[R-sig-ME] comparing random effect variation between data sets

[Ben Bolker]

  A couple of quick thoughts:

* as a crude test, you could get the profile confidence intervals
for the random-effect SDs in each model and compare them
* a more formal test would put both species into the same model
and allow for different variances.  This might work:

library(lme4)
library(gamm4)
alldat <- rbind(data.frame(species="X",speciesX),
                data.frame(species="Y",speciesY))

comb <- gamm4(distance ~ s(time,k=4,by=species),
                  random=~(1|CODE)+(dummy(species,"Y")|CODE))

You could test of comb against the model with just random=~(1|CODE),
or look at the confidence intervals of the second RE term.


On Thu, Jun 25, 2015 at 9:07 AM, John Morrongiello
<jrmorrongiello at gmail.com> wrote:
> Hi all
>
> I have movement data for two species of fish (say species X and species Y).
> The response variable is 'distance from home' (distance) and multiple
> measurements (5-17) are made for 20-30 individuals per species over a
> period of 200 days. Distance is continuous and strictly positive so I am
> using a gamma distribution with a log link. My collegues would like me to
> fit seperate models for each species to explore temporal patterns in
> movement (time). We believe there will be non-linear patterns in movement
> through time so have I have fitted GAMMs
>
> I'm happy with the two models and their interpretation. I would, however,
> like to compare the level of among-individual variation in movement between
> species. The species-average trends are pretty similar but it looks like
> there is more variance among individuals for species X than species Y.
>
> I was thinking of extracting the individual level random effects from the
> two models and performing a variance test on these. My logic is that given
> the response variables are the same and the model structures are the same,
> I could make a meaningful comparison of random effects. The overall model
> intercepts are different but I thought this is not an issue here as I'm
> only interested in the variance of the random effects, not their average.
>
> Is this approach valid, or am I violating assumptions/ these data are not
> directly comparable ? If not, what would be a potential way to perform this
> comparision? Below is the code I'm thinking of using.
>
> Thanks very much for your time
>
> John
>
> #####GAMMs
> X1<-gamm4(distance ~ s(time,k=4), random=~(1|CODE),data=speciesX,
> family=Gamma(link=log))
> Y1<-gamm4(distance ~ s(time,k=4), random=~(1|CODE),data=speciesY,
> family=Gamma(link=log))
>
> ######extract random effects from the two models
> X1ranef<-ranef(X1$mer)###extract random effects from model
> X1ranef<-data.frame(matrix(unlist(X1ranef$'CODE'),ncol=1))##convert list of
> random effects to data frame
> names(X1ranef)<-c('CODE')
>
> Y1ranef<-ranef(X1$mer)###extract random effects from model
> Y1ranef<-data.frame(matrix(unlist(Y1ranef$'CODE'),ncol=1))##convert list of
> random effects to data frame
> names(Y1ranef)<-c('CODE')
>
> ######perform variance test
> var.test(X1ranef$CODE,Y1ranef$CODE)
>
>         [[alternative HTML version deleted]]
>
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models