[R-sig-ME] about warning "In mer_finalize(ans) : singular convergence (7)"

[Douglas Bates]

On Fri, Feb 10, 2012 at 5:06 AM, Toni Hernandez-Matias
<ahmatias at gmail.com> wrote:
> Dear Douglas,
>
> thank you very much for your message. If necessary I can send you my data
> set. Before let me show you the output of the verbose=TRUE and the summary
> of the model results (see below).
> I wonder wether the problem is that the variance estimated for the main
> random effect (study area: coded as 'ter') takes the value of 0 (third
> column in the results of the 'verbose'). If that would be the problem, I
> wonder whether it would be acceptable (dessing) that I would ommit this
> random effect in the models and only considering the random effect of the
> transects (coded as 'trans'). On the other hand, I don't understand why this
> problem only happens with some of the independent variables (other models
> fitted fine).
>
> Thank you very much in advance,
>
> Toni
>
> mod05<-lmer(cagaders~E_arb_alt+(1|ter)+(1|ter:trans),data=conill,family=poisson,verbose=TRUE)
>
>  0:     1580.1040: 0.521157 0.212762 0.0146621 -0.00550703
>  1:     1578.7769: 0.529048 0.217137 0.0121796 -0.00198275
>  2:     1574.8963: 0.536915 0.221586 0.00963723 -0.00542672
>  3:     1546.2308: 0.659172 0.290027 -0.0296622 -2.71216e-05
>  4:     1544.4481: 0.659755 0.290395 -0.0299398 -0.00438446
>  5:     1512.2918: 0.827224 0.356975 -0.453220 0.000576404
>  6:     1506.1926:  1.14082 0.0210690 -0.431339 -0.00365397
>  7:     1500.5942:  1.16992  0.00000 -0.828777 0.00123612
>  8:     1500.3060:  1.20528  0.00000 -0.790690 0.00198426
>  9:     1499.5992:  1.25168  0.00000 -0.779210 0.000659413
> 10:     1499.3691:  1.28723  0.00000 -0.803391 0.000445819
> 11:     1499.2722:  1.32511  0.00000 -0.837827 0.000455882
> 12:     1499.2705:  1.33004  0.00000 -0.843619 0.000496015
> 13:     1499.2704:  1.33032  0.00000 -0.844199 0.000502862
> 14:     1499.2704:  1.33030  0.00000 -0.844231 0.000503448
> Mensajes de aviso perdidos

The fact that the second parameter is stuck at 0 indicates that the
random effect associated with ter is inert given that the random
effect associated with ter:trans is in the model.  You should try
fitting a model of the form

cagaders ~ E_arb_alt + (1|ter:trans)

> In mer_finalize(ans) : singular convergence (7)
>
>
>
> SUMMARY
> Generalized linear mixed model fit by the Laplace approximation Formula:
> cagaders ~ E_arb_alt + (1 | ter) + (1 | ter:trans)   Data: conill  AIC  BIC
> logLik deviance
> 1507 1525 -749.6     1499
> Random effects:
> Groups    Name        Variance Std.Dev.
> ter:trans (Intercept) 1.7697   1.3303  ter       (Intercept) 0.0000
> 0.0000  Number of obs: 648, groups: ter:trans, 66; ter, 11
>
> Fixed effects:
>             Estimate Std. Error z value Pr(>|z|)    (Intercept) -0.8442310
> 0.1839531  -4.589 4.45e-06 ***
> E_arb_alt    0.0005023  0.0029302   0.171    0.864    ---
> Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
> Correlation of Fixed Effects:
>         (Intr)
> E_arb_alt -0.247
>
>
>
>
>
>
> On Thu, Feb 9, 2012 at 9:03 PM, Douglas Bates <bates at stat.wisc.edu> wrote:
>>
>> On Thu, Feb 9, 2012 at 11:08 AM, Toni Hernandez-Matias
>> <ahmatias at gmail.com> wrote:
>> > Dear all,
>>
>> > I am trying to fit a set of models with lmer function.
>> > My aim is to investigate the relationship between the abundance of a
>> > mammal
>> > species (count) and several environmental variables.
>> > The sample size is 648, but the observations are not independent and the
>> > random effect is nested: I have 10 study areas, within each area I
>> > performed 6 transects. I have 9-10 observations in all transects. So an
>> > example of a model with a single independent variable is:
>> >
>> > mod05<-lmer(cagaders~E_arb_alt+(1|ter)+(1|ter:trans),data=conill,family=poisson)
>>
>> > When running this model I get the warning:
>> > In mer_finalize(ans) : singular convergence (7)
>>
>> Try using verbose=TRUE to determine where the parameter values are
>> going during the iterative optimization process.
>>
>> If your data could be made available, even in an anonymized form, we
>> could check the model fit against other optimizers that may be more
>> successful.
>>
>> > I don't see an apparent reason for this warning.
>> > I would be very grateful if someone can help me to solve this problem
>> > and
>> > to know wether the results in the fitted model are credible.
>> >
>> > Thank you very much in advance,
>> >
>> > Toni
>> >
>
>
> --
> *********************************************************
>
> Antonio Hernandez Matias
>
> Departament de Biologia Animal (Vertebrats)
> Facultat de Biologia
> Universitat de Barcelona
> Av. Diagonal, 645
> Barcelona      08028
> Spain
> Telephone: +34-934035857
> FAX: +34-934035740
> e-mail: ahernandezmatias at ub.edu
>
> ***********************************************************