Skip to content

Error in mer_finalize(ans) : Downdated X'X is not positive definite, 8.

2 messages · Yoav Avneon, Ben Bolker

#
Dear All,
I have conducted an experiment in order to examine predation pressure in the
surroundings of potential wildlife road-crossing structures.  I have
documented predation occurrence (binary?) in these structures and calculated
several possible explanatory variables describing the spatial heterogeneity
in several scales.  At the landscape scale I have calculated the percentage
of different land-uses (7) in buffers around the structures with changing
radii (100m, 250m, 500m, 1000m and 2000m).  For each radii I have generated
a set of all possible models from the given 7 variables related to this
radii.    
I have tried to account for random effects. A spatial random effect ? the
structure itself as a repeated measurement, by adding the term
(1|Road_Structure).  A temporal random effect ? the observation session (is
it correct to say that this examines possible learning in time ?), by adding
the term (1|Session.ord).   
I have generated 4 sets for each radii: 1) with both random effects, 2+3)
only one (each) random effect, and 4) none.    
I have tried to model the relationships using glm, glmer and lmer.
The problem is that in the full model (all 7 variables) of some sets with
random effects I get the following error:
*Error in mer_finalize(ans) : Downdated X'X is not positive definite, 8.*

The only record of this error I have found in the internet is for data with
missing observation.  This doesn't comply with my data set.  I think it
might be a singularity problem but I don't know how to check it, and as a
result, I obviously don't know how to fix it.
I have attached example R scripts (for the radius 2000) and the data table.
Any help is highly appreciated !
Thanks and all the best,
#
Yoav Avneon <Avneony <at> bgu.ac.il> writes:
Questions like this (and follow-ups) should probably go to
r-sig-mixed-models at r-project.org.

  I don't have time to dig into this in detail, but I would start
by (1) centering and scaling all of your variables (see ?scale, and
Schielzeth 2010 _Methods in Ecology and Evolution_ and (2) checking
the correlations among your variables; if they are strongly correlated
you have lots of options (none of them perfect), much discussed on
the internet, including discarding some variables or using PCA.