[R-sig-ME] glmer: how are “non-integer #successes in a binomial glm!” actually modeled?

[Ben Bolker]

   In this case I think you should be fine, if it's really reasonable
to imagine that a trial is "partly successful".  lme4 is following
standard GLM practice in handling binomial models -- the variance of
the response is assumed to be p*(1-p)/W .  The warning is there
because *most* people who input non-integer values are making a
mistake (i.e., treating proportional data without a denominator as
binomial).

 We have seen a few cases where non-integer responses give weird
answers in lme4 (e.g. https://github.com/lme4/lme4/issues/180 ), but
as long as your answers seem reasonable I think you should feel
confident.


On Mon, Nov 2, 2015 at 12:39 PM, Bob Wiley <rwwiley at gmail.com> wrote:
> This is hopefully a clear question but I fear the answer may not be
> simple... I have not been able to find anyone who can answer this for me. I
> am using the weights= argument in glmer (family = binomial) and get the
> non-integer successes warning. I know what this means and why I get it--
> some of the weights I have produce values like 4.5 out of 5. This is not an
> error, its because on some trials people were awarded partial credit,
> essentially.
>
> The estimates of the models seem good to me. And if I round-down or
> round-up (i.e. model a 4.5 out of 5 as 4/5 or 5/5) the estimates only
> change a little bit. So this makes me confident that whatever lme4 is
> doing, it is reasonable. That being said... what IS it doing in these
> cases? Should I be doing something other than what
> I am to model this data,
> because of the simple fact that people could receive partial credit on some
> trials?
>
> Thanks so much for any help on this confusing issue...
>
>         [[alternative HTML version deleted]]
>
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models