nlme: New variance function structure varConstProp
Dear R developers,
recently I have written a wishlist bug report for nlme containing a patch that
adds the variance function structure
s2(v) = t1^2 + t2^2*v^2
where v denotes the variance covariate, s2(v) denotes the variance function
evaluated at v, and t, t1 and t2 are the variance function coefficients. The
covariate can also be the fitted response.
The idea that the residual variance has an additive component and a component
proportional to the observed or fitted response is found in analytical
chemistry [2, 3], pharmacokinetics , and has recently also been introduced
to chemical degradation kinetics [5, 6].
As discussed in a recent monograph on mixed effects models [7, p. 55] and in a
manuscript dedicated to the comparison of different variants of this error
model , there are two principal possibilities to implement the idea of
combining additive and proportional error. The variant implemented in the
proposed patch assumes that the sources of the additive and the proportional
component are independent. An alternative variant would be
s2(v) = (t1 + t2*abs(v))^2
However, I have less frequently found that alternative proposal in the
literature, and I also think that it makes sense to assume independent sources
of the two components.
I think that adding the possibility to specify this error model greatly
enhances the potential of the nlme package in the area of generalised linear
models (gls), generalised nonlinear models (gnls) and, last but not least,
generalised nonlinear mixed effects models (nlme).
In addition, the patch is rather unobtrusive, as it only adds a further
variance function structure. The documentation contains an example application
using gnls(), and there is a test using nlme() and group dependent values for
t1 and t2.
Therefore, I would recommend the proposed patch for your kind consideration.