On Mon, 06 Nov 2000 18:36:16 +0100, Joseph McDonnell
<[EMAIL PROTECTED]> wrote:
> Gentlemen,
>
> I agree with both of you. Several correspondents had already pointed out that a
> logistic regression approach would be more appropriate in this situation.
> However, I was trying to steer the discussion in a slightly different direction,
> which I suspect may be the subject of Gerhard's interest (and of mine, but
> Gerhard can speak for himself). Perhaps I should rephrase the question to read
> 'What are the consequences of performing an analysis on a discreet dependent
> variable as if it were continuous?' And I'm thinking of situations other than
> just linear/logistic regression. For example, in an artificial reproduction
< snip, rest >
- the original question had been posted in the Usenet group for SPSS,
where Neil Henry wrote one direct answer to the actual question
(instead of wandering off to discuss logistic, etc.)
and I commented on his reply. Below, I am re-posting my comment and
Neil's.
========== copied from 2 Nov 2000, comp.soft-sys.stat.spss
To Gerhard Luecke's request,
"Can anyone name some references where the problem of using a
DICHOTOMOUS variable as a DEPENDENT variable in an ANOVA is
discussed?"
On Thu, 02 Nov 2000 15:28:02 -0500, "Neil W. Henry"
<[EMAIL PROTECTED]> wrote:
> Why would anyone consider this a PROBLEM?
> In large samples the oneway anova F test and the ordinary chi-square
> test from the crosstabulation will be equivalent.
>
As Neil says, one-way, between groups *testing* is the same.
If you were interesting in "linear trend" then the scaling would make
a difference, depending on what units are additive; similarly, scaling
matters more for two-way analyses (any analyses with multiple
variables).
There is a bit of general discussion on the topic of
probit vs. logit vs. other transformations
in DJ Finney's classical textbook,
"Statistical methods in biological assay."
================= end of Nov 2 posting
To give more detail:
( 1:1, 3:1, 9:1, 27:1 ) - is an example which is
equal-ratio-steps, or,
equal steps in the log-odds metric.
These are proportions of ( 50%, 75%, 90%, 96.4% ).
These are obviously different from equal-steps
in proportions, with differences of 25/ 15/ 7 ...
and are slightly different from equal
steps of a probit (normal) distribution
Fitting 0/1 responses with either a linear trend
across means or a multivariable model
can run into similar, artifactual evidence of 'non-fit'
(which, by the way, should be ignored).
--
Rich Ulrich, [EMAIL PROTECTED]
http://www.pitt.edu/~wpilib/index.html
http://www.pitt.edu/~wpilib/index.html
=================================================================
Instructions for joining and leaving this list and remarks about
the problem of INAPPROPRIATE MESSAGES are available at
http://jse.stat.ncsu.edu/
=================================================================
