[R] Structural Equation Models(SEM)

[Ralf Finne]

Thank you Jeremy for your information.
The world is changing though.  We live in an
increasing economic pressure. One symptom is that
we are forced to use smaller samples for economy.
This explains the interest for research in how the
methods perform on small samples.  The cited 
large simulation study shows GLS is more efficient
for small and badly distributed samples. 
So perhaps it would be worth the effort to make
an addition to sem(sem)

Yours
Ralf Finne
Emeritus
Novia University of Applied Science
Vasa Finland


>>> Jeremy Miles <jeremy.miles at gmail.com> 12/02/09 8:22 PM >>>
In the world of SEM, GLS has pretty much fallen by the wayside - I
can't recall anything I've seen arguing for it's use in the past 10
years, and I also can't recall anyone using it over ML.   The
recommendations for non-normal distributions tend to be robust-ML, or
robust weighted least squares.  These are more computationally
intensive, and I *think* that John Fox (author of sem) has written
somewhere that it wouldn't be possible to implement them within R,
without using a lower level language - or rather that it might be
possible, but it would be really, really slow.

However, ML and GLS are pretty similar, if you dug around in the
source code, you could probably make the change (see,
http://www2.gsu.edu/~mkteer/discrep.html for example, for the
equations; in fact GLS is somewhat computationally simpler, as you
don't need to invert the implied covariance matrix at each iteration).
 However, the fact that it's not hard to make the change, and that no
one has made the change, is another argument that it's not a change
that needs to be made.

Jeremy



2009/12/2 Ralf Finne <Ralf.Finne at novia.fi>:
> Hi R-colleagues.
>
> I have been using the sem(sem) function.  It uses
> maximum likelyhood as optimizing. method.
> According to simulation study in Umeå Sweden
>
(http://www.stat.umu.se/kursweb/vt07/stad04mom3/?download=UlfHolmberg.pdf
> Sorry it is in swedish, except the abstract)
> maximum likelihood is OK for large samples and normal distribution
> the SEM-problem should be optimized by GLS (Generalized Least
Squares).
>
>
> So to the question:
>
> Is there any R-function that solves SEM with GLS?
>
>
> Ralf Finne
> Novia University of Applied Science
> Vasa  Finland
>
> ______________________________________________
> R-help at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide
http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.
>



-- 
Jeremy Miles
Psychology Research Methods Wiki: www.researchmethodsinpsychology.com