[R-sig-ME] Nested longitudinal

[Ben Bolker]

Robert Kushler <kushler at ...> writes:

> 
> 
> I'm breaking my own rules by offering answers before getting answers to my own
> questions.  

  Hmmm.  I didn't see any outstanding unanswered questions from you 
on the list -- did I miss some?  Or have you not asked them yet?

> 1) The simple general answer to your basic question is that the "/"
> character is used to represent nested factors, while a "*" is used
> to indicate "crossed" factors that might interact.  Your use of "+"
> in the model formula below means you are *assuming* that the factors
> do not interact - and assuming something doesn't make it true.

  [snip]

>  4) It seems to me that there will be very strong serial correlation
> in the 15 measurements on an individual subject.  Unfortunately lmer
> doesn't include this as a modeling option.  Your current syntax does
> one of the following: (a) fits a linear "time effect" with a random
> slope and intercept or (b) if time is a factor you are trying to
> estimate an "unstructured" (sorry, Doug Bates) 15 by 15 matrix.
> Both approaches are problematic here.  I suggest you collapse the
> data by time and record y = number of minutes (out of 15) spent in
> light (or dark - doesn't matter) areas, and then use "cbind(y,15-y)"
> as the response.  You probably should also try some alternatives to
> the binomial family (e.g., "quasibinomial").

   lmer doesn't include it at least in part because it's difficult to
fit into a conditional GLMM framework -- the most sensible way to
define this would be to allow a per-observation random effect (which
would then also make the binomial overdispersed by definition), and
then specify that the individual-level random effects were themselves
temporally correlated.

  Usually when you find packages that can incorporate temporal
correlation in a GLMM they either require that you specify the full
statistical model yourself (WinBUGS/JAGS, AD Model Builder), *or* they
are in some sense marginal models (glmmPQL in the MASS package [I
know] and ASREML [I think] allow 'R-side' structures such as temporal
correlation in GLMMs, but they use penalized quasi-likelihood for
estimation, which may under some circumstances be problematic).

  For what it's worth you can't use quasi- families in glmer(); you
can either add an individual-level random effect, or use MASS::glmmPQL
if you want quasi- (see http://glmm.wikidot.com/faq for more
discussion of overdispersion in GLMMs).

 
> Regards,   Rob Kushler
> 
> On 4/24/2012 12:40 AM, arun wrote:

 [snip]

> > A brief introduction about the work: It is a light/dark choice
> > test conducted in insect larvae.  The response is binary (0-
> > present in dark area, 1-present in light area) and the experiment
> > is run for 15 min, so there are 15 measurements per individual
> > larva at 1 min intervals.  The factors which affect this study are
> > Strain (2 levels-G and S), wavelength of light (4 levels-blue,
> > green, UV, red), and starting response at 0 min (two levels-
> > animal present in dark-D or light-L).  This is how I think it is
> > nested.  Strain nested inside Wavelength, Subject (individual)
> > nested within strain, Starting response within subject, and time >
> > within Starting response.  The data looks like this:

[snip]

> >
> >
> > (fm2<-lmer(Response~Wavelength+Startingresponse+Strain+ time +
> (time|Subject),family=binomial, data=Behavdat))
> >
> > I am not sure how to specify the nested structures
> >   within the
> >   model.