| |
 |
|
|
Science Forum Index » Space - Consult Forum » Chi^2 square independence test & post-hoc test...
Page 1 of 1
|
| Author |
Message |
| sigbert... |
 Posted: Fri Jun 06, 2008 3:20 am |
|
|
|
Guest
|
Hi,
for a project we doing a lot of chi^2 independence tests between
categorical variables. Of course, we not only want to know that the
variables are dependent, but also which cells are the ones which are
responsible for the dependence. I found three ways how to flag the
significant cells and we decided to use the fact that the standardized
adjusted residuals are approximately N(0,1) distributed. Since the
approximation conditions for the independence test are sometimes not
fulfilled, we decided to use the Monte-Carlo version of the test.
Therefore I would like to apply the Monte-Carlo method also to the
standardized residuals. If I do so then I get an asymmetric acceptance
area around 0. My question is now: should I use this asymmetric
acceptance region around 0 or should I better consider the squared
residuals and the corresponding acceptance area?
Thanks in advance Sigbert |
|
|
| Back to top |
|
| Richard Ulrich... |
| Posted: Fri Jun 06, 2008 3:23 pm |
|
|
|
Guest
|
On Fri, 6 Jun 2008 06:20:07 -0700 (PDT), sigbert
<sigbert at (no spam) wiwi.hu-berlin.de> wrote:
Quote: Hi,
for a project we doing a lot of chi^2 independence tests between
categorical variables. Of course, we not only want to know that the
variables are dependent, but also which cells are the ones which are
responsible for the dependence. I found three ways how to flag the
significant cells and we decided to use the fact that the standardized
adjusted residuals are approximately N(0,1) distributed. Since the
approximation conditions for the independence test are sometimes not
fulfilled, we decided to use the Monte-Carlo version of the test.
Therefore I would like to apply the Monte-Carlo method also to the
standardized residuals. If I do so then I get an asymmetric acceptance
area around 0. My question is now: should I use this asymmetric
acceptance region around 0 or should I better consider the squared
residuals and the corresponding acceptance area?
Thanks in advance Sigbert
You have a lousy situation. Are you sure that you
want to pretend to great precision, when you have
multiple testing and correlated results? In addition,
and what is probably more important for readers, for each
of those tests, the power and sensitivity are strongly
tied to the marginal frequencies. - That is, you might want
to use "Odds ratio" in addition to p-level, in deciding
what extremes are worth flagging.
If you want best precision, you will use the asymmetric
confidence limits, like the ones that always result
from binomial comparisons with small frequencies.
Monte Carlo has little to do with it, except that this was
the way you generated them. On the other hand, the
p-level was generated by the squared differences, and
in my opinion, the whole description would probably
be easier and simpler if you dropped the Monte-Carlo
generation entirely. Use the standardized residuals.
--
Rich Ulrich
http://www.pitt.edu/~wpilib/index.html |
|
|
| Back to top |
|
| |
|
Page 1 of 1
All times are GMT - 5 Hours
The time now is Sun Nov 23, 2008 4:03 am
|
|