[R-sig-ME] Species as both fixed and random effect

Paul Debes paul.debes at utu.fi
Tue May 10 12:42:42 CEST 2016

Hi Liliana,

If your interest is in differences in among-y variance among the seven  
species, and you have single measurements for each combination of the  
seven species by 10 samples (five per species and treatment?), then you  
can do this at the residual (not the random) level by specifying different  
variances for species. However, it would probably be better to have more  
than 10 samples for each species. You can use the 'gls' function of the  
nlme package to do this, or if you include random effects in addition,  

I'm not sure about your experimental design, but if adequate and your  
interest is in firstly Treatment effects, secondly Species effects, and  
thirdly if the Treatment effects differ among Species, then you should do  
this based on the Treatment by Species interaction. Making inferences  
about differences in treatment effects among different groups in the  
absence of formally tested interactions is an horribly often-made mistake  
in many scientific fields:


 From the information you provide, it does not make sense to me to specify  
Species also as a random term. However, why are data from DIFFERENT  
species considered to be not independent? I don't think including Species  
as a random term would account for any non-independence among species.

I hope this helps,

On Tue, 10 May 2016 13:09:28 +0300, Liliana D'Alba Altamirano  
<liliana.dalba at ugent.be> wrote:

> I have a question about the use of a factor as both fixed and random  
> effect.  Specifically, I want to test the effects of an experimental  
> treatment and the species from which the samples originate on the  
> response variable "y".
> I have seven different species with about 10 samples from each.  On one  
> hand I want to be able to make inferences about the differences in "y"  
> between different species, but I also want to look at the variance among  
> the values of "y" at different species (random). In addition, two  
> separate reviewers have explicitly asked for species to be included as  
> random effect to account for the non-independence among data from  
> different species.
> So, my model looks like this:
> y ~ treatment + species + (1|species)
> The question is whether this is an example of statistical malpractice or  
> it is correct to do it.
> I am aware of the discussion about this topic posted here:
> https://stat.ethz.ch/pipermail/r-sig-mixed-models/2014q3/022365.html
> in which Thierry Onkelinx mentioned that it is ok to specify a factor as  
> fixed and random effect but only when it is treated as continuous  
> (fixed) variable.
> However, that did not fully answer my question and unlike Year (which  
> can be used as continuous variable), it does not make sense to treat  
> Species as continuous just so the model produce good estimates.
> Thank you, I appreciate your help.
> Liliana
> 	[[alternative HTML version deleted]]
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models

Paul V. Debes
DFG Research Fellow

Division of Genetics and Physiology
Department of Biology
University of Turku
PharmaCity, 7th floor
Itainen Pitkakatu 4
20014 Finland

Email: paul.debes at utu.fi

More information about the R-sig-mixed-models mailing list