[R-sig-ME] Understanding and analysing split split design using lmer()

Guillaume Adeux gu|||@ume@|mon@@2 @end|ng |rom gm@||@com
Tue Sep 10 10:53:24 CEST 2019

OK, I will respond in a practical sense with my philosophy (I am not a
statistician), and will let others fill in on a more statistical note.

First of all, there is no real gain in considering block as random, it will
only "cost" you 3 df as fixed, which is not dramatical. Indeed, random
effects are modelled as variance and four points don't seem very reliable
to estimate a variance.

What do you mean by oveparameterized? If it is a singular fit, this just
implies that certain random effects are estimated as 0. I usually tend to
keep everything because I want my model to reflect my experimental set-up.
Note that removing the random effects which are estimated at 0 will not
change any of the fixed effect coefficients or SE. If the model does not
converge, I consider that as more problematic and selection of random
effects might be necessary (drop the one with the lowest variance).
If you overlooked my statement about (1|block:crop:tillage:CC) (i.e. you
introduced it in the model even though you don't have pseudoreplication),
then this might explain it also.

Don't farm naked, plant cover crops,


Le mar. 10 sept. 2019 à 10:40, Rabin KC <kc000001 using umn.edu> a écrit :

> Thank you for your suggestion, Guillaume.
> However, the model you suggest is overparameterized in my case. I don't
> have enough observations to fit the maximal model.
> Also regarding blocks as fixed, I would like to make block inferences in a
> broad sense. Is there a particular reason that 4 levels are not enough to
> be modeled as a random effect? Or is it due to the risk of getting
> imprecise treatment estimates as mentioned in this topic below?
> https://stat.ethz.ch/pipermail/r-sig-mixed-models/2010q2/003712.html
> I look forward to hearing from you
> Rabin
> On Tue, Sep 10, 2019 at 1:51 AM Guillaume Adeux <
> guillaumesimon.a2 using gmail.com> wrote:
>> I agree with Peter Claussen.I am a weed ecologist and often deal with
>> these experimental set-ups.
>> The correct full intercept model ought to be:
>> lmer(biomass~block+
>> crop*tillage*cover+(1|block:crop)+(1|block:crop:tillage)+(1|block:crop:tillage:CC))
>> Note that block is treated as fixed because 4 levels are not sufficient
>> to estimate a random effect.
>> Also note that (1|block:crop:tillage:CC) should only be introduced in the
>> model if there is pseudoreplication at the elementary plot level, as is
>> often done.
>> For appropriate test of effects, look into the {monet} or {afex} package.
>> For appropriate post-hoc multiple comparisons, look into the {emmeans}
>> package.
>> Cheers,
>> GA2
>> Le lun. 9 sept. 2019 à 20:36, Rabin KC <kc000001 using umn.edu> a écrit :
>>> Hello R mixed models community,
>>> I am trying to fit a model for a split- split-plot design using lmer. I
>>> know how to do this with aov and ssp.plot function in agricolae, but I
>>> need
>>> to use lmer in this case.
>>> I am interested in the response variable biomass of cover crops. I have
>>> the
>>> main plots as crops (corn and soybean). The subplot is tillage, which is
>>> randomized within the crops and has 3 levels (conventional-till, no-till,
>>> and strip-till). Within tillage(sub plot), 3 cover crop strategies are
>>> randomly assigned (AR, ARCC, ARCCFR). Therefore 1 rep (block) has 18
>>> treatments. The whole experiment is replicated 4 times, therefore, total
>>> experimental units equal to 72 units.
>>> Data arranged in a csv file is attached in this email.
>>> Now the analysis part:
>>> I know how to do this using aov/ ssp.plot( agricolae). The code is as
>>> follows:
>>> model<-with(data,ssp.plot(rep,crop,tillage,cover,biomass)).
>>> Same thing with aov:
>>> model1<aov(cover_coverage~rep+crop*tillage*cover+Error(rep/crop/tillage),data=data)
>>> For my  analysis using lmer, and lme,  I have the following code:
>>> model2<-lmer(biomass~crop*tillage*cover+ (1|rep),data=data)
>>> model3<-lme(fixed=biomass~crop*tillage*cover, data=data, random=~1|rep)
>>> Now my confusion and questions are:
>>> 1. I have used crop, tillage, and cover as fixed effects. And only the
>>> rep
>>> (block) as random. Is the random effect properly assigned for this
>>> design?
>>> 2. I have read a lot about crossed and nested design. Crawley and Oehlert
>>> give particular examples about it, but I have trouble understanding if
>>> this
>>> design has nested factors or crossed factors.
>>> I would be very relieved to receive feedback with the matter in hand.
>>> Looking forward, and thank  you  a lot,
>>> Sincerely,
>>> Rabin
>>> _______________________________________________
>>> 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