[R] problem with anova() and syntax in lmer

[David C. Howell]

The answer to your first question is "yes, the order does make a 
difference." I have not worked with lmer, but the standard anova applied 
to lm() will provide what are called Type I sums of squares. Each effect 
is adjusted for all prior effects.

Look at John Fox's car package. I don't know if it will handle lmer 
models, but it is worth trying. Note that for car the function is Anova, 
not anova.

Good luck,
Dave Howell


Gilles San Martin wrote:
> Dear R user
> 
> I have 2 problems with lmer.
> The statistical consultance service of my university has recomended to me to 
> expose those problems here.
> 
> Sorry for this quite long message.
> Your help will be greatly appreciated...
> 
> Gilles San Martin
> 
> 
> 1) anova()
> 
> I fit a first model :
> model1 <- lmer(eclw~1 + density + landsc + temp + landsc:temp + (1|region) + 
> (1|region:pop) + (1|region:pop:family), data=fem1)
> 
> I fit the same model but I'm just changing the order of 2 fixed factors 
> (here : "temp" and "landsc") :
> model2 <- lmer(eclw~1 + density + temp + landsc + landsc:temp + (1|region) + 
> (1|region:pop) + (1|region:pop:family), data=fem1)
> 
> Then, if I apply the anova() function on these 2 models, the given Sum of 
> Squares are different for the fixed effects whose place has been changed:
> 
>> anova(model1)
> Analysis of Variance Table
>             Df  Sum Sq Mean Sq
> density      1 21941.3 21941.3
> landsc       1  4800.7  4800.7
> temp         1 10119.9 10119.9
> landsc:temp  1   292.2   292.2
> 
>> anova(model2)
> Analysis of Variance Table
>             Df  Sum Sq Mean Sq
> density      1 21941.3 21941.3
> temp         1 10441.1 10441.1
> landsc       1  4479.5  4479.5
> temp:landsc  1   292.2   292.2
> 
> How is it possible? Do the fixed effects need to be writen in a particular 
> order ?
> My dataset is unbalanced. Somebody tells to me that this could have some 
> importance for this problem.
> 
> 
> 
> 2) syntax
> 
> I have a quite complex model and we have not been able to find accurate 
> documentation about the syntax corresponding to my model.
> 
> I have  :
>  - 2 fixed factors : "landsc" & "temp" and their interaction " landsc:temp"
>  - 1 continuous covariate considered as fixed
>  - 3 nested random factors : "region", "pop" and "family" with family nested 
> in pop and pop nested in region*landsc
> 
> I'm mainly interrested in the effect of "landsc" ane "landsc:temp" on the 
> variable I'm studying.
> 
> I had used the following synthax :
> model3 <- lmer(eclw~1 + density + landsc + temp + landsc:temp + (1|region) + 
> (1|region:pop) + (1|region:pop:family), data=fem1)
> 
> But somebody told to me that the folowing one could be more correct , and 
> I'm in doubt now:
> model4 <- lmer(eclw~1 + density + landsc + temp + landsc:temp + (1|region) + 
> (pop|region) + (family|pop), data=fem1)
> 
> The variables are coded with unique levels from inner nested factors as 
> recomended here (Bates & Pinheiro : lme for SAS PROC MIXED users)  :
> http://biostat.hitchcock.org/FacultyandStaff/OnlineManuals/PDF%20Files/lmesas.pdf
> 
> Which syntax is the right one and describe de nested structure correctly? 
> And what could be the meaning of the wrong model?
> Is there somewhere general information about lmer synthax that we could have 
> missed  (not just simple examples)?
> (I just have an article D. Bates from Rnews vol5/1 and a book of Mr Galwey 
> in addition to the lme4 package help).
> 
> 
> I have also tried lme  (without the covariate) :
> But the denominator DF seem very strange to me considering the containment 
> method that is used, so I wonder also if the syntax that I use is correct :
> 
>> model5 <-lme(eclw~landsc + temp + landsc*temp , random= ~ 
>> 1|region/pop/family ,method="REML", data=femr)
>> anova.lme(model5)
>             numDF denDF  F-value p-value
> (Intercept)     1   332 546.0825  <.0001
> landsc          1     9   2.8841  0.1237
> temp            1   332  25.7565  <.0001
> landsc:temp     1   332   0.4316  0.5117
> 
> The number of levels of the factors are : temp : 2 ; landsc : 2 ; region : 2 
> ; pop : 12 ; family : 34
> If I'm not wrong the containment method use the same denominator DF as the 
> classical Anova approach.
> So here landsc would have to be tested against landsc*region with (2-1) * 
> (2-1) = 1 denominator DF.
> And the same for temp...
> 
> 
> 
> 
> ________________________________
> 
> Gilles San Martin y Gomez
> 
> Biodiversity Research Centre
> Ecology & Biogeography Unit
> University of Louvain-La-Neuve (UCL)
> Croix du Sud 4/5
> B-1348 Louvain-la-Neuve
> Belgium
> 
> Tel. +32 (0)10 47 21 73
> E-mail: gilles.sanmartin at gmail.com
> 
> ______________________________________________
> R-help at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.
> 

-- 
David C. Howell
PO Box 770059
627 Meadowbrook Circle
Steamboat Springs, CO
80477