Skip to content

Help with multilevel model Poisson

4 messages · Thierry Onkelinx, Andrea Céspedes

#
Hello



I am currently working the shrinkage phenomenon in multilevel models.



I have a problem of convergence in the model when I add more random
coefficients to the models, after three coefficients,I have this error
message:



 Warning message:

In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv,  :

  Model failed to converge with max|grad| = 0.0278398 (tol = 0.001,
component 1)



I have a three-level model, with Poisson distribution, the structure is: to
my subject (rat) I register the vocalizations emitted in four specific
moments of an experiment that is repeated for three days. And I have 31
subjects. All my variables are dichotomous.



I tried several alternatives to solve that error, but the only effective
thing was to specify the optimizer = ?bobyqa? and more iterations, for my
luck it worked,

F.aleat.int <- glmer(y ~ 1+DIA2+DIA3+M1+M4+M5+M6+M9+M10+M11+ (1 |
ID_DIA:ID_SUJ) + (1+M1+M5+M11 | ID_SUJ), family=poisson, data=base,
control=glmerControl(optimizer="bobyqa",optCtrl=list(maxfun=2e4)))

But the estimates of random effects are much greater than the obtained
using other packages (MCMCglmm, glmmLasso, glmmsr and hglm),

For example, for one variable I have 2.1 and in the other packages I have
0.8

How can I explain that:

-        Due to the nature of the model? Or the optimizer as such?

-        Would you appreciate it if you could tell me how I can solve that?

-        Is it because all my variables are dichotomous?



Regards
3 days later
#
Dear Andrea,

How many observation per subject? 12? That is too few to fit a random slope
model with 10 parameters.

ir. Thierry Onkelinx
Statisticus / Statistician

Vlaamse Overheid / Government of Flanders
INSTITUUT VOOR NATUUR- EN BOSONDERZOEK / RESEARCH INSTITUTE FOR NATURE AND
FOREST
Team Biometrie & Kwaliteitszorg / Team Biometrics & Quality Assurance
thierry.onkelinx at inbo.be
Havenlaan 88 bus 73, 1000 Brussel
www.inbo.be

///////////////////////////////////////////////////////////////////////////////////////////
To call in the statistician after the experiment is done may be no more
than asking him to perform a post-mortem examination: he may be able to say
what the experiment died of. ~ Sir Ronald Aylmer Fisher
The plural of anecdote is not data. ~ Roger Brinner
The combination of some data and an aching desire for an answer does not
ensure that a reasonable answer can be extracted from a given body of data.
~ John Tukey
///////////////////////////////////////////////////////////////////////////////////////////

<https://www.inbo.be>


Op wo 9 okt. 2019 om 16:38 schreef Andrea C?spedes <ancelis.07 at gmail.com>:

  
  
#
Hi everyone

No, it?s 12 observations per day and it?s three days, so I have 36
observations per subject

Thanks


El mi?., 9 oct. 2019 a las 9:04, Thierry Onkelinx (<thierry.onkelinx at inbo.be>)
escribi?:

  
  
#
36 observations per level is still very little to fit a 10 parameter random
effect

Op wo 9 okt. 2019 18:07 schreef Andrea C?spedes <ancelis.07 at gmail.com>: