[R] nlme w/no groups and spatially correlated residuals

Ben Bolker bolker at ufl.edu
Fri Feb 12 04:19:18 CET 2010

Seth <sjmyers <at> syr.edu> writes:

> I would like to specify a spherical correlation structure for spatially
> autocorrelated residuals in a model based upon the logistic function of a
> response that is a proportion (0 to 1) (so usual binary logistic regression
> is not an option).  There is no need for a g-side random effect with
> grouping in this model. Am I correct that nlme requires this (meaning a
> correlated error structure only is not permissible)?  I have tried to
> replicate the 'abuse' of the lme function I've seen for similar problems
> (specifying that all observations belong to one group), but this does not
> seem to work for nlme.  Any legitimate work arounds?

  The proportion part might be a bit tricky (you can only 
reasonably assume that the variation is normally distributed
if the variance is pretty small), but I think gnls() is what you're
looking for if you want non-linear least squares with correlation
and/or heteroscedasticity but without random effects.

  Ben Bolker

More information about the R-help mailing list