[R-sig-ME] random effect variance greater than output variable variance

[Norman DAURELLE]

Dear Ben, 

is there something important about the variance being attached to the intercept ? 

If a factor influences the output variable, the deviation it accounts for is the same when it is applied to the intercept or not (meaning wether X, R1 and R2 = 0) isn't it ? 

Concerning the number of years, I mainly set them as random effects because I thought you used random effects when you know that you do not observe all levels. 

additionnally, Paul Johnson wrote this in a recent previous answer : 

[ The minimum number of blocks/groups required to support a random effect is discussed in Ben Bolker's GLMM FAQ wiki: 

https://bbolker.github.io/mixedmodels-misc/glmmFAQ.html#should-i-treat-factor-xxx-as-fixed-or-random 

"One point of particular relevance to ‘modern’ mixed model estimation (rather than ‘classical’ method-of-moments estimation) is that, for practical purposes, there must be a reasonable number of random-effects levels (e.g. blocks) – more than 5 or 6 at a minimum." ] 

So, if you are of the opinion that practicality prevails over philosophy about when a variable should be used as a random or fixed effect, apparently 5 is ok. 

It may be worth noting that I don't wish to predict the outcome, simply to describe the effct of my fixed effects on the output. 

And I think that using year as a fixed or a random effect changes the values for the fixed effects (disease and rainfall), so I would actually rather have a reasoning that makes theoretical good sense than something that is pragmatic for use behind the treatment of the year factor. 

Regards, 

Norman 




De: "ben pelzer" <benpelzer using gmail.com> 
À: "r-sig-mixed-models" <r-sig-mixed-models using r-project.org> 
Envoyé: Mercredi 9 Novembre 2022 10:29:11 
Objet: Re: [R-sig-ME] random effect variance greater than output variable variance 

Dear Norman, 

The random intercepts in your model are related to (as always) value zero 
of your predictors, i.e. X=0, R1=0 and R2=0. These 0-values may be far out 
of the actual or possible range of values. This means that the intercept 
variance is about variance between locations, say, with zero disease and 
zero rainfall in both seasons. If you do not want this, rescale X, R1 and 
R2 so that value zero is "in range". 

Further, and that is also a reason that I respond, I was wondering if it is 
a good idea to estimate random effects if there are so few units, like only 
five "fixed" years e.g. A.f.a.i.k. one should have at least 20 units or so, 
but maybe you or someone else could correct me. 

Regards, Ben. 

[[alternative HTML version deleted]] 

_______________________________________________ 
R-sig-mixed-models using r-project.org mailing list 
https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models 

	[[alternative HTML version deleted]]