In article <Rv%t4.2498$[EMAIL PROTECTED]>,
William Chambers <[EMAIL PROTECTED]> wrote:
>Rubin said:
>>As someone who has worked on the foundations, I suggest you
>>look at the real problem. In principle, you start out by
>>considering every possible theoretical model, and you use
>>the data to combine with your outlook to produce results.
>No. I do not look at every possible model. I prefer hypothesis testing, But
>if someone wants to use CR to examine every possible model, they will come
>up with one single result,,, the causal model that is implicit IN THE DATA.
>They should cross validate their findings, however, because even CR is
>subject to chance variation.
As normally used, hypothesis testing is just plain WRONG.
That lower-dimensional null hypothesis is rarely tenable,
and even if it is, such as the speed of light in vacuum
being constant, it is never directly tested. Also, one
has to balance incorrect acceptance, whateve that means,
with incorrect rejection.
>>In practice, you cannot do this exactly, as it would
>>require an infinitely large and infinitely fast computer
>>operating with zero cost. But it does tell you that much
>>of the current statistical religion is wrong.
>This warning should be sent to the traditional structural equation modeling
>folks using LISREL etc. They are the ones who have the problem with
>ambiguity. With CR the data support only one model, unlike the case with
>LISREL.
The data can never support only one model.
If you are simply suggesting that we have to know everything to
>know anything then you are saying that only an omniscient God can have
>anything to say in science, I do not think things are that bad, There is a
>reasonable faith that keeps science from grinding to a halt because we do
>not understand every tick in the Laplace clockwork universe,
The Laplace clockwork universe is not even accepted by many.
While it is not possible to do things precisely, one can
make good approximations.
>> The polarization effect is there. Have you paid any
>>>attention to the posts on corresponding correlations/regressions?
>>WHY should one look at correlations or regressions?
>Because we can not do experiments on many of the variables of interest in
>the social sciences, Manipulation of these variables may be impossible or
>unethical, So we have to make do with measurement and correlation/regression
>methods, I cover something of the history of this problem in my latest
>paper, with a bit about the place of astronomy as a nonexperimental but
>mathematical science,
So what? There are many situations in the biological and
social sciences where current theories are all nonlinear.
Are there any cases where the analysis of a set of genes IS
linear? There are many ways to analyze data not relying on
correlations and regressions.
>>Are
>>these linear relations even approximately correct? Using
>>linear approximations is reasonable for SMALL effects, but
>>those using correlations and regressions usually have large
>>ranges for their variables.
>First, larger variances are desirable in correlational research, It is why
>use factor analysis, The eigenvector/value solutions maximize variances,
Do you understand factor analysis? Anderson and I wrote one
of the few mathematically sound papers on the subject, appearing
in the proceedings of the Third Berkely Symposium.
..............
>What you are saying, however, is that it is pointless to look for linear
>relationships because they do not exist in nature,
They can be useful approximations, or the can be poor. In
any case, they must be justified.
I do not think things are
>that bad, And certainly 90% or more of the statistical analyses done assume
>the possibility of linearity,
Practitioners of ritual abound; most find it unable to
leave the position, no matter how easily it is shown to
be inappropriate.
Some would even say that nonlinear
>relationships can be broken down into a set of discrete linear stages or
>subsequences,
This is nonsense, unless you allow an infinite set.
Such slicing may be better than just giving up on linearity,,
>especially since most people cannot understand linear theories, much less
>nonlinear ones,
I see no reason why a non-linear theory is any harder to
understand than a linear one. Most physical relationships
have major aspects of non-linearity.
..............
--
This address is for information only. I do not claim that these views
are those of the Statistics Department or of Purdue University.
Herman Rubin, Dept. of Statistics, Purdue Univ., West Lafayette IN47907-1399
[EMAIL PROTECTED] Phone: (765)494-6054 FAX: (765)494-0558
===========================================================================
This list is open to everyone. Occasionally, less thoughtful
people send inappropriate messages. Please DO NOT COMPLAIN TO
THE POSTMASTER about these messages because the postmaster has no
way of controlling them, and excessive complaints will result in
termination of the list.
For information about this list, including information about the
problem of inappropriate messages and information about how to
unsubscribe, please see the web page at
http://jse.stat.ncsu.edu/
===========================================================================
