[R-sig-ME] confusing error message

[Ben Bolker]

Victoria Buller <Victoria.Buller at ...> writes:

> Can someone explain to me what this error message means? -
> 
> () : Downdated VtV is not positive definite
> 
> I keep getting it after using this code for my data-

> > m1<-glmer(a.martiensseni~distance.tree+herbangular.trans+
    shrubangular.trans+canopyangular.trans+soilangular.trans+
    leaflitangular.trans+sandangular.trans+woodangular.trans+
    waterangular.trans+cobblesangular.trans+rocksangular.trans+
    archaangular.trans+archbangular.trans+archcangular.trans+
    archdangular.trans+archeangular.trans+alt+slope+start.time+
    (1|site)+(1|location),family=binomial)

> I get that it is to do with my random effects... but I don't
> understand why there is a problem. 'Site' is just an area known as
> A,B,C,D or E and location is the location within the site the frog
> was found, so either transect 1,2, or 3. I am trying to analyse frog
> presence or absence data with habitat characteristics...  can anyone
> help????

  Site is just on the edge of having enough levels to estimate
a random effect reliably, and transect has not enough (the rule
of thumb is a *minimum* of 5-6: see e.g. http://glmm.wikidot.com/faq ).
If you have three transects within each site, I suspect you want
(1|site/location) or (equivalently) (1|site)+(1|site/location).

  I hope you have a pretty big data set: one rule of thumb is
that you need the effective size of the data set to about
10 times the number of parameters; for binary data 'effective size'
is (approximately) the minimum of the number of presences and
the number of absences, so the minimum of # presences/# absences
should be at least 200 or so ...

   Frank Harrell's book is a good reference for what to do
if you have more predictors than your data can support.

  Ben Bolker