Skip to content
Prev 9597 / 20628 Next

Minimum number of levels for mixed model

nrm2010 <nrm2010 at ...> writes:
I would suggest

 ~ treat*year + [other covariates] + (1|treat:year:rep)

It's hard to imagine why you wouldn't want to model treatment
as a fixed effect in any case.  It might be nice to model
year as a fixed effect, but it's not practical (unless you want
to do something fancy like putting a Bayesian prior on the
variance).

You might consider adding a term (1|treat:rep) as well (this
would have an effective sample size of 12), to model persistent
differences among plots within treatments.

  It might be a good idea to follow the advice of Murtaugh 2007
(Ecology), and aggregate (take the mean of) all the points within
each treat:year:rep combination.  Your model will simplify to

 ~ treat*year + ... + (1|treat:rep)

because the (1|year:treat:rep) term will be equivalent to
the residual error term.  All you will lose will be the estimate
of the variance within (1|treat:rep:year) combinations.