On Sun, 29 May 2005, John Fox wrote:
Dear Spencer,
-----Original Message-----
From: Spencer Graves [mailto:[EMAIL PROTECTED]
Sent: Sunday, May 29, 2005 4:13 PM
To: John Fox
Cc: r-help@stat.math.ethz.ch; 'Jacob van Wyk'; 'Eric-Olivier Le Bigot'
Subject: Re: [R] Errors in Variables
Hi, John:
Thanks for the clarification. I know that the
"errors in X problem"
requires additional information, most commonly one of the
variances or the correlation. The question I saw (below)
indicated he had tried "model of the form y ~ x (with a given
covariance matrix ...)", which made me think of "sem".
If he wants "the least (orthogonal) distance", could
he could get it indirectly from "sem" by calling "sem"
repeatedly giving, say, a variance for "x", then averaging
the variances of "x" and "y" and trying that in "sem"?
I'm not sure how that would work, but seems similar to averaging the
regressions of y on x and x on y.
Also, what do you know about "ODRpack"? It looks
like that might solve "the least (orthogonal) distance".
I'm not familiar with ODRpack, but it seems to me that one could fairly
simply minimize the sum of squared least distances using, e.g., optim.
Exactly. In fact this is easily reduced to a function of one variable
(the slope, known to lie between the y or x and x om y regressions) and so
optimize() would be more appropriate. I did do that in S once upon a long
time, but it seemed too esoteric to package up (and it would take me
longer to find the code than to do it again).
My paper quoted originally deals with the case of known variances for each
of x and y (and heteroscedasticity in both). It was written for chemists,
and contains all the formulae one needs. In their applications knowing
(at least approximately) the variances is a reasonable assumption.
Brian
Thanks again for your note, John.
Best Wishes,
Spencer Graves
John Fox wrote:
Dear Spencer,
The reason that I didn't respond to the original posting (I'm the
author of the sem package), that that without additional
information
(such as the error variance of x), a model with error in
both x and y
will be underidentified (unless there are multiple indicators of x,
which didn't seem to be the case here). I figured that what
Jacob had
in mind was something like minimizing the least
(orthogonal) distance
of the points to the regression line (implying by the way
that x and y
are on the same scale or somehow standardized), which isn't
doable with sem as far as I'm aware.
Regards,
John
--------------------------------
John Fox
Department of Sociology
McMaster University
Hamilton, Ontario
Canada L8S 4M4
905-525-9140x23604
http://socserv.mcmaster.ca/jfox
--------------------------------
-----Original Message-----
From: [EMAIL PROTECTED]
[mailto:[EMAIL PROTECTED] On Behalf Of
Spencer Graves
Sent: Saturday, May 28, 2005 4:47 PM
To: Eric-Olivier Le Bigot
Cc: r-help@stat.math.ethz.ch; Jacob van Wyk
Subject: Re: [R] Errors in Variables
I'm sorry, I have not followed this thread, but I
wonder if you
have considered library(sem), "structural equations modeling"?
"Errors in variables" problems are the canonical special case.
Also, have you done a search of "www.r-project.org"
-> search -> "R site search" for terms like "errors in
variables regression"? This just led me to "ODRpack",
which is NOT a
CRAN package but is apparently available after a Google
search. If it
were my problem, I'd first try to figure out "sem"; if that seemed
too difficult, I might then look at "ODRpack".
Also, have you read the posting guide!
http://www.R-project.org/posting-guide.html? This suggests, among
other things, that you provide a toy example that a potential
respondant could easily copy from your email, test a few
modifications, and prase a reply in a minute or so.
This also helps clarify your question so any respondants are more
likely to suggest something that is actually useful to you.
Moreover,
many people have reported that they were able to answer their own
question in the course of preparing a question for this
list using the
posting guide.
hope this helps. spencer graves
Eric-Olivier Le Bigot wrote:
I'm interested in this "2D line fitting" too! I've been looking,
without success, in the list of R packages.
It might be possible to implement quite easily some of the
formalism
that you can find in Numerical Recipes (Fortran 77, 2nd ed.),
paragraph 15.3. As a matter of fact, I did this in R but
only for a
model of the form y ~ x (with a given covariance matrix
between x and
y). I can send you the R code (preliminary version: I
wrote it yesterday), if you want.
Another interesting reference might be Am. J. Phys. 60, p.
66 (1992).
But, again, you would have to implement things by yourself.
All the best,
EOL
--
Dr. Eric-Olivier LE BIGOT (EOL) CNRS
Associate Researcher
~~~o~o~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
~~~~o~o~~~
Kastler Brossel Laboratory (LKB)
http://www.lkb.ens.fr
Université P. & M. Curie and Ecole Normale Supérieure, Case 74
4 place Jussieu 75252 Paris CEDEX 05
France
~~~o~o~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
~~~~o~o~~~
office : 01 44 27 73 67 fax:
01 44 27 38 45
ECR room: 01 44 27 47 12 x-ray room:
01 44 27 63 00
home: 01 73 74 61 87 For int'l calls: 33 + number
without leading 0
On Wed, 25 May 2005, Jacob van Wyk wrote:
I hope somebody can help.
A student of mine is doing a study on Measurement Error models
(errors-in-variables, total least squares, etc.). I have an old
reference to a "multi archive" that contains
leiv3: Programs for best line fitting with errors in both
coordinates.
(The date is October 1989, by B.D. Ripley et al.) I have done a
search for something similar in R withour success. Has this been
implemented in a R-package, possibly under some sort of
assumptions
about variances. I would lke my student to apply some regression
techniques to data that fit this profile.
Any help is much appreciated.
(If I have not done my search more carefully - my
apologies.) Thanks
Jacob
Jacob L van Wyk
Department of Mathematics and Statistics University of
Johannesburg
APK P O Box 524 Auckland Park 2006 South Africa
Tel: +27-11-489-3080
Fax: +27-11-489-2832
______________________________________________
R-help@stat.math.ethz.ch mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide!
http://www.R-project.org/posting-guide.html
------------------------------------------------------------
----------
--
______________________________________________
R-help@stat.math.ethz.ch mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide!
http://www.R-project.org/posting-guide.html
______________________________________________
R-help@stat.math.ethz.ch mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide!
http://www.R-project.org/posting-guide.html
______________________________________________
R-help@stat.math.ethz.ch mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide! http://www.R-project.org/posting-guide.html
--
Brian D. Ripley, [EMAIL PROTECTED]
Professor of Applied Statistics, http://www.stats.ox.ac.uk/~ripley/
University of Oxford, Tel: +44 1865 272861 (self)
1 South Parks Road, +44 1865 272866 (PA)
Oxford OX1 3TG, UK Fax: +44 1865 272595
______________________________________________
R-help@stat.math.ethz.ch mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide! http://www.R-project.org/posting-guide.html
