I did not see the original post.  From the information in the one
paragraph Rich quotes, Erica's data set has fewer cases than her
contingency table has cells:  which implies either that it's a humungous
great table, or a really tiny data set, or both.  In any case, it does
not sound as though Erica has enough information to obtain a reliable
measure of association under any circumstances or via any coefficient.

If the description above (very many cells, or very few cases) is not
true of her data set, there must be some serious computational or
data-processing error in her procedures somewhere.
 I'd be curious to know how she gets expected values less than 1 in
EVERY cell, as she claims.
  -- Don Burrill.

On Sun, 15 Feb 2004, Rich Ulrich wrote:

> On Sat, 14 Feb 2004 12:16:13 +0800, "Erica So"
> <[EMAIL PROTECTED]> wrote:
>
> > One of the requirement on conducting Chi-Square test of association
> > is 'no more than 20% of the expected values should be less than
> > five', however, all of my data are less than 1. Is there any other
> > association coefficient allow expected values less than 1 in SPSS?
> >
>
> I've seen that happen with tests that are used for
> checking whether a Pseudo-Random Number Generator
> (that is, computer algorithm)  is really working.  What
> I remember reading about it, is that for *that*  application,
> the chisquared test is considered  useful, despite the
> rule-of-thumb.

I suspect its utility resided in data sets where only a few cells had
tiny expected frequencies, and in which those cells had observed
frequencies of zero.  The reason for the arbitrary rule (no cells, or
few cells, with expected frequency less than 5) is that one's data can
only be integers;  if <exp> = 0.2, say, the contribution of that cell to
the total chi-square value is 0.2 if <obs> = 0 (in which case there's
probably not much of a problem) but if <obs> = 1 (the next bigger
possible value) that contribution is 3.2, and if <obs> = 2 the
contribution is 16.2;  one can thus obtain a "significant" value of
chi-square that is wholly due to only one or two observations.  -- DFB.

> Even for that, I think I would look at both tests that SPSS
> provides, the Pearson and the Likelihood chisquared.
>
> For anything other than that, I would be suspicious,
> and I would try to convert the problem.  (Do they have
> to be categories?  Can't there be fewer categories?)
> For some problems that I can imagine, you could
> make statements about exact probabilities

 ------------------------------------------------------------
 Donald F. Burrill                              [EMAIL PROTECTED]
 56 Sebbins Pond Drive, Bedford, NH 03110      (603) 626-0816
.
.
=================================================================
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