If X- is zero, then there is no variation in X.  Thus, there can be no
covariation... The issue is logical, not computational.  I can't think of
any method that can show association in the absence of variation in one of
the variables...

WBW


On 21 Aug 2003, Garbriel wrote:

> I have a question. Thanks a lot for your help,
> It is about statistics in computing phi correlation coefficient: ( which
> is a special formula for the Pearson product-moment correlation in the
> particular case of two binary variable.)
>
> We used the method described in
>
> http://www.cmh.edu/stats/definitions/phi.htm
>
> See the following contigency table ,
>
>      Y+    Y-
>    -------------
> X+ |a11 | a12|
>    ------------
> X- |a21 | a22|
>    -----------
>
> Defination of Phi,
>
>       a11*a22 - a21*a12
> phi= ---------------------------------------------
>       sqrt((a11+a12)*(a21+a22)*(a11+a21)*(a12+a22))
>
> phi>0.7 strong correlation.
> phi>0.3 weak correlation
>
>
> However, there is limitations in this calculation,
>
> since if X- is zero, so, a21=0, a22=0;
> *************************************
> There will appear 0/0.
>  (and also big problem in p_value calculation).
>
> How to solve this? Is there any other method to quatitfy this when
> denominator (4 sums)  appear 0? ( I think phi correlation coeeficient is
> not qualified to this situation.) And also compute p value or other
> statistics term in this situation?
> .
> .
> =================================================================
> Instructions for joining and leaving this list, remarks about the
> problem of INAPPROPRIATE MESSAGES, and archives are available at:
> .                  http://jse.stat.ncsu.edu/                    .
> =================================================================
>

.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
.                  http://jse.stat.ncsu.edu/                    .
=================================================================

Reply via email to