[R-sig-ME] syntax for indicating fixed covariates in nlmer ?

Sebastian P. Luque spluque at gmail.com
Tue Mar 31 15:10:04 CEST 2009

On Wed, 11 Mar 2009 07:30:38 -0500,
Douglas Bates <bates at stat.wisc.edu> wrote:


> I think Laurent and Dieter might have been asking a different
> question.  In a nonlinear mixed-effects model the nonlinear model
> function incorporates covariates and "nonlinear model parameters".
> For example, the four-parameter logistic model is a function of a
> covariate, often something like log-dose, and four parameters that
> could be the asymptote on the left, the asymptote on the right, the
> midpoint and a scale parameter.

> example(SSfpl)

> produces a plot showing these.

> I think the question was how to express an NLMM in which one of these
> parameters, say the midpoint, xmid, incorporates the effect of a
> covariate like treatment group.

> If that is the question then the answer is "not easily at present".
> The development version of nlmer, in the branches/allcoef section of
> the SVN repository, has the capability of doing that.  That's the good
> news.  The bad news is that the syntax of the formula has changed a
> bit and I don't want to release the development branch until I can
> resolve problems with GLMMs in that branch.

> I am having some bizarre problems with GLMMs there - the sort of
> problem that will seem trivial once I know the answer but right now is
> very frustrating.  For some reason the current code fits Poisson GLMMs
> like a charm and diverges on Bernoulli GLMMs and binomial GLMMs.

> What I will do is to polish up the documentation of the nlmer function
> over the next few days so the brave (or foolhardy, depending on your
> point of view) user can fit those models.  Then I will try to get the
> binomial GLMMs happy again.

To make sure I understand the situation, is this equivalent to the case
where an nlme Gompertz growth model (3 parameters: Linf, b, and k) was
specified as:

nlme(y ~ Linf * exp(-b * exp(-k * x)), data=growthdata,
     fixed=Linf + b + k ~ 1, random=Linf ~ 1 | population,
     start=c(Linf=400, b=0.9, k=0.1))

so currently we can't do this in nlmer?



More information about the R-sig-mixed-models mailing list