Skip to content
Prev 6864 / 7420 Next

[EXTERNAL] Re: Using GLMs or GLMMs for diversity metrics?

Hello Tom,
Yes..I think it is a very useful statistical method, the gllvm package 
is great, and the people who have programmed it are very helpful (which 
is also important).
Well...opinions may (and will) differ....but the way I see random 
effects is something that is a random grab out of a large number of 
possible values. Say you have 10000s of birds out there in the wild, and 
somehow you decided to catch and tag a random selection of them....say 
100. You measure multiple observations from the same bird, and then you 
use a random intercept 'Bird' to model a different mean value per bird, 
and en-passant the random intercept models dependency. And you then want 
to generalise that back to all those birds out there. Three for the 
price of one. But you do assume that the random effects are iid. As in 
'independent and identical' distributed. With emphasis on independent.? 
Whether that assumptions holds, I don't know.

Now to your 'year' as random effect....if you collect 5, 10 or 20 years 
of sequential data...then that is not a random grab out of a 
large?number of possible years. They are (mostly) sequential.? And if 
you plot the estimated random effects versus time, then I am pretty sure 
that you are going to see a nice temporal trend in those estimated 
random effects, which is not iid. What is the consequence of this? I 
don't know. That will be very data and model specific. Sometimes, if you 
do the right thing and you compare it with a GLMM in which year is a 
random effect, then you see minimal differences in the fixed effects, 
but sometimes you do see differences. And finally..if you use Year with 
only (say) 5 levels as a random effect....to what exactly do you 
generalise it to?

Also...if you are going to do simulations/predictions from the model 
(e.g. as in DHARMa model validation).....assuming idd for a random 
effect that is in fact AR1 (or something else) may give you some trouble.
I guess there is where the R-INLA fun starts provided that you have 
enough spatial locations. And if you don't...then yeah...maybe use year 
as a random effect after all?


But as I said...I'm pretty sure that there are plenty of people with a 
different opinion.


This may be a relevant link (within the context of GLLVM) as well:

https://openresearch-repository.anu.edu.au/server/api/core/bitstreams/a7eecd7d-e78a-410d-ab4e-7c12222a6b4a/content
Thanks......urgently begging for second editions.

Kind regards,

Alain
Message-ID: <087ac8b0-bbec-4670-beac-cb0be9a2eaa8@highstat.com>
In-Reply-To: <SA1PR09MB8128D0A5708CB75D5D2B9134F3452@SA1PR09MB8128.namprd09.prod.outlook.com>