Skip to content

comparing 3 levels of fixed factor in lme4

11 messages · John Maindonald, Obermeier Andrew, Luca Borger +2 more

#
In lme4, in models with 3 levels of the fixed factor, each of these gets a t value comparing it to a reference level.

How is this done? 

It is my understanding that the t value can only be used to compare 2 means.
#
On Jul 30, 2012, at 08:04 , Obermeier Andrew wrote:

            
Then your understanding is wrong, and you need to read a text on basic linear modelling theory. Nothing specifically mixed-model or even R relevant about that.
#
Perhaps Andrew has vaguely at the back of his mind the notion
that for comparing >2 means, one should be using a multiple
range test or an anova test, at all events if the aim is to achieve
an experiment-wise 5% level.  Tests based on the individual
t-statistics are not independent.  This is of course a somewhat
controversial area.  
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.
http://www.maths.anu.edu.au/~johnm
On 30/07/2012, at 8:38 PM, peter dalgaard wrote:

            
#
Thank you John Maindonald.

For more than 2 levels of the experimental condition, I learned that usually we find an F to test mean differences across the levels of the condition.

In lme4, the model summary reports a t value, and I am replicating a study that uses lme4 to compare 3 levels of a fixed factor. My advising professor has told me that the t value can only be used to compare 2 means.

Andrew Obermeier
On Jul 30, 2012, at 8:00 PM, John Maindonald <john.maindonald at anu.edu.au> wrote:

            
#
If you have a specific book to suggest I would appreciate it. I have Pinhiero and Bates (2000), Mixed Effects Models in S and S-Plus, but have not found specific mentioned of how the t value is derived there yet.

I admit my ignorance in this area, and apologize if my questions are dumb, but nevertheless would appreciate some civility.

Andrew Obermeier
#
Well, you can use the t-statistics for comparison-wise tests!
the issue is whether one ought to do this, or whether one
should do some kind of overall test.  As I see it, all depends 
on the purpose that is in mind.

As Professor Dalgaard says, the issue is much the same as 
for lm models.  

You can for example do:
numDF denDF F-value p-value
(Intercept)     1    67   72.39  <.0001
fert            2    67    3.94  0.0241
variety         1    67   25.48  <.0001

As the design is balanced, the order of terms does not affect the anova F-test.
But as the design is balanced, you be better to do:
Error: Block
          Df Sum Sq Mean Sq F value Pr(>F)
Residuals  1   3528    3528               

Error: Within
          Df Sum Sq Mean Sq F value  Pr(>F)
fert       2   7019    3509    3.94   0.024
variety    1  22685   22685   25.48 3.7e-06
Residuals 67  59657     890                

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.
http://www.maths.anu.edu.au/~johnm
On 30/07/2012, at 9:38 PM, Obermeier Andrew wrote:

            
#
Hello,

 >I have Pinhiero and Bates (2000), Mixed Effects Models in S and 
S-Plus, but have not found specific mentioned of how the t value is 
derived there yet.

Did you check ch. 2.4 "Hypothesis tests and confidence intervals"?

HTH

Cheers,
Luca



------------------------------------------------------------
Luca Borger
Postdoctoral Research Fellow
Centre d'Etudes Biologiques de Chiz?
CNRS (UPR1934); INRA (USC1339)
79360 Villiers-en-Bois, France

Tel:    +33 (0)549 09 96 13
Fax:    +33 (0)549 09 65 26
email:  lborger at cebc.cnrs.fr
Skype:  luca.borger at skype.com
Web:    http://www.cebc.cnrs.fr/Fidentite/borger/borger.htm
         http://cnrs.academia.edu/LucaBorger
------------------------------------------------------------
# Forthcoming book chapter
# Dispersal Ecology and Evolution (ch. 17)
# http://ukcatalogue.oup.com/product/9780199608904.do
###
# New reprint! Animal Migration: A synthesis (ch. 8):
# http://ukcatalogue.oup.com/product/9780199568994.do

Le 30/07/2012 13:51, Obermeier Andrew a ?crit :
#
On 7/30/2012 1:51 PM, Obermeier Andrew wrote:
[...]
You may consider to read Pinheiro & Bates more carefully or consult an 
introductory text about model selection and likelihood ratio tests, e.g.:

Johnson, J., G. & Omland, K. S. (2004) Model Selection in Ecology and 
Evolution. Trends in Ecology and Evolution, 19, 101-108.
>
 > Andrew Obermeier

... so please add a little background (or an email signature) so that it 
is possible to find an appropriate scientific and/or technical level in 
our answers.

Hope it helps


Thomas
#
Thank you.

I'll read Pinhiero and Bates more carefully.

Andrew Obermeier
Doctoral Candidate, Temple University, Japan.
On Jul 30, 2012, at 9:24 PM, Thomas Petzoldt <Thomas.Petzoldt at TU-Dresden.de> wrote:

            
#
On Jul 30, 2012, at 13:51 , Obermeier Andrew wrote:

            
You probably need something from the level just below MEMSS. There's Maindonald & Braun, although I don't have a copy to hand just now. You might also take a look at the Faraway PDF (or  buy his more recent book) at

http://cran.at.r-project.org/other-docs.html

Specifically, take a look chapter 15 and chapter 3.1--3.

(And apologies if I came across a bit abrasive, but you did ask the same question about four times...)
#
Thank you very much Professor Dalgaard. 

I'll get the books by Maindonald & Braun and also Faraway, and meanwhile read the pdf you suggested.

Yes, MEMSS is a bit difficult for me.

I'm very sorry for all the noise I have made.

Sincerely,

Andrew Obermeier
Doctoral Candidate, Temple University Japan