[R-sig-ME] (no subject)
john.maindonald at anu.edu.au
Thu Dec 11 11:21:52 CET 2014
On point 2, it is the random effects that are assumed to be normally
distributed. You can use ranef() to get estimates of these. Unless
however the design is pretty much balanced, the estimates can be
highly non-normal, because of the interplay between effects at
different levels of the design, or in a crossed design, different random
factors. Also, think which random effects matter for purposes of the
model estimates in which you are interested.
In a classical design where one set of treatments are applied at the
level of plots, with perhaps another at the level of subplots, it is the plot
random effects that likely mainly matter for inference wrt treatments
applied to plots (and in that classical context, you probably have a
balanced design, i.e., all treatment differences estimated with similar
In severely unbalanced designs, maybe one can get somewhere by
simulating from the fitted model, plotting ordered simulated plot
effects agains the estimated effects, and hoping for a roughly linear
scatter. Others may be able to comment — what literature is there
addressing this general issue?
John Maindonald email: john.maindonald at anu.edu.au
phone : +61 2 (6125)3473 fax : +61 2(6125)5549
Centre for Mathematics & Its Applications, Room 1194,
John Dedman Mathematical Sciences Building (Building 27)
Australian National University, Canberra ACT 0200.
On 11 Dec 2014, at 20:00, ONKELINX, Thierry <Thierry.ONKELINX at inbo.be> wrote:
> Dear Emily,
> It sounds like you could use some reading on this topic. I would recommend Zuur et al (2009), Bolker (2008) and Pinheiro & Bates (2000).
> 1. No. The assumption of normality is only for the **residuals** of a **linear** (mixed) model. So neither the fixed effects nor the response has to be normally distributed. Note that the residuals of a **generalized** linear (mixed) model don't have the assumption of normality.
> 2. Yes, these are assumed to be normally distributed. Small deviations from normality are not problematic. Strong deviations usually indicate that you are missing an important covariate. Maybe the assumption of i.i.d. is not valid due to the spatial structure of the plots (spatial autocorrelation).
> 3. Plot the residuals against available covariates. There shouldn't be a pattern in the residuals.
> 4. The deviance of a single model is not that informative. It is mainly used to compare models.
> Note that DBH+Elevation+(DBH*Elevation) can be written as DBH * Elevation. Andaddingspacestocodemakesitmuchmorereadable.
> Best regards,
> ir. Thierry Onkelinx
> Instituut voor natuur- en bosonderzoek / Research Institute for Nature and Forest
> team Biometrie & Kwaliteitszorg / team Biometrics & Quality Assurance
> Kliniekstraat 25
> 1070 Anderlecht
> + 32 2 525 02 51
> + 32 54 43 61 85
> Thierry.Onkelinx at 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
> -----Oorspronkelijk bericht-----
> Van: R-sig-mixed-models [mailto:r-sig-mixed-models-bounces at r-project.org] Namens Emily L. De Stigter
> Verzonden: donderdag 11 december 2014 8:38
> Aan: r-sig-mixed-models at r-project.org
> Onderwerp: [R-sig-ME] (no subject)
> I'm just out of undergrad and working in an ecology lab as the leading statistical investigator for a project studying conifers in Northern California. I'm using lme4 to do a GLMM. My response variable is binary:
> status of the trees (live vs dead). I also have two fixed effects: diameter at breast height (DBH) and elevation, plus the interaction of the two. My random effect is plot.
> Here are my questions:
> Do the fixed effects (DBH and elevation) need to be normally
> distributed? Both DBH and elevation are not normally distributed and the
> basic transformations I've tried did not correct the issue.
> The random effect I have (plot) is not normally distributed as I know
> that it needs to be. I have tried a couple different transformations (log,
> sqrt...) but, again, nothing the corrected the issue. What should I try
> next to fix it?
> How should I best check that my model is fitting the data appropriately?
> How do I interpret the deviance of the model?
> Here's some of my R code:
> PSME stands for Psuedotsuga menziesii, one of the species of interest.
> Thanks for reading and let me know if anyone has any follow-up questions. I appreciate greatly any and all advice I get. I realize these questions are not exactly related to this forum, so I would also be glad for some suggestions on resources to check out.
> Thanks again,
> [[alternative HTML version deleted]]
> R-sig-mixed-models at r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
> Disclaimer Bezoek onze website / Visit our website<https://drupal.inbo.be/nl/disclaimer-mailberichten-van-het-inbo>
> R-sig-mixed-models at r-project.org mailing list
More information about the R-sig-mixed-models