixed model

[Mike Marchywka]

----------------------------------------
> Date: Tue, 9 Nov 2010 18:25:18 -0800
> From: djmuser at gmail.com
> To: Paul.Rheeder at up.ac.za
> CC: r-help at r-project.org
> Subject: Re: [R] "repeated"repeated measures in ANOVA or mixed model
>
> Hi:
>
> This sounds like a 'doubly repeated measures problem'. Are any treatments
> assigned to individuals or is this a purely observational study?
>
> Is the time horizon of the between-visit factor (much?) longer than that of
> the within-visit factor? You could try to assess the strength of correlation
> of measurements between visits; if it's close to zero, you might be able to
> get away with treating visit as a non-repeated measures factor, which would
> simplify the analysis.
>
> If not, then the between-visit factor is nested within subject and the
> within-visit factor is nested within visits (obviously :) within subjects,
> so you would have one within-subject correlation structure to deal with for
> visits and another for times within visits. The question comes down to how
> easily the form of the overall covariance matrix can be specified.
>
> It's within the realm of possibility that the within-visit relationship is
> nonlinear in time (as in any number of pharmacokinetic models). If the
> visits are more or less uncorrelated in time, it might be reasonable to
> combine the data over visits in the hope that leads to a better fitting
> model. In that (comparatively happy) situation, the nlmer() function in the
> lme4 package (or the nlme() function in package nlme) would be a good place
> to look.
I thought the OP question was just related to R rather than analysis but
now that you have opened it up,  I often find myself coming out against
graphics in favor or numbers but sometimes
a picture is worth a thousand words and maybe some scatterplots would let
you come up with post hoc hyphotheses to test. Non-linear doesn't mean unrelated
and, for example, you could have a parabolic or at least saturating dose response curve that 
was or was not anticipated a priori or for that matter withdrawal/rebound
effects in a time series of clinical effects even if the drug/metabolite  elimination kinetics
are monotonic in time. Personally I guess I'd be open minded, stare at
some pictures, employ the informal US FDA thought that " a p-value is no
substitute for a brain" and see what the numbers say when you start testing ideas. 
The goal of a model fit is to explain, not rationalize, the data but this
is always difficult post hoc. 

For that matter, it is not a waste of time to try to run negative controls.
That is, do you find that your measures have some periodicity that relates
to the state of your lab equipment or time of day etc. Confirmation bias
can be a problem if you get numbers you like early on. 

>
> If it is necessary for you to deal with two types of nontrivial
> within-subject correlation, then I'm not so sure that nlme/lme4 is the right
> path to follow, but I'm not a mixed model expert so I could easily be wrong
> about that and others with more expertise are welcome to chime in with
> alternatives.
>
> You might also consider sending the initial mail and follow-ups to the
> R-mixed-models list, to which you can subscribe from here if you're not a
> list member:
>
> http://www.r-project.org/mail.html
>
> Scroll down to the bottom of the web page to find the list of special
> interest groups (SIGs).
>
> Sounds like an interesting problem...
>
> Cheers,
> Dennis
>
> On Mon, Nov 8, 2010 at 10:02 PM, Paul Rheeder  wrote:
>
> > dear List
> > I have a dataset with blood measurements at 5 points in TIME
> > (0,30,60,90,120) taken on 3 VISITS (same subjects). the interest is to
> > compare these measurements between Visits, overall and at the different time
> > points.
> >
> I have problems setting up repeated measures ANOVA with 2 repeated measures
> > (VISIT and TIME) (and then doing post hoc testing) or doing it with a linear
> > mixed model ( both VISIT and TIME are repeated).
> > Any suggestions?
> > Paul
> >
> >
> > Prof P Rheeder
> > School of Health Systems and Public Health
> > Faculty of Health Sciences
> > University of Pretoria
> > Room 6:12
> > HW Snyman North
> > Tel: 012 354 1488
> > Fax: 012 354 1750
> > Mobile: 082 779 3054
> >