Skip to content

Unbalanced ANOVA in R?

6 messages · Fredrik Karlsson, vito muggeo, John Fox +1 more

Original
Fredrik Karlsson
# 
Hi all

I'm trying to complete a textbook example originally designed for SPSS
in R, and I therefore need to find out how to compute an unbalanced
ANOVA in R.

I did a search on the mailinglist archives an found a post by Prof.
Ripley saying one should use the lme function for (among other things)
unbalanced ANOVAs, but I have not been able to use this object.
My code gives me an error.. Why is that ?
Any ideas? How do I get an ANOVA computation that can handle uniqual
sampe sizes in R?

/Fredrik

-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-
r-help mailing list -- Read http://www.ci.tuwien.ac.at/~hornik/R/R-FAQ.html
Send "info", "help", or "[un]subscribe"
(in the "body", not the subject !)  To: r-help-request at stat.math.ethz.ch
_._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._
Brian Ripley
#
On Fri, 8 Mar 2002, Fredrik Karlsson wrote:

            

      
      
      
    

        
I said use lme, *not* aov and lme.  Note, though, that the advice applied
to multistrata unbalanced anova, and this appears to be a two-way layout in
one stratum.  You can use aov for that. (You can't use lme unless there are
two or more strata.)  Just be careful how you interpret the output: in SAS
parlance aov gives you `Type I' sums of squares.

The reason why this is more difficult in R is that R unlike SPSS does not
guess which of the many possible interpretations you mean.  We would need
to know a lot more about the actual statistical problem to tell you what
would be a good analysis in R.

Maybe you should look at a textbook designed for use with S/R not SPSS?

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      vito muggeo
    

    

      
      #
    

  


  

      
The lme() function doesn't need aov(). So you would type:

lme(DELAY ~ DOSE + TRIALS,data=epinuneq)

however see ?lme o some book about mixed models: you need to specify the
random part and "clustering" variable (e.g. id-subject in a repeated
measurement design). Something like

lme(DELAY ~ DOSE + TRIALS, random=~1|your_cluster_variable, data=epinuneq)

best,
vito

After all, I don't think that simple "unbalanced ANOVA" and mixed models are
the same!


----- Original Message -----
From: "Fredrik Karlsson" <fredrik.karlsson at ling.umu.se>
To: "r-help-request Mailing List" <r-help at r-project.org>
Sent: Friday, March 08, 2002 6:56 AM
Subject: [R] Unbalanced ANOVA in R?

            

      
      
      
    

        
-.-.-

            

      
      
      
    

        
http://www.ci.tuwien.ac.at/~hornik/R/R-FAQ.html

            

      
      
      
    

        
_._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._.
_._
-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-
r-help mailing list -- Read http://www.ci.tuwien.ac.at/~hornik/R/R-FAQ.html
Send "info", "help", or "[un]subscribe"
(in the "body", not the subject !)  To: r-help-request at stat.math.ethz.ch
_._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._

  
  


  


      

    

    

      
      

        


  

        

      John Fox
    

    

      
      #
    

  


  

      
Dear Fredrik,

        
At 06:56 AM 3/8/2002 +0100, Fredrik Karlsson wrote:

            

      
      
      
    

        
The lme function is designed for linear mixed models, that is, models 
including both fixed and random effects. It's hard to tell from the 
information that you provide whether a mixed model is appropriate for your 
data -- in particular, is TRIALS a repeated-measures factor? If you indeed 
need to fit a mixed model, then you haven't specified it correctly. See the 
help page for lme for details.

On the other hand, if you just want to fit a fixed-effects Anova with no 
interactions, then you can use the aov or lm functions. Again, see 
help(aov) and help(lm). The summary method for aov and the anova method for 
both aov and lm objects will give you sequential ("type I") sums of 
squares. The Anova function in the car package can compute "Type II" and 
"III" sums of squares for lm objects. If you use the latter, be careful 
with the contrast coding.

I hope that this helps,
  John

-----------------------------------------------------
John Fox
Department of Sociology
McMaster University
Hamilton, Ontario, Canada L8S 4M4
email: jfox at mcmaster.ca
phone: 905-525-9140x23604
web: www.socsci.mcmaster.ca/jfox
-----------------------------------------------------

-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-
r-help mailing list -- Read http://www.ci.tuwien.ac.at/~hornik/R/R-FAQ.html
Send "info", "help", or "[un]subscribe"
(in the "body", not the subject !)  To: r-help-request at stat.math.ethz.ch
_._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._

  
  


  


      

    

    

      
      

        


  

        

      Fredrik Karlsson
    

    

      
      #
    

  


  

      
Thank you Professor Ripley for the quick and informative response.

Being unable to choose a different textbook (even though I must admit
that this SPSS obsession is starting to annoy me).
I solved the problem though.

I changed the subject for this post in order to change the topic into
a more general one:

What does aov() do and what does it NOT do? 
Prof. Ripley mentioned that you should use the lme() function  for
multistrata anova, so that's one. What else? (Realising that this is a
question with an open set of answers).

My impression is that what I am suppose do do in SPSS under some heading
n-way anova, ancova,  one-way glm or whatever is handled in S/R by aov()

Does this mean you can use aov() to compare two groups with different
ns?

Many open questions, I know, but the SPSS -> R translation is troubling
for me.

Please, could someone sort this out for me?

/Fredrik

        
On Fri, Mar 08, 2002 at 06:24:27AM +0000, Prof Brian D Ripley wrote:

            

      
      
      
    

        

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      Brian Ripley
    

    

      
      #
    

  


  

      
On Fri, 8 Mar 2002, Fredrik Karlsson wrote:

        

            

      
      
      
    

        
You could read a textbook on S/R which would explain what aov does.  As I
understand it John Fox's car package is designed (amongst other things) to
make R easier for diehard SPSS users, so you ought to find that useful.

            

      
      
      
    

        
No, for *unbalanced* multistrata anova (and that's the second time you
have quoted me out of context: please be more careful!).  For balanced
desigms, aov()  is the preferred tool.  It's also general for unbalanced
fixed-effects models, but for some lm will be able to interpret.

            

      
      
      
    

        
You can, but that's the job of t.test.

            

      
      
      
    

        
It's troubling for us.  SPSS is often grossly misleading in its
terminology.