In cleaning out my INBOX I came across this message, to which I had
intended to reply some time ago.  A possibly useful reference (if you
have access to it), along the lines you asked about, is
 Goodman, L.  Partitioning of chi-square, analysis of marginal
contingency tables, and estimation of expected frequencies in
multidimensional contingency tables.  Journal of the American
Statistical Association 66 (1971), 339-344.

An approach that I have sometimes found useful is the following.

1.  If H0:  [rows & columns are independent] has been rejected, display
the standardized residuals for each cell.  If these are all about the
same size, then the non-independence is, so to speak, distributed more
or less evenly throughout the table and it may be difficult to identify
useful patterns to extract.  Usually, however, there are a few cells
with large residuals (and therefore supplying large contributions to the
total chi-square value).  If this is the case, it may be useful to
remove the effects of these cells (one cell at a time, or several cells
at a time, as seems convenient and/or appropriate), as follows.

2.  For the cell whose effect is to be removed:  substitute for the
observed frequency in that cell a fictitious frequency obtained from the
expected frequency.  (Actually, for reasons I can adduce later but won't
bother with now, it is usually efficient to overcorrect somewhat.)
Using the fictitious frequency, recalculate the chi-square value, and
observe the contribution to the total chi-square (or the standardized
residual, if you prefer) in the "adjusted" cell.  What you want to do is
to adjust the fictitious frequency until that standardized residual is
zero, or as close to zero as the chi-square routine will permit.  (If,
as in MINITAB, one is restricted to integer values in the contingency
table, you won't get exactly zero, but you can get to a negligible
value.)  With a little practice, one can get to the optimal adjustment
in three or four successive approximations.

3.  With the contribution of this cell reduced to zero, or as near to it
as you can get, examine the standardized residuals of the other cells
(and the total chi-square value) in the adjusted table, to see if there
are other potentially interesting effects to be extracted.  If so,
repeat the procedure for one or more additional cells.
 (Note that every time you've adjusted the frequency in a cell, the
number of degrees of freedom is reduced by 1.  Your software will not
know that, so you'll have to refer the reported value of chi-square to a
table after you start this process.  On the other hand, any individual
cell contribution can be referred to the critical value of chi-square
with one d.f. (which is conservative:  since you have a table with r*c
cells and only (r-1)*(c-1) d.f. to start with, it can be argued that the
initial critical value for one cell is for chi-square with (r-1)(c-1)/rc
d.f.; tables of chi-square for fractional d.f. are not common, however).

It is demonstrable that a single cell in a r-by-c table, if it departs
strongly from the expected value, can make other effects in the table
less easy to see.  One can think of reducing the effect of the current
cell to zero as rather like removing an object from the foreground of a
landscape, so that objects in the middle distance or in the background
can the more easily be perceived.

You may find the following example interesting.  Some years ago, a PhD
candidate in linguistics was analyzing the frequencies with which
aphasic patients mispronounced vowels in ordinary English.  This more or
less naturally led to a square contingency table in which the rows
represented the "target" vowel (the correct sound for the word being
spoken) and the columns represented the actual vowel sound produced;  it
was as I recall an 11-by-11 table.  But of course it didn't only contain
errors:  its major diagonal contained frequencies with which the vowels
were correctly pronounced.  And as one might expect, the total
frequencies on that diagonal represented rather more than 75% of the
data.  An ordinary chi-square analysis wuld show large positive
residuals on that diagonal, and a highly significant chi-square value;
while effects occurring in the errors, which were the focus of the
study, were pretty well obliterated.  And it wouldn't do to substitute
zero frequencies on the diagonal;  you'd have much the same effect in
reverse.  Replacing the diagonal observed frequencies with fictitious
frequencies, chosen so as to make the contribution of that whole
diagonal as near to zero as possible, permitted one to analyze the
errors these speakers actually made:  one could identify vowel sounds
that more frequently tended to be substituted incorrectly as well as
those that were less frequently substituted than would be expected.
 (And one could even interpret the fictitious frequencies, as those
errors that accidentally produced the correct vowel, in a manner of
speaking.)
 The candidate did have some physiological reasons for expecting certain
kinds of errors to be made preferentially, and was able to show evidence
that these errors did occur relatively frequently.
  -- DFB.

On Thu, 30 Oct 2003, VOLTOLINI wrote:

> Hi folks,
>
> My doubt is about the chi-square test.
>
> I heard from a colleague that....
>
> When the Ho of independence is rejected in a contingency table you can
> compare the partial chi-square values with the chi-square value from
> the table to check if there is one or more partial chi-square values
> that could reject the Ho alone.
>
> Is it true?
>
> I cannot find this idea in any classic stats text book :(
>
> Thanks for any help !!!
>
> Voltolini

 ------------------------------------------------------------
 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