[R-sig-ME] Non-significant fixed effect allows approximation of variance-covariance matrix
bbolker at gmail.com
Sun Mar 13 18:08:49 CET 2016
Hard to say in general without looking at the data in detail.
Perhaps the Geomorphic information helps explain some of the variation
in the response that is not otherwise well explained (e.g., leaving out
the Geomorphic information makes the site-level BLUPs/conditional modes
less well approximated by a normal distribution). This is not
necessarily inconsistent with a non-significant effect overall ...
I would (1) plot your data to see if there are any outliers or
patterns that might not be getting picked up by the model; (2) consider
glmmLasso, if you really want to use the model for prediction; (3) try
the model in lme4, which is a tiny bit more stable.
heterogeneous variance models are not built in to lme4; you can make
it happen for *each* site by adding an observation-level random effect
that only applies at one level of a site, by adding an observation-level
my_data$obs <- factor(seq(nrow(my_data))
and then adding
... but this is tedious when there are many sites.
Can you address heteroscedasticity by transformation instead ...?
How many sites and how many total observations in your data set ... ?
On 16-03-12 08:33 PM, Jacob Bukoski wrote:
> Dear all,
> I am using lme() to run a mixed effects model on soil carbon observations,
> with fixed effects specified for latitude (continuous), and dominant genera
> of tree (factor with five levels), and random effects specified for site.
> The data is heteroscedastic, which I can account for well with a varIdent
> weights specification; however, when I do so I receive a "non-positive
> definite approx. var-covar matrix" output from the call to $apVar.
> When I add in a third fixed effect (Geomorphic setting, a factor with three
> levels), the variance-covariance matrix is approximated successfully, but
> the levels of the added third fixed effect are non significant.
> I've been trying to read up on why this might be occurring, but can't for
> the life of me figure out why a more complex model (including
> non-significant predictors) would allow for the successful approximation of
> the variance-covariance matrix.
> I'm hoping to use the model for predictive purposes, and ideally would not
> include non-significant effects in its final form.
> Does anyone have any ideas on why this might be occurring, or
> intermediate-level resources per non-positive definite variance-covariance
> matrices that I could look into?
> Many kind thanks,
> P.S. If it helps, the model specification is here:
> *lme.C.density <- lme(C.density ~ Latitude + Genus + Geomorph,
> random=(~1|Site), weights=varIdent(form=~Genus|Site), data=model.c.dens,
> method = "REML")*
More information about the R-sig-mixed-models