Skip to content

chisq.test

8 messages · Omphalodes Verna, arun, Peter Ehlers +4 more

Original
arun
# 
Hi,

The error is due to less than 5 observations in some cells.

You can try,
fisher.test(tabele)
??? Fisher's Exact Test for Count Data

data:? tabele 
p-value = 0.0998
alternative hypothesis: two.sided 

A.K.



----- Original Message -----
From: Omphalodes Verna <omphalodes.verna at yahoo.com>
To: "r-help at r-project.org" <r-help at r-project.org>
Cc: 
Sent: Tuesday, June 26, 2012 2:27 PM
Subject: [R] chisq.test

Dear list!

I would like to calculate "chisq.test" on simple data set with 70 observations, but the output is ''Warning message:''

Warning message:
In chisq.test(tabele) : Chi-squared approximation may be incorrect


Here is an example: 

??? ??? tabele <- matrix(c(11, 3, 3, 18, 3, 6, 5, 21), ncol = 4, byrow = TRUE)
??? ??? dimnames(tabela) <- list(
??? ??? "SEX" = c("M","F"),
??? ??? "HAIR" = c("Brown", "Black", "Red", "Blonde"))
??? ??? addmargins(tabele)
??? ??? prop.table(tabele)
??? ??? chisq.test(tabele)
Please, give me an advice / suggestion / recommendation.

Thanks a lot to all, OV

??? [[alternative HTML version deleted]]


______________________________________________
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.
Peter Ehlers
#
On 2012-06-26 11:27, Omphalodes Verna wrote:
Do this:

   ct <- chisq.test(tabele)
   ct$expected

If that does not give you a sufficient hint, then you need
to review the assumptions underlying the chisquare test.

Peter Ehlers
David L Carlson
# 
The warning means that you have many cells with expected values less than 5
(4 of 8 cells in this case) so that the chi square estimate may be inflated.
The good news is that the probability of the inflated chi square is .0978
which you probably would not consider to be significant anyway. If you want
to get a simulated p value using Monte Carlo simulation (see the references
in the manual page for chisq.test), just change the call to

chisq.test(tabele, simulate.p.value=TRUE, B=2000)

When I run this five times, I get probability estimates ranging from .09795
to .1089.

Alternatively, get more data.

----------------------------------------------
David L Carlson
Associate Professor of Anthropology
Texas A&M University
College Station, TX 77843-4352
David Winsemius
#
On Jun 26, 2012, at 2:27 PM, Omphalodes Verna wrote:

            

      
      
      
    

        
Read any introductory stats book regarding  small cell sizes:

      [,1] [,2] [,3] [,4]
[1,]   11    3    3    5
[2,]    3   18    6   21

            

      
      
      
    

        

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      Rolf Turner
    

    

      
      #
    

  


  

      
On 27/06/12 08:54, arun wrote:

            

      
      
      
    

        
NO, NO, NO!!!!  It's not the observations that matter, it is
     the ***EXPECTED COUNTS***.  These must all be at least
     5 in order for the null distribution of the test statistic to be
     adequately approximated by a chi-squared distribution.

         cheers,

             Rolf Turner

            

      
      
      
    

        

      
      
    

  
  


  


      

    

    

      
      

        


  

        

      John
    

    

      
      #
    

  


  

      
On Wed, 27 Jun 2012 16:58:29 +1200

        
Rolf Turner <rolf.turner at xtra.co.nz> wrote:

        

            

      
      
      
    

        
Pretty sure the point was that in a situation where the expected counts
are too low for a reliable chi-square, that an alternate test such as
the nonparametric Fisher's Exact Test may be the way to go, especially
if there isn't nay more data to get. That way you don't have to worry
about expected counts.

JDougherty

  
  


  


      

    

    

      
      

        


  

        

      Peter Ehlers
    

    

      
      #
    

  


  

      
On 2012-06-26 23:02, John wrote:

            

      
      
      
    

        
That may well be; nevertheless, the post included the statement
Rolf quotes: "... less than 5 _observations_ in some cells" (my
emphasis). And Rolf's point is quite correct - it's the
_expected_ counts that the approximation cares about.

Peter Ehlers