[R] Heteroscedasticity in a percent-cover dataset

[Cade, Brian]

Quit trying to eliminate heteroscedasticity in your data - there is
information there in the pattern of changing variances.  I would suggest
instead that  you go directly after modeling the change in entire
distributional form by using quantile regression (package quantreg).  So,
for example, depending on your sample size and model complexity you might
estimate 0.05 to 0.95 quantiles by increments of 0.05 to categorize how the
distribution of percent cover changes conditional on your predictor
variables.  Heterosecdasticity will be accomodated by changing coefficients
(slopes) for some of your predictors for different values of tau denoting
the quantiles.  See Cade and Noon (2003) for a good introduction for
ecologists.  For percent or proportion data you can use a simple logit
transformation of the dependent variable (see Bottai et al. 2010) to keep
the response bounded appropriately, and incorporate predictor variables any
way you would in any other linear (or generalized linear) model.  There
even are mixed-effects versions of quantile regression now (package lqmm)
but I haven't used them enough to speak to their veracity and value.

Brian

Brian S. Cade, PhD

U. S. Geological Survey
Fort Collins Science Center
2150 Centre Ave., Bldg. C
Fort Collins, CO  80526-8818

email:  cadeb at usgs.gov <brian_cade at usgs.gov>
tel:  970 226-9326


On Thu, Apr 14, 2016 at 11:49 PM, Lai Wen Ya Samantha <s.lai at u.nus.edu>
wrote:

>
>
> Hi,
>
> I am currently trying to do a GLMM on a dataset with percent cover of
> seagrass (dep. var) and a suite of explanatory variables including algal
> (AC) and epiphyte cover (EC), rainfall, temperature and sunshine hours.
>
> M2=glmer(SG~AC+EC+TP+SS+RF+(1|Location/fSi/fTr),
> family=binomial,data=data,nAGQ=1)
>
> As the dependent variable is percent cover, I used a binomial error
> structure. I also have a random effect due to nested of the data collection
> strategy. However, I keep getting heteroscedasticity issues as shown in the
> image below. I have tried using an arcsine transformation (with a lme), but
> the scatter of residuals are still very much similar.
>
> What else can I do to try to resolve the heteroscedasticity in my data? Any
> help will be very much appreciated!
>
>
> <http://r.789695.n4.nabble.com/file/n4719735/Heteroscedasticity.png>
>
>
>
>
> [http://r.789695.n4.nabble.com/file/n4719735/Heteroscedasticity.png]
>
>         [[alternative HTML version deleted]]
>
> ______________________________________________
> R-help at r-project.org mailing list -- To UNSUBSCRIBE and more, see
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide
> http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.
>

	[[alternative HTML version deleted]]