[R-sig-ME] Convergence in glmmTMB but not glmer

[Ben Bolker]

On 10/20/20 2:02 PM, Thierry Onkelinx wrote:
> Daniel sent me the data in private.
> 
> A couple of remarks on the dataset.
> - the response is non-integer. You'll need to convert it to integer 
> (total number) and use an appropriate offset term (log(nights)).
> - make sure the factor covariate is a factor and not an integer.

   If the response is non-integer, that makes my comment about 
overdispersion not necessarily relevant (check again after re-fitting).

   It's often a good idea when using an offset such as log(nights) to 
*also* (alternatively) try using log(nights) as a predictor: using 
log(nights) assumes that the number of counts is strictly proportional 
to the number of nights measured (log(counts) ~ log(nights) + <stuff> -> 
counts ~ nights*exp(stuff) , whereas using log(counts) allows for some 
saturation effects (log(counts) ~ alpha*log(nights) + <stuff> -> counts 
~ nights^alpha*exp(stuff))


> 
> Please see if that solves the problem. What happens if you use a nbinom 
> distribution as Ben suggested?
> 
> Personally, I don't like to "standardise" covariates. It makes them much 
> harder to interpret. I prefer to center to a more meaningful value than 
> the mean. And rescale it by changing the unit. E.g. Age ranges from 1 to 
> 15 with mean 6.76. I'd use something like AgeC = (Age - 5) / 10. This 
> gives a similar range as the standardisation of Age. But one unit of 
> AgeC represents 10 year. And the intercept refers to Age = 5. Making the 
> parameters estimates easier to interpret IMHO.

   Yes, although 'strict' standardization (scaling by predictor SD or 
2*predictor SD) allows direct interpretation of the parameters as a kind 
of effect size (Schielzeth 2010), whereas 'human-friendly' 
standardization trades interpretability for the comparison of magnitudes 
being only an approximation.


> 
> Best regards,
> 
> 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 using inbo.be <mailto:thierry.onkelinx using inbo.be>
> Havenlaan 88 bus 73, 1000 Brussel
> www.inbo.be <http://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 di 20 okt. 2020 om 19:40 schreef Ben Bolker <bbolker using gmail.com 
> <mailto:bbolker using gmail.com>>:
> 
>         As Thierry says, the data would allow us to give a more detailed
>     answer.  However:
> 
>         * the overall goodness-of-fit is very similar (differences of
>     ~0.001
>     or less on the deviance scale)
> 
>         * the random-effects std deve estimate is similar (2% difference)
>         * the parameter estimates are quite similar
>         * the standard errors of the coefficients look reasonable for
>     glmmTMB
>     and bogus for lme4 (in any case, if there's a disagreement I would be
>     more suspicious of the platform that gave convergence warnings)
> 
>         There's also strong evidence of dispersion (deviance/resid df > 6);
>     you should definitely do something to account for that (check for
>     nonlinearity in residuals, switch to negative binomial, add an
>     observation-level random effect ...)
> 
>          You might try the usual set of remedies for convergence problems
>     (see ?troubleshooting, ?convergence in lme4), e.g. ?allFit.  Or try
>     re-running the lme4 model with starting values set to the glmmTMB
>     estimates.
> 
>         Overall, though, I would trust the glmmTMB results.
> 
>     On 10/20/20 12:56 PM, Daniel Wright wrote:
>      > Hello,
>      >
>      > I'm having convergence issues when using glmer in lme4, but not
>     glmmTMB.
>      > I'm running a series of generalized linear mixed effect models
>     with poisson
>      > distribution for ecological count data. I've included a random
>     effect of
>      > site (n = 26) in each model. All non-factor covariates are
>     standardized.
>      >
>      > The coefficient estimates of models run in glmer and glmmTMB are very
>      > similar, but models run in glmer are having convergence issues.
>     Any advice
>      > would be appreciated, as I'm not sure if I can rely on my results
>     from
>      > glmmTMB.
>      >
>      > Attached are example of outputs from glmmTMB vs glmer:
>      >
>      >
>      > _______________________________________________
>      > R-sig-mixed-models using r-project.org
>     <mailto:R-sig-mixed-models using r-project.org> mailing list
>      > https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>      >
> 
>     _______________________________________________
>     R-sig-mixed-models using r-project.org
>     <mailto:R-sig-mixed-models using r-project.org> mailing list
>     https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>