[R-sig-ME] glmmadmb instead of glmer to avoid convergence warnings in lme4

[Ben Bolker]

  The convergence warnings are false positives, as much discussed on the
list: see https://github.com/lme4/lme4/blob/master/README.md

  Poisson or NB models are not completely silly for modeling a success
scale, but they may not be appropriate either -- depends on what your
data actually look like.  If you have enough data, an ordinal model
could be appropriate.

On 14-06-27 07:25 AM, Lisa Rummel wrote:
> Dear mixed models experts,
> 
> I'm working with a dataset to analize the success of translocation
> programs of animal species. For this purpose, the success of various
> programs has been evaluated at a scale from 0 to 10. Now we want to
> investigate, in which way different factors like the duration of the
> program or the number of released animals, influence the success
> rate.
> 
> We´re using a linear mixed-effects model (package lme4 1.1-6) with
> "success" as response variable, "duration" in years as predictor
> variable and species as random effect. But when I run the model, I
> obtain the following convergence warning:
> 
> 
>> model10 <- glmer(succes~duration+(1|species), family=poisson)
> Warning message: In checkConv(attr(opt, "derivs"), opt$par, ctrl =
> control$checkConv,  : Model failed to converge with max|grad| =
> 0.0136827 (tol = 0.001)
> 
> 
> In order to avoid the convergence failure, we changed the
> distribution into a negative binomial distribution, using a glmmadmb
> model.
> 
> 
>> model11 <- glmmadmb(succes~duration+(1|species), data,
>> family="nbinom") summary(model11)
> 
> Call: glmmadmb(formula = succes ~ duration + (1 | species), data =
> data, family = "nbinom")
> 
> AIC: 183.8
> 
> Coefficients: Estimate Std. Error z value Pr(>|z|) (Intercept)
> 1.944321   0.123230   15.78   <2e-16 *** duration    -0.000988
> 0.013070   -0.08     0.94 --- Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01
> ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
> 
> Number of observations: total=33, species=21 Random effect
> variance(s): Group=species Variance StdDev (Intercept)  0.01666
> 0.1291
> 
> Negative binomial dispersion parameter: 15.451 (std. err.: 24.659)
> 
> Log-likelihood: -87.8784
> 
> 
> Our question is, if using the glmmadmb model instead of the glmer is
> an applicable solution for the convergence failure in this case?
> Also, we're not quite sure if we can use the poisson/negative
> binomial distribution for the variable "success", because it`s
> limited to discrete values from 0 to 10. Perhaps there is kind of a
> truncated distribution we could use?
> 
> Thank you in advanced, Lisa
> 
> _______________________________________________ 
> R-sig-mixed-models at r-project.org mailing list 
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>