[R-sig-ME] need help with predicted values from lmer with newdata

[Ben Bolker]

Joshua Wiley <jwiley.psych at ...> writes:

> 
> Hi Paul,
> 
> I think this may be a bug in predict.merMod.  It is not terribly
> elegant, but I suggest changing line 102 from:
> 
> re_new <- unlist(re_List[m])
> 
> to
> 
> re_new <- as.vector(do.call("rbind", re_List[m]))
> 
> I believe the issue with the current is that the random effects are
> unlisted which essentially 'stacks' the vector, so you have the random
> effects one at a time by effect, not by level (i.e., all intercepts,
> then all slopes).  Later this is post multiplied by a matrix that is
> sorted by level, the dimensions are correct so the matrices multiply
> fine, but the results are not what I believe was intended.
> 
> I cannot see a way to add a patch to R-forge or submit a bug report so
> I cced Doug on this.

  Thanks Josh. This should be fixed now.

 It turned out had to re-vamp things a bit to fix this
in a general and robust way.  I'm not quite sure how this slipped through
the cracks, since I *thought* that I had tested predict() with a multi-response
case (i.e. the sleepstudy data) and gotten sensible answers.  However,
I was testing with graphical output and not with explicit stopifnot() tests,
so it's possible that it was working originally and that I subsequently
broke something ...

  Paul, would you be willing to generate a trimmed-down version of your
example that simply generates multi-trait data?

  For future reference, the lme4 bug tracker lives at:

https://r-forge.r-project.org/tracker/?atid=298&group_id=60&func=browse

  In addition, any advice on how to improve ?predict.merMod ("I can't
tell what those numbers are. Can't be sure from ?predict.merMod") would
be welcome ...

  Ben Bolker