In article <[EMAIL PROTECTED]>,
EugeneGall <[EMAIL PROTECTED]> wrote:
>>Andrew Morse wrote:
>>>Who was the first to say "Correlation does not imply causation" in so many
>>>words? I know that the idea dates back to David Hume, but Hume did his
>>>work about a century before the term "correlation" acquired its modern
>>>statitical meaning.
>>It certainly wasn't Hume, who's argument would be 'Causation is nothing but
>>correlation'
Wright understood this, and other, statistical problems.
He was astute enough to evaluate his models on calculated
"population" data algebraically, and to see that the usual
correlation methods would not work. Essentially, he saw
the difference between structural models and prediction
models, apparently well before anyone else. He still
fitted his models crudely by using covariances, and quite
often by correlations, which was not always best, but this
was fitting population moments to sample moments, not an
unreasonable method of inference.
Econometricians saw the same problem in the late 30s, in
ignorance of Wright's work, including economic modeling!
My results with Anderson on limited information estimation,
and the later work of Theil and others on two-stage and
three-stage least squares, are right in line with Wright's
work, although derived from attempts to use more powerful
methods on structural models. Knowledge of what Wright did
would have speeded up the work somewhat in the 30s and 40s.
Other than not using more efficient methods of inference,
I find it hard to see why anyone should reject path
coefficients. Prediction models are even confusing,
fitting them yields biased estimators, and structural
models cover much more.
I heard Wright speak once, and recognized that he had the
structural approach long before those with whom I was
working. It might have helped get started, but their
problems were in inference in more complicated situations,
not structural modeling, which was reached in a slightly
different manner.
>>William Provine's biography of Sewall Wright includes a discussion with
>>Wright
>>about his introduction of path analysis (in 1918 I think) in a paper entitled
>>'Correlation & Causation' A Humean critic said the title was redundant.
>I've now had a chance to check William B. Provine's superb biography, "Sewall
>Wright and Evolutionary Biology (1986 U of Chicago Press).
>Wright first described path analysis in 1918, but formally presented his method
>of path coefficients in 1921 in
>Wright, S. B. Correlation & causation. J. Agr. Res. 20: 557-585.
>On p. 142-144 of Provine's biography, Provine writes:
>"Wright's method of path coefficients did not gain immediate acceptance or
>attention, even after the application to the previously insuluble problems of
>systems of mating. The first substantial criticims appeared in Genetics ...
>The critic was Henry E. Niles, from Pearl's Department of Biometry and Vital
>Statistics, ... Johns Hopkins University. Niles was a thorough-going Humean
>familiar with modern correlation analysis; he naturally found even Wright's
>title, "Correlation and Causation," offensive. For Niles, as for Karl Pearson
>(who stated the belief in all editions of his Grammar of Science), causation
>was nothing more than correlation. Thus to contrast them was nonsense:
> "Causation" has been popularly used to express
> the condition of association, when applied to
> natural phenomena. There is no philosophical
> basis for giving it a wider meaning than partial
> or absolute association. In no case has it
> been proved that there is an inherent necessity
> in the laws of nature. Causation is correlation
> ... Perfect correlation, when based upon
> sufficient experience, is causation in the
> scientific sense (Niles 1922, 259, 261)
>Provine provides further discussion of the debate between Niles and Wright.
>Wright insisted that correlation and causation were different but that
>correlations could be used to test models describing causal connections among
>variables.
>At age 94, Sewall Wright defended his interpretation of causation and
>correlation presented in his 1918 & 1921 method of path coefficients which in
>response to a 1993 critique by Karlin.
>Wright argued that path analysis can be used
>to test causal models involving correlations and
>partial correlations. Path analysis was developed as a deductive testing
>procedure, not an inductive method for generating causal explanations.
>Correlations can be used to refute and decide among different causal models.
>His defense can be found in
>Wright, S. 1983. On "path analysis in genetic epidemiology: a critique" Am. J.
>Hum Genetics 35: 757-768.
--
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
=================================================================
Instructions for joining and leaving this list and remarks about
the problem of INAPPROPRIATE MESSAGES are available at
http://jse.stat.ncsu.edu/
=================================================================
