[R-sig-ME] Why || doesn't zero out the correlations in lmer
Phillip Alday
me @end|ng |rom ph||||p@|d@y@com
Sat Oct 9 05:46:48 CEST 2021
On 10/8/21 17:07, Simon Harmel wrote:
> Thank you, this is extremely helpful to know. You mentioned it's well
> documented, any possible links to share?
The lme4 documentation, e.g.
https://www.rdocumentation.org/packages/lme4/versions/1.1-27.1/topics/lmer
> (Because of the way it is implemented, the ||||-syntax /works only for
> design matrices containing numeric (continuous) predictors/; to fit
> models with independent categorical effects, see |dummy
> <https://www.rdocumentation.org/link/dummy?package=lme4&version=1.1-27.1>|
> or the |lmer_alt| function from the afex package.)
> Also, does nlme::lme() behave in the same manner in this regard?
I am unaware of nlme supporting the double-bar syntax at all, but
specifing the correlation structure to be diagonal (pdDiagonal? it's
been a while) will force all correlations to zero.
>
> On Fri, Oct 8, 2021 at 4:46 PM Phillip Alday <me using phillipalday.com> wrote:
>> This is a well-documented issue: || doesn't zero correlations between a
>> categorical variable's levels. As far as I know, there are
>> software-development/technical reasons for this, not statistical ones.
>>
>> The afex package has an implementation that zeroes everything out.
>>
>> On 8/10/21 4:32 pm, Simon Harmel wrote:
>>> Dear Colleagues,
>>>
>>> I have a 'factor' predictor called 'type' (with 4 levels). In the
>>> random part, I have used `||` so the levels of 'type' can't correlate
>>> with each other.
>>>
>>> But I wonder why still correlations are reported in the output?
>>> Thanks, Simon
>>>
>>> lmer(y~type + (type || ID), data = data)
>>>
>>> Random effects:
>>> Groups Name Std.Dev. Corr
>>> ID type0 0.4276
>>> type1 0.7012 0.81
>>> type2 0.7115 0.72 0.97
>>> type3 0.7655 0.83 1.00 0.98
>>>
>>> _______________________________________________
>>> R-sig-mixed-models using r-project.org mailing list
>>> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>>>
More information about the R-sig-mixed-models
mailing list