[R-sig-ME] Warning message when using lmer

[Ben Bolker]

On 05/20/2011 03:30 AM, Thackeray, Stephen J. wrote:
> Hi list members,
> 
> I am currently trying to run a series of linear mixed effects models
> using the lmer function in lme4, in order to compare models with
> fixed effects with null models. The models contain a series of nested
> and a crossed random effects in order to represent the structure of
> the hierarchic sampling campaign from which the data were collected.
> The code I'm using is:
> 
> mod1<-lmer(log10(MPP_TCB_mm3.l+0.1)~log10(TP) 
> +(1|Country/WBID/StationID/SampleNumber)+(1|Analyst),REML=FALSE,data=phyto.data1)
>
> 
null<-lmer(log10(MPP_TCB_mm3.l+0.1)~1
> +(1|Country/WBID/StationID/SampleNumber)+(1|Analyst),REML=FALSE,data=phyto.data1)
>
> 
AIC(mod1,null)
> 
> However, when I run mod1 I get: Warning message: In mer_finalize(ans)
> : singular convergence (7)
> 
> I'm not familiar with this warning and so I am unsure of what to
> change to address the issue. Could anyone advise me on this? I could
> provide the data set if necessary.
> 
> All the best
> 
> Steve
> 

  This is a fairly generic, and fairly difficult, "you may have
identifiability/fitting problems due to the amount/structure of your
data" problem.

  To get started,

RSiteSearch("{singular convergence}", restrict="R-sig-mix")

  You could try simplifying your model slightly (e.g. drop one of the
levels in your 4(!)-layer nested design) and see if that helps.  (If you
get similar answers either way, and simplifying the model eliminates the
warning, then you can decide whether you're more comfortable living with
the warning or with the simplified model.)

  Ben Bolker