One could easily get the impression that these sorts of use are
almost the only ways in which factor analysis that goes beyond
principal components kinds of uses, and structural equation
modelling, are used. It is very hard to find examples of the use
of these techniques where there is some genuine and plausible
attempt at model validation. It is also very hard to devise and
implement good diagnostic statistics and plots. Such concerns are
well articulated in David Freedman's paper (with extensive
following discussion) "As Others See Us; A Case Study in Path
Analysis", in J. Educ. Statistics 12: 101-223
I think that Freedman's point is more general, and extends, for
example, to regression analysis, even supported by good diagnostics
-- he argues generally against drawing causal conclusions from
nonexperimental data. Although many of his points are well taken, I
find his argument extreme.
Thanks for that comment.
The arrow surely goes in the other direction. The points which
Freedman makes on regression in his "Statistical Analysis and Shoe
Leather" paper (Sociological Methodology 1991, pp. 291-358) apply also
to Path Analysis, Structural Equation Modelling, etc. There are
however additional points to be made, which have to do with the near
impossibility of checking empirically the hidden relationships that
are assumed.
Freedman does sometimes get extreme. I do not think it is fair to say
that he rules out the drawing of causal conclusions from
non-experimental data. Rather I understand his argument to be that
the data on its own, however analysed, are not enough. There is
extensive contextual information which, as he illustrates from John
Snow's work on the spread of cholera in the mid-19th C in London, has
to come together with the data. In Path Analysis, Structural Equation
Modelling etc., one has to be very sure of the contextual information
(what are are paths? what direction do they go?) for the analytic
results to have any credence. This is in a context where the
statistical assumptions, on variables that are mostly hidden, readily
get out of hand.
I'd very much like to learn of published examples of the use of
these methods which do address these concerns. If these can be
found, they would be good references to include when and if these
techniques are implemented.
I don't generally think of myself as an advocate of either factor
analysis or of structural-equation models, but I recognize that the
better practitioners of these methods pay some attention to the
descriptive adequacy of their models. Take a look, for example, at
Bollen, Structural Equations With Latent Variables (Wiley, 1989) and
Bollen and Long, eds., Testing Structural Equation Models (Sage,
1993).
John
|----------------------------------------------------|
| John Fox jfox at McMaster.ca |
| Department of Sociology McMaster University |
|----------------------------------------------------|
-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-
I guess if this discussion is to continue, it might be well for it to
continue off-list. It is well for list members to be aware that there
are issues here to consider. One service that R may perform is to bring
together communities of statisticians and statistical users which have
become too largely separate.
-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-
r-help mailing list -- Read http://www.ci.tuwien.ac.at/~hornik/R/R-FAQ.html
Send "info", "help", or "[un]subscribe"
(in the "body", not the subject !) To: r-help-request at stat.math.ethz.ch
_._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._