[R-sig-ME] Overdispersion and R2 in GLMM

[Ben Bolker]

On 16-06-30 12:40 PM, Teresa Oliveira wrote:
> Dear all,
> 
> I want to estimate RSF

  Can you tell us/remind us what RSF is?  (Remember that this is a very
general mailing list about mixed models, with readers in many different
fields)

> and to obtain the relative probabilities I will use
> the coefficients obtained with GLMMs. I am new in this so I hope I can
> express myself well.
> To construct the GLMM's, I am using glmer(). I will to use the MuMIn
> package to perform model selection [using dredge()].

  I'm getting less and less enthusiastic about this approach, but
comments will have to wait for another day.

> I have two doubts:
> 1) For the top models I get, in order to understand if they are really
> meaningful, I need to estimate R2 [r.squaredGLMM()] and to test for
> overdispersion [overdisp.glmer() with "RVAideMemoire" package], right? Is
> there anything else I must consider (besides de AIC)?

  I would generally suggest looking at graphical diagnostics (e.g. see
?plot.merMod) and plotting the predicted responses of the model along
with the original data.  R^2 values are not bad, but they're very much
harder to interpret in a (G)LMM context.

https://tinyurl.com/glmmFAQ#how-do-i-compute-a-coefficient-of-determination-r2-or-an-analogue-for-glmms

(the same base URL also has more information about computing overdispersion)

> 
> 2) How do I interpret the results in both tests?
> For a model with all variables I wanted to include (so, before performing
> model selection), I estimated R2 and overdispersion.
> 
> For overdispersion I got this:
> "> overdisp.glmer(lm_set3)
> Residual deviance: 8537.397 on 32658 degrees of freedom (ratio: 0.261)"
> Which value for ratio is acceptable?

  What kind of model are you fitting?
  In general overdispersion values (a) are only relevant
> 
> For R2 I got this:
> "> r.squaredGLMM(lm_set3)
> The result is correct only if all data used by the model has not changed
> since model was fitted.
>         R2m         R2c
> 0.006139516 0.788967246 "
> Is it normal to get values so different? Should I consider both?

  This means that your fixed effects explain almost nothing, but the
combination of fixed and random effects explains almost everything.
Does seem a little concerning.

  Explore your data graphically!

  cheers
    Ben Bolker

> 
> Thank you very much for your time and help!
> 
> Best regards,
> Teresa
> 
> 	[[alternative HTML version deleted]]
> 
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>