[R-sig-ME] corARMA in lme

[Ben Bolker]

Arnaud Mosnier <a.mosnier at ...> writes:

> 
> Dear mixed-models useRs,
> 
> I made a mixed model of the form
> 
> lme(Distance ~ season, random = ~1|yearTagged/tag, correlation =
> corAR1(form= ~Time|yearTagged/tag), data = dat)
> 
> or
> 
> lme(Distance ~ season, random = ~1|yearTagged/tag, correlation =
> corARMA(form= ~Time|yearTagged/tag, p=1, q=1), data = dat)
> 
> In the AR1 model, the Phi value obtained for the corAR1 structure is 0.
> In the ARMA(1,1) model, both Phi1 and Theta1 are equal to 0.
> 
> I find this strange. I believed first that it was caused by the fact that
> yearTagegd and tag variables were not considered as factor, but it did not
> change anything if I convert them to factor.
> The temporal autocorrelation is not obvious in the residuals of the model
> without a temporal structure included (i.e. without corAR1 or corARMA), but
> can I conclude that there is not correlation based on Phi value ?

  A value of exactly zero does seem a little odd.  I don't have an
immediate answer, but in your position I would (1) try specifying
starting values (i.e., test whether lme simply got stuck immediately:
I suspect that 0 is the default starting value); (2) see about turning
on 'verbose' or 'trace' flags; (3) simulate some data with an
appropriate structure that does contain an AR term and see whether you
can recover it with this model fitting approach ...