[R-sig-ME] R-sig-mixed-models Digest, Vol 35, Issue 4

[David Airey]

Thanks for the responses.

We'll start simulations with nlme(). Once we get things working, I'll  
ask about using nlmer().

[Motivating science (skip if not interested in drug genetics!): There  
is likely a need for nlmer(), because the projected data set is 100  
curves, each curve estimated by 12 dose levels, and the context is a  
genome scan. If feasibility doesn't kill this simulation project, then  
it might be a good testing bed for nlmer(). The curves can be grouped  
(nested) in one of two possible alleles at each genetic marker along  
each chromosome. There are only two possible alleles at each marker,  
because the subjects are from a set of recombinant inbred lines  
derived from two parental standard inbred mouse strains. The twist  
here is that each curve is estimated by 24 independent mice per line,  
2 mice per dose level, rather than exposing mice to multiple doses to  
achieve a true repeated measures design. The key is that a curve per  
RI line can still be estimated with independent mice, because mice  
within RI line are isogenic (genetic clones). The environment of each  
mouse will of course be assumed equivalent, otherwise this design is  
problematic. If you have 24 mice per RI line, how do you spend them in  
the context of drug quantitative genetics? The dose response curve is  
extremely important to know. Do you put 8 independent mice in each of  
three dose levels and use a linear model? Or can you distribute the  
mice 2 mice per dose level, and use a nonlinear model such as the four  
parameter logistic function? What is typically done in drug response  
quantitative genetics with mice is to use a _single_ dose level that  
was optimized for differences among mice in a small pilot study.  
That's feasible but arguably inadequate.]

-Dave

I asked:

> "I have a project to conduct simulations for which I must consider a
> nonlinear mixed model (dose-response, 4 parameter logistic model).
> Because I have the Pinheiro and Bates book, which does such a great
> job explaining things, whereas lme4 documentation is too thin for me,
> I'd like to stick with the NLME package for now. Is that in some way
> unadvisable?"


Ben Bolker wrote:

>  Respectfully, I would say that's a big "but" in the second paragraph.
> lme4 has enormous potential, and is already much better than nlme for
> some things (crossed designs, large data sets, GLMMs), but its edges  
> are
> still rough.  Accessor methods are less available/fully
> developed/documented, and in particular it's my impression that
> nonlinear models have been less thoroughly exercised in lme4 ... Using
> lme4 would be a form of contributing back to the community
> (beta-testing), but if you're in a hurry my HO would be that sticking
> with nlme sounds like a good idea.  (Others should as always feel free
> to disagree.)

Harold Doran wrote:

> Yes and no. There are no particular problems with nlme that would  
> impede you from completing scientific work properly. With that said,  
> lmer is substantially improved from its primative cousins in many  
> ways; primarily it is more capable of estimating models with large  
> data.
>
> You are correct that documentation is thinner than what is available  
> for nlme. But, there is plently of help on this list. I'm quite sure  
> with good questions, you'll find all the support on this list you  
> could possibly need.


--
David C. Airey, Ph.D.
Pharmacology Research Assistant Professor
Center for Human Genetics Research Member
Vanderbilt University School of Medicine