[R-sig-ME] [FORGED] warning error question

Ben Bolker bbo|ker @end|ng |rom gm@||@com
Tue Mar 24 01:50:50 CET 2020



On 2020-03-23 6:33 p.m., Rolf Turner wrote:
> 
> On 24/03/20 3:41 am, Anahí Fernández wrote:
> 
>> hi!! I run this model in lme4:
>> "M.4=glm(Cuenta~carga*categ.asoc+(1|campo/foto)
>>            +offset(log(area.foto)),family=poisson(link =
>> "log"),data=tipocat)"
>> And I have this warning message: "Warning message:
>> In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv,  :
>>    Model failed to converge with max|grad| = 0.00432818 (tol = 0.001,
>> component 1)"
>> I don´t know what is that means, could you help me, please!!
>> My datatable is attached...
>>
>> Cheers, Anahí
> 
> (a) Since the function you invoke is glm() this would appear to be
> off-topic for r-sig-mixed-models.  OTOH your formula does indeed seem to
> involve random effects.  Did you *really* call glm()?  Or did you
> actually call glmer()?  If so you, you should be ashamed of yourself for
> such sloppiness in posing your question.  People are providing help out
> of the goodness of their hearts; don't impose on their good nature by
> expecting them to be telepathic.

  Rolf, can you tone it down slightly? I agree that the OP could be more
careful, but "you should be ashamed of yourself" seems way too strong.

> (b) Assuming that you really did call glmer() --- my impression is that
> such warnings are usually false positives and may usually (???) be
> safely (???) ignored.  However I'm no expert; you should perhaps wait
> for confirmation of this from the more knowledgeable.
> 
> (c) Your "datatable" was *NOT* attached.  Most attachments get stripped
> by the system (for security reasons).  There are exceptions.  *READ* the
> posting guide, which you appear not to have done.

  I did get the data from a previous interchange (Anahí, can you post
the data set somewhere publicly accessible?  CSV is strongly preferred
to XLSX ...).

  The bottom line here is that your baseline category has only a single
'Cuenta' value in it and only two unique 'carga' values, leading to
extreme estimates - this is essentially the analogue of 'complete
separation' in the logistic regression, and has the same solutions
(regularize somehow if you want sensible answers).

  cheers
   Ben Bolker


            Cuenta
categ.asoc     1   2   3   4   5   6   7   8   9
  highly      10   0   0   0   0   0   0   0   0
  isolated    78  20   8   4   1   0   1   0   0
  moderately  58   0   1   0   0   0   0   0   0
  poorly     120  47  24  16  12   9   0   1   1

round(coef(summary(M.4)),3)
                           Estimate Std. Error z value Pr(>|z|)
(Intercept)                  -2.852      1.264  -2.257    0.024
carga                         8.029     13.626   0.589    0.556
categ.asocisolated            1.776      1.255   1.415    0.157
categ.asocmoderately          1.299      1.256   1.034    0.301
categ.asocpoorly              1.929      1.251   1.542    0.123
carga:categ.asocisolated     -9.548     13.612  -0.701    0.483
carga:categ.asocmoderately   -9.418     13.614  -0.692    0.489
carga:categ.asocpoorly       -9.006     13.615  -0.661    0.508

> 
> cheers,
> 
> Rolf Turner
>



More information about the R-sig-mixed-models mailing list