[R-sig-ME] fixed effects estimates too small in binomial GLMM with low "success"-rate

Henrik Singmann henrik.singmann at psychologie.uni-freiburg.de
Sun May 18 00:38:20 CEST 2014

Am 18.05.2014 00:19, schrieb Ben Bolker:
> On 14-05-17 06:07 PM, Henrik Singmann wrote:
>> Dear list,
>> I would really appreciate your input on an issue in which the fixed
>> effects estimates of a binomial GLMM with relatively low rate of
>> "successes" are too low (compared to the observed effect) by around .5%.
>    Don't know, but one comment and one thought:
>    (1) the convergence warning goes away with the most recent version
> of lme4 (1.1-7)

This is true, thanks.
>    (2) I'm _guessing_ that you will find that this phenomenon is
> due to Jensen's effect (i.e., a nonlinear averaging effect) --
> that would mean its magnitude should be proportional to the
> random-effects variances and to the strength of the nonlinear
> (inverse link) relationship.
>    (I think there was a similar, possibly neglected, question along
> these lines on the mailing list earlier ...)

Thanks for the pointer. In fact, when entering experiment as fixed effect (which removes the random effect with the smallest variance), the estimates are way more precise. They are almost perfect. Perhaps I should then use this approach (although I would prefer treating experiment as random). Or what would you do?

>    Ben Bolker

Dr. Henrik Singmann
Albert-Ludwigs-Universität Freiburg, Germany

More information about the R-sig-mixed-models mailing list