Skip to content

Help with Linear Model

1 message · Phillip Alday

#
Whether or not you're interested in a parameter has little to no
influence on whether or not it should be in your model nor whether or
not it is a fixed or random effects. Instead, that is determined by the
structure of your data. In your case, it would in many ways be nicer to
have years as a random effect (as you only care about the *variance*
due to year and not any particular fixed year), but you simply don't
have the data structure to do so with so few levels.

In your case, the year parameter is simply a nuisance parameter -- it
captures important variation, even if it's variation you don't care
about. Now you can marginalize over that effect either before modelling
(e.g. by aggregating / averaging over the time dimension) or after
modelling (your choice of test type (II, III) or using predicted
marginal / least-square means (e.g. via package lsmeans) or implicitly
(just completely omitting it from the model specification). I would
tend to leave the year effect in the model if possible -- I'm a big fan
of modelling as much as you can so that your model is as good as
possible, even if there are only a few predictors you actually care
about.?

That said, if you don't have enough data for such a complex model, then
you should leave out less-interesting dimensions in your model
specification (and potentially aggregate across them beforehand). The
Bates et al preprint on parsimonious mixed models really emphasizes
this point -- it doesn't do much good to model everything and have a
degenerate and/or overfitted model. You know your data best and so you
know what tradeoffs are best for your case.

Best,
Phillip
On Fri, 2016-11-04 at 09:51 -0600, Joseph Aidoo wrote: