[R-sig-ME] A question on item random effects

[Ben Bolker]

On 14-07-09 06:03 AM, Anastasiia Romanova wrote:
> Dear all,I would be very grateful if somebody helped me answering the
> following question.. I'm looking at some categorical (naming
> accuracy) data with glmer, and I would like to use items as a random
> factor. I have 4 participants that name the items. The problem is
> that items vary across participants quite considerably, which leads
> to one third of the items being named only by 1 participant. I have
> heard that you need at least 2 subjects to  give a response on one
> item to be able to include this item into the random effect analysis.
> Is that right? And if it is, can I avoid this problem somehow? Thank
> you very muchBest wishes,Anastasiya 

  A reproducible example might be nice, but I'll take a shot.

  In theory it should be possible to fit a random effect of item as long
as _any_ items are measured repeated times (they don't _all_ have to be
measured more than once).  In practice it may be difficult if only a
small fraction of items have been measured repeatedly.  It's hard to say
with precision what "small" means in the previous sentence, or how this
will come out in your case.  If you are in fact pushing your data too
hard, the symptoms will be convergence warnings and singular fits
(getME(.,"theta") including zero values, or VarCorr() including
variances of zero or correlations of +/- 1).  The singular fit is not
_necessarily_ wrong, but increases the chances of numerical errors and
complicates interpretation. One choice would be to use blme::bglmer() to
force your data away from the singular edge (but you would have to read
the corresponding paper to understand what you were actually doing).

  cheers
    Ben Bolker