[R] lmer: Estimated variance-covariance is singular, false convergence

[Bert Gunter]

Your data do not support your model complexity. You need to simplify the
model. If you do not know what this means, I suggest you consult with a
local statistician.

Bert Gunter
Genentech Nonclinical Statistics


-----Original Message-----
From: r-help-bounces at r-project.org [mailto:r-help-bounces at r-project.org] On
Behalf Of Martin Klipp
Sent: Wednesday, February 13, 2008 2:06 PM
To: r-help at r-project.org
Subject: [R] lmer: Estimated variance-covariance is singular,false
convergence

Dear R Community!

We analyse the impact of climbing activity on cliff vegetation. During
our fieldwork, we recorded 90 Transects in 3 climbing sites. The aim is
to see, if the plant cover (response: Cover) is influenced only by
crevice availability (predictor: Cracs), or, additional, by the distance
to the climbing route (predictor: Distance). Six plots are nested within
one Transect (ID.Transekt), the Transects are nested within the climbing
site (Site).

I tried to fit two linear mixed models with lmer(), and to compare them
by anova():


 > cracks.lmer <- lmer (Cover ~ Cracks + Distance + (Cracks + Distance |
 > + Site / ID.Transekt) , method = ("ML"))

Warning message:
In .local(x, ..., value) :
   Estimated variance-covariance for factors 'ID.Transekt:Site', 'Site'
is singular


 > cracks.lmer1 <- lmer (Cover ~ Cracks + (Cracks | Site / ID.Transekt) ,
 > + data = Lmm.spalten.corr, method = ("ML"))

Warning messages:
1: In .local(x, ..., value) :
   Estimated variance-covariance for factor 'Site' is singular

2: In .local(x, ..., value) :
   nlminb returned message false convergence (8)

For this one, i changed the maxIter and msMaxIter up to 50000000, but
the Warning messages remained the same.
I have many "0" values for "Cover" and "Cracks" and tried to exclude
some of them in different ways (as all transects without vegetation
cover), but this did not help. The current version deals with all
transects wich have at least two plots with crevices.
The correlation Intercept/Cracks ist -1.000 for both models.
I attachet a symmary of the used data.frame, a randomly selected subset
and the summarys of cracks.lmer and cracks.lmer1 in a *.txt file
(modified with kate). Some xyplots are attached as *.png. (The error
message on the plots means "Error using packet 5, NA/NaN/Inf in externem
Funktionsaufruf", excluding these transects does not change the lmer
message.)

My question is: Is it possible and does it make sense to fit these date
with lmer? I do get outputs, but they dont seem really reliable to me.

I hope that this is not a stupid and unnecessary question, but i did't
really find a answer in the current mailing-lists. Thanks a lot!

Martin Klipp

R version 2.6.2 (2008-02-08)
i486-pc-linux-gnu
Kubuntu 6.06.1, precompieled packages from the CRAN-mirror.
lme4_0.99875-9
Matrix_0.999375-4
lattice_0.17-2
nlme_3.1-86