[R-sig-ME] Choice of fixed effects and random effects in glmm

[Ebhodaghe Faith]

Many thanks, Philip and apologies for delayed response. Your text gives
some really great insights but I'm still studying it and trying to clearly
grasp some aspects of the message.

Cheers
Faith

On Wed, 9 Dec 2020, 1:24 p.m. Phillip Alday, <me using phillipalday.com> wrote:

> Hi  Faith,
>
> if you have a model like
>
> y ~ 1 + x * location
>
> (i.e. fixed-effects only), then this still accounts for
> repeated-measurements within the location. The motivation for doing a
> mixed model like
>
> y ~ 1 + x + (1+x|location)
>
> is that this reduces the complexity of the fixed effects, especially
> when you don't want to interpret the effects of the individual levels of
> location. This reduction in complexity means that you have fewer things
> in your fixed-effects table, which is nice, but it also means that there
> are fewer parameters in the model (because you model the variance across
> locations instead of the mean at each location) and so it becomes easier
> to fit such models when you have lots of locations. (There are also some
> other more subtle differences in terms of partial pooling, but we can
> leave those aside for now).
>
> But if you want to interpret the effect of location or particular
> locations, e.g. "the effect of x at location A", then that extra
> complexity in the fixed-effects table isn't really a problem. In that
> case, I would recommend just treating location as  a fixed effect.
>
> My commentary thus far is simplifying a lot of detail -- there are
> exceptions to almost all of the rules I'm stating. I don't know enough
> about your data, research question and inference goals to be able to
> tell if you are one of the exceptions.
>
> Thierry Onkelinx has a nice blog post on when it's okay to have
> something as both a fixed and a random effect:
>
> https://www.muscardinus.be/2017/08/fixed-and-random/
>
> The short answer is "only when you have discrete data, not for
> categorical nor continuous data", where discrete data are things like
> "time samples" which have a numerical structure but are not truly
> continuous. Distinct locations are usually simply categorical because
> they don't have a numerical structure (though I guess locations
> expressed as e.g. latitude or longitude might).
>
> Best,
>
> Phillip
>
> On 09/12/2020 07:06, Ebhodaghe Faith wrote:
> > Hi All.
> >
> > I'm running a Generalised Linear Mixed Model for a study in which I
> > repeatedly collected data within each of 14 different locations. My
> > predictor variable is 'location' and at same time wish to account for
> > repeated measures within locations. Is it okay in this case to select
> > 'location' as both fixed effects and random effects?
> >
> > Thanks in advance for kind response.
> >
> > Cheers
> > Faith
> >
> >       [[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]]