[R-sig-ME] Not really interested in correlations or random effects variances

[Phillip Alday]

You've discovered the compound symmetric model.

I owe this list some comments on this, but if you search the archives
for some discussion between Maarten Jung and me (one thread was titled
"Appropriate model reduction sequence for factorial design in glmmTMB"
from 2020Q2), you'll find some thoughts on this.

Best,
phillip

On 2/10/20 10:52 am, Michele Scandola wrote:
> Hi all,
> 
> let's hypothesize that I have a full-crossed experiment, that actually it
> may be analysed with a standard ANOVA, but because I want to increase the
> power of the analyses, and add as random slope the time passed
> trail-by-trail since the beginning of the experiment, to take into account
> for fatigue effects, I want to analyse these data with a multilevel linear
> model.
> 
> Let's say that in this experiment I have two within-subjects factors that
> are interacting. If I am not really interested neither in the variances
> between the random slopes nor in the variances of the random effects, may I
> totally avoid random slopes, using in their place random intercepts?
> 
> What I mean is, instead of this full slopes model: y~C1*C2+(C2*C2|ID)
> can I use this "full intercept model":
> y~C1*C2+(1|ID:C1:C2)+(1|ID:C1)+(1|ID:C2)+(1|ID)?
> 
> Is there any specific reason why it should not be recommended the second
> model?
> 
> Thanks a lot for your help, best regards,
> 
> Michele Scandola
>