[R-sig-ME] Need help on zero-inflated beta model and post-hoc test

[Yiru Cheng]

Hi, Ben

Thanks for your clarification! The raw residuals for the simulated data
don't look normal as you said. Since the expected and observed residuals
have no sig. deviation, I guess I can go ahead and do the post-hoc test
using this model.

Yiru


>
>
> Message: 1
> Date: Sun, 2 Feb 2020 16:08:36 -0500
> From: Yi-Ru Cheng <yc2975 using columbia.edu>
> To: r-sig-mixed-models using r-project.org
> Subject: [R-sig-ME] Need help on zero-inflated beta model and post-hoc
>         test
> Message-ID:
>         <CAOmDo73Fv5Ew0ZtqXUQhTjv5N==Jty=
> aEYfdM5n2ULQ_OVj6og using mail.gmail.com>
> Content-Type: text/plain; charset="utf-8"
>
> Hi, everyone
>
>
> I would like to compare the relatedness of individuals among plots. Since
> relatedness is bound between 0 and 1 and most of them are zeros, I am
> trying to fit with the zero-inflated beta model in the gamlss and glmmTMB
> packages. After fitting the model, I want to perform the post-hoc test
> between different plots. However, I have some conundrums here.
>
>
> 1.  gamlss
>
> The beta model in the package gamlss gives a pretty good residual plot, but
> the post-hoc tests in emmeans and multcomp don't work with the model...Here
> is my model.
>
>
>
> m1 <- gamlss(r ~ plot + random(year), family = BEINF, data = df)
>
> emmeans(m1, pairwise ~ plot, type="response")
>
> hist(resid(m1))
>
>
>  "Error in emm_basis.gamlss(object, trms, xlev, grid, misc = attr(data,  :
>
>   gamlss models with smoothing are not yet supported in 'emmeans'"
>
>
>
> 2. glmmTMB
>
> I then tried the beta family in glmmTMB. For some reason, the simple
> residual plot looks pretty skewed, but it looks alright with the scaled
> residual plot in DHARMa. Otherwise, the post-hoc test in emmeas works fine.
>
>
>
> m2 <- glmmTMB(r ~ plot + (1|year), ziformula =~ 1,  data=df,
> family=beta_family(link="logit"))
>
> emmeans(m2, pairwise ~ plot, type="response")
>
> hist(resid(m2))  #skewed
>
> DHARMa::simulateResiduals(m2)  ⇒ no significant deviation
>
>
>
> I’m not quite sure why the residual plots give different patterns in two
> models.  Is it not correct to look at the residual distribution for
> diagnosis? Could I see it’s a green light to use the model in glmmTMB based
> on the simulated residuals in the DHAMRMa package even though the simple
> residual plot looks skewed? Or is there any other post-hoc test designed
> for gamlss with the smoothing process? Any suggestions would be
> appreciated. Thanks in advance!
>
>
> Best,
>
> Yiru
>
>         [[alternative HTML version deleted]]
>
>
>
>
> ------------------------------
>
> Message: 2
> Date: Sun, 2 Feb 2020 19:57:06 -0500
> From: Ben Bolker <bbolker using gmail.com>
> To: r-sig-mixed-models using r-project.org
> Subject: Re: [R-sig-ME]  Need help on zero-inflated beta model and
>         post-hoc test
> Message-ID: <e7a36b0d-339b-5564-065c-c529d7084c0d using gmail.com>
> Content-Type: text/plain; charset="utf-8"
>
>
>
> On 2020-02-02 4:08 p.m., Yi-Ru Cheng wrote:
> > Hi, everyone
> >
> >
> > I would like to compare the relatedness of individuals among plots. Since
> > relatedness is bound between 0 and 1 and most of them are zeros, I am
> > trying to fit with the zero-inflated beta model in the gamlss and glmmTMB
> > packages. After fitting the model, I want to perform the post-hoc test
> > between different plots. However, I have some conundrums here.
> >
> >
> > 1.  gamlss
> >
> > The beta model in the package gamlss gives a pretty good residual plot,
> but
> > the post-hoc tests in emmeans and multcomp don't work with the
> model...Here
> > is my model.
> >
> >
> >
> > m1 <- gamlss(r ~ plot + random(year), family = BEINF, data = df)
> >
> > emmeans(m1, pairwise ~ plot, type="response")
> >
> > hist(resid(m1))
> >
> >
> >  "Error in emm_basis.gamlss(object, trms, xlev, grid, misc = attr(data,
> :
> >
> >   gamlss models with smoothing are not yet supported in 'emmeans'"
> >
> >
> >
> > 2. glmmTMB
> >
> > I then tried the beta family in glmmTMB. For some reason, the simple
> > residual plot looks pretty skewed, but it looks alright with the scaled
> > residual plot in DHARMa. Otherwise, the post-hoc test in emmeas works
> fine.
> >
> > m2 <- glmmTMB(r ~ plot + (1|year), ziformula =~ 1,  data=df,
> > family=beta_family(link="logit"))
> >
> > emmeans(m2, pairwise ~ plot, type="response")
> >
> > hist(resid(m2))  #skewed
> >
> > DHARMa::simulateResiduals(m2)  ⇒ no significant deviation
>
>    Quick answer: I don't think there's any reason to expect the
> residuals of a Beta GLMM *not* to be skewed.  Try simulating a simple
> example (you may even be able to use simulate() on your fitted model) to
> see for yourself.  DHARMa does a plot of expected vs actual
> distributions of residuals, not a raw plot of residuals.
>
>
> >
> >
> >
> > I’m not quite sure why the residual plots give different patterns in two
> > models.  Is it not correct to look at the residual distribution for
> > diagnosis? Could I see it’s a green light to use the model in glmmTMB
> based
> > on the simulated residuals in the DHAMRMa package even though the simple
> > residual plot looks skewed? Or is there any other post-hoc test designed
> > for gamlss with the smoothing process? Any suggestions would be
> > appreciated. Thanks in advance!
> >
> >
> > Best,
> >
> > Yiru
> >
> >       [[alternative HTML version deleted]]
> >
> > _______________________________________________
> > R-sig-mixed-models using r-project.org mailing list
> > https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
> >
>
>
>
>
> ------------------------------
>
> Subject: Digest Footer
>
> _______________________________________________
> R-sig-mixed-models mailing list
> R-sig-mixed-models using r-project.org
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>
>
> ------------------------------
>
> End of R-sig-mixed-models Digest, Vol 158, Issue 3
> **************************************************

	[[alternative HTML version deleted]]