Skip to content

covariate data errors

4 messages · Andy Jacobson, John Fox, Duncan Mackay +1 more

#
Greetings,

	I would like to fit a multiple linear regression model in
which the residuals are expected to follow a multivariate normal
distribution, using weighted least squares.  I know that the data in
question have biases that would result in correlated residuals, and I
have a means for quantifying those biases as a covariance matrix. I
cannot, unfortunately, correct the data for these biases.

	It seems that this should be a straightforward task, but so
much of the literature is concerned with the probability model in
which the residuals are uncorrelated that I can't find a good
reference.  So in order of importance, please, can someone point me to
a definitive reference for least squares with correlated residuals,
and is there a standard R package to handle this case?

	Many thanks in advance,

	Anthony
#
Dear Anthony,

The gls (generalized least squares) function in the nlme package should do 
what you want. (I assume that your analysis leads you to expect an 
error-covariance matrix of a specific form with some free parameters to 
estimate.)

Generalized least squares estimation is a common topic in regression texts. 
You'll find a brief appendix on the subject from my R and S-PLUS Companion 
to Applied Regression, in the context of time-series regression, at 
<http://www.socsci.mcmaster.ca/jfox/Books/Companion/appendix-timeseries-regression.pdf>.

I hope that this helps,
  John
At 11:40 PM 6/12/2003 -0400, Andy Jacobson wrote:
-----------------------------------------------------
John Fox
Department of Sociology
McMaster University
Hamilton, Ontario, Canada L8S 4M4
email: jfox at mcmaster.ca
phone: 905-525-9140x23604
web: www.socsci.mcmaster.ca/jfox
#
Hello,
I'm not qualified to judge about definitiveness, but you may find the
discussion of "Time Series Regression and Generalized Least Squares" in
Chapter 14 of "Applied Regression Analysis, Linear Models, and Related
Models" by John Fox (1997, Sage Publications) helpful. He also has an
accompanying package "car" on the CRAN R site.
Duncan

*****************************************
Dr. Duncan Mackay
School of Biological Sciences
Flinders University
GPO Box 2100
Adelaide
S.A.    5001
AUSTRALIA

Ph (08) 8201 2627    FAX (08) 8201 3015

http://www.scieng.flinders.edu.au/biology/people/mackay_d/index.html


-----Original Message-----
From: r-help-bounces at stat.math.ethz.ch
[mailto:r-help-bounces at stat.math.ethz.ch]On Behalf Of Andy Jacobson
Sent: Friday, 13 June 2003 1:10 PM
To: r-help at stat.math.ethz.ch
Subject: [R] covariate data errors


Greetings,

	I would like to fit a multiple linear regression model in
which the residuals are expected to follow a multivariate normal
distribution, using weighted least squares.  I know that the data in
question have biases that would result in correlated residuals, and I
have a means for quantifying those biases as a covariance matrix. I
cannot, unfortunately, correct the data for these biases.

	It seems that this should be a straightforward task, but so
much of the literature is concerned with the probability model in
which the residuals are uncorrelated that I can't find a good
reference.  So in order of importance, please, can someone point me to
a definitive reference for least squares with correlated residuals,
and is there a standard R package to handle this case?

	Many thanks in advance,

	Anthony

______________________________________________
R-help at stat.math.ethz.ch mailing list
https://www.stat.math.ethz.ch/mailman/listinfo/r-help
#
Do you mean correlations in the *errors*?  The residuals are always 
correlated.  What does this have to do with your subject line -- it is 
errors in the dependent variable I think you mean?

If you have correlated errors, you should be using generalized least
squares not least squares or weighted least squares.  (That is covered in
all good books on regression: I don't know your level, but Seber's has a
comprehensive account.)  There are several R functions to fit GLS,
including gls(nlme) and lm.gls(MASS).
On Thu, 12 Jun 2003, Andy Jacobson wrote: