[R-sig-ME] Estimation of variance components in random- and mixed-effects models
Thierry Onkelinx
th|erry@onke||nx @end|ng |rom |nbo@be
Mon Jun 28 21:30:11 CEST 2021
Dear Amy,
Since you have only three cohorts and a single temperature for each cohort,
you can't include both in the model. I recommend using cohort rather than
temperature. You can plot the cohort effect against their temperature to
get an indication of potential correlation between cohort effect and
temperature.
Since you add cohort as a fixed effect, you take the common effect of all
observations of the same cohort into account. So no problem with
pseudoreplication. That would be a problem when you include only
temperature as a numeric instead of cohort as a factor.
Best regards,
ir. Thierry Onkelinx
Statisticus / Statistician
Vlaamse Overheid / Government of Flanders
INSTITUUT VOOR NATUUR- EN BOSONDERZOEK / RESEARCH INSTITUTE FOR NATURE AND
FOREST
Team Biometrie & Kwaliteitszorg / Team Biometrics & Quality Assurance
thierry.onkelinx using inbo.be
Havenlaan 88 bus 73, 1000 Brussel
www.inbo.be
///////////////////////////////////////////////////////////////////////////////////////////
To call in the statistician after the experiment is done may be no more
than asking him to perform a post-mortem examination: he may be able to say
what the experiment died of. ~ Sir Ronald Aylmer Fisher
The plural of anecdote is not data. ~ Roger Brinner
The combination of some data and an aching desire for an answer does not
ensure that a reasonable answer can be extracted from a given body of data.
~ John Tukey
///////////////////////////////////////////////////////////////////////////////////////////
<https://www.inbo.be>
Op ma 28 jun. 2021 om 21:21 schreef Amy Huang <amyrbss using gmail.com>:
> Thank you very much for your responses and references. Sorry that I missed
> mentioning a lot of information.
>
> I am using lme4, and the fixed predictors are all numeric. Only having 3
> levels of cohorts is indeed the major issue. After removing insignificant
> predictors in the 2nd model, the only factor left is temperature: offspring
> trait ~ temperature + (1 | cohort/female/clutch), which gives the
> convergence warning.
>
> Now I treat cohort as a fixed effect, but when I include both cohort and
> temperature as fixed effects (in the 2nd model), a warning appears:
> "fixed-effect model matrix is rank deficient so dropping 1 column /
> coefficient". When I remove cohort (2nd model), the two models become very
> similar and give similar results.
> offspring trait ~ cohort + (1 | female/clutch)
> offspring trait ~ temperature + (1 | female/clutch)
> But am I not introducing pseudoreplicates if I do not include cohort as a
> factor?
>
> PS. The section "How do I compute a coefficient of determination (R2), or
> an analogue, for (G)LMMs?" in the GLMM FAQ also gives me some insight.
> However, the links provided there seem to be not working.
>
> Best regards,
> Amy Huang
>
> Am Mo., 28. Juni 2021 um 19:32 Uhr schrieb Ben Bolker <bbolker using gmail.com>:
>
> > See also:
> >
> >
> >
> https://stats.stackexchange.com/questions/37647/what-is-the-minimum-recommended-number-of-groups-for-a-random-effects-factor
> >
> > https://www.biorxiv.org/content/10.1101/2021.05.03.442487v2
> >
> > (I should these links, and the blog post link, to the GLMM FAQ ...)
> >
> > On 6/28/21 1:17 PM, Thierry Onkelinx wrote:
> > > Another issue is that you have too few levels to fit "cohort" as a
> > > random effect. I wrote a blogpost on this a few years ago:
> > > https://www.muscardinus.be/2018/09/number-random-effect-levels/
> > > <https://www.muscardinus.be/2018/09/number-random-effect-levels/>
> > >
> > > Best regards,
> > >
> > > ir. Thierry Onkelinx
> > > Statisticus / Statistician
> > >
> > > Vlaamse Overheid / Government of Flanders
> > > INSTITUUT VOOR NATUUR- EN BOSONDERZOEK / RESEARCH INSTITUTE FOR NATURE
> > > AND FOREST
> > > Team Biometrie & Kwaliteitszorg / Team Biometrics & Quality Assurance
> > > thierry.onkelinx using inbo.be <mailto:thierry.onkelinx using inbo.be>
> > > Havenlaan 88 bus 73, 1000 Brussel
> > > www.inbo.be <http://www.inbo.be>
> > >
> > >
> >
> ///////////////////////////////////////////////////////////////////////////////////////////
> > > To call in the statistician after the experiment is done may be no more
> > > than asking him to perform a post-mortem examination: he may be able to
> > > say what the experiment died of. ~ Sir Ronald Aylmer Fisher
> > > The plural of anecdote is not data. ~ Roger Brinner
> > > The combination of some data and an aching desire for an answer does
> not
> > > ensure that a reasonable answer can be extracted from a given body of
> > > data. ~ John Tukey
> > >
> >
> ///////////////////////////////////////////////////////////////////////////////////////////
> > >
> > > <https://www.inbo.be>
> > >
> > >
> > > Op ma 28 jun. 2021 om 16:31 schreef Ben Bolker <bbolker using gmail.com
> > > <mailto:bbolker using gmail.com>>:
> > >
> > > Are you using lme4? (I'm 99% sure you are, but it's good to be
> > > explicit.)
> > >
> > > Are all of your fixed predictors numeric (rather than
> > > factor/categorical) ?
> > >
> > > Note that a convergence warning is a *warning*, not an error:
> > have
> > > you checked the troubleshooting steps in ?lme4::convergence (in
> > > particular, scaling and centering your predictor variables might
> > > help ...)
> > >
> > > cheers
> > > Ben Bolker
> > >
> > >
> > > On 6/28/21 10:17 AM, Amy Huang wrote:
> > > > Dear all,
> > > >
> > > > I am examining maternal effects, and my data have three
> hierarchy
> > > levels:
> > > > clutches of the same female, females, and cohorts. My
> explanatory
> > > variables
> > > > are at the female level (female length, age) and at the cohort
> > level
> > > > (temperature).
> > > >
> > > > I would like to estimate the variance components of each
> > > hierarchy level
> > > > (i.e. relative amount of variance at each level) and then to
> find
> > > out which
> > > > factors (female length, age, temperature) explain most of the
> > > variance. For
> > > > these, I have two models:
> > > > offspring trait ~ 1 + (1 | cohort/female/clutch)
> > > > offspring trait ~ temperature + female length + age + (1 |
> > > > cohort/female/clutch)
> > > >
> > > > The major problem is that I only have 3 cohorts (and so 3
> > > temperatures).
> > > > From the first model I am able to get the information, but from
> > > the second
> > > > one there is an error message: "Model failed to converge with 1
> > > negative
> > > > eigenvalue: -2.0e+01". The error pops up probably because I have
> > both
> > > > temperature (fixed) and cohort (random) included. Is my approach
> > > correct?
> > > > And is there a way to fix this error?
> > > >
> > > > Thank you so much for your time.
> > > >
> > > > Best regards,
> > > > Amy Huang
> > > >
> > > > [[alternative HTML version deleted]]
> > > >
> > > > _______________________________________________
> > > > R-sig-mixed-models using r-project.org
> > > <mailto:R-sig-mixed-models using r-project.org> mailing list
> > > > https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
> > > <https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models>
> > > >
> > >
> > > _______________________________________________
> > > R-sig-mixed-models using r-project.org
> > > <mailto:R-sig-mixed-models using r-project.org> mailing list
> > > https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
> > > <https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models>
> > >
> >
> > _______________________________________________
> > R-sig-mixed-models using r-project.org mailing list
> > https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
> >
>
> [[alternative HTML version deleted]]
>
> _______________________________________________
> R-sig-mixed-models using r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>
[[alternative HTML version deleted]]
More information about the R-sig-mixed-models
mailing list