R help - Jan 2010 - Poisson regression: computation of linear combination of coefficients. Should one use the scaled var-cov matrix?

windows XP
R 2.10

When computing the variance of a linear combination of the coefficients from a
Poisson regression (i.e. glm with log link and offset) should one use the scaled
or unscaled covariance matrix? For a simple linear regression (i.e. lm), I
believe we use the unscaled matrix; for Poisson regression I believe we use the
scaled matrix.

Questions: (1) Am I correct about the use of the scaled matrix for the Poisson
regression?
               (2) What is the difference between the scaled and unscaled
matrices?

Thanks,
John



John David Sorkin M.D., Ph.D.
Chief, Biostatistics and Informatics
University of Maryland School of Medicine Division of Gerontology
Baltimore VA Medical Center
10 North Greene Street
GRECC (BT/18/GR)
Baltimore, MD 21201-1524
(Phone) 410-605-7119
(Fax) 410-605-7913 (Please call phone number above prior to faxing)

Confidentiality Statement:
This email message, including any attachments, is for th...{{dropped:6}}

John Sorkin wrote:
> windows XP
> R 2.10
> 
> When computing the variance of a linear combination of the
coefficients from a Poisson regression (i.e. glm with log link and
offset) should one use the scaled or unscaled covariance matrix? For a
simple linear regression (i.e. lm), I believe we use the unscaled
matrix; for Poisson regression I believe we use the scaled
matrix.
> 
> Questions: (1) Am I correct about the use of the scaled matrix for
> the
Poisson regression?
> (2) What is the difference between the scaled and unscaled matrices?
> Thanks,
> John
The scaled covariance matrix is the unscaled one multiplied by the 
dispersion, which is 1 for the Poisson and binomial families. I.e. they 
are one and the same in this case. For the Gaussian case, the dispersion 
is obtained from the the residual SS. In either case, the scaled version 
  is the one to use, although you might use it unscaled in Poisson.

One little mystery is why vcov.glm (which you might as well have used), 
does this:

{
     so <- summary.glm(object, corr = FALSE, ...)
     so$dispersion * so$cov.unscaled
}

rather than just return cov.scaled? (Historical reasons, I suspect).
> 
> 
> 
> John David Sorkin M.D., Ph.D.
> Chief, Biostatistics and Informatics
> University of Maryland School of Medicine Division of Gerontology
> Baltimore VA Medical Center
> 10 North Greene Street
> GRECC (BT/18/GR)
> Baltimore, MD 21201-1524
> (Phone) 410-605-7119
> (Fax) 410-605-7913 (Please call phone number above prior to faxing)
> 
> Confidentiality Statement:
> This email message, including any attachments, is for ...{{dropped:15}}

Peter,
I should have added that because I have over dispersion, I am ruining a
quasipoisson regression.
John




John David Sorkin M.D., Ph.D.
Chief, Biostatistics and Informatics
University of Maryland School of Medicine Division of Gerontology
Baltimore VA Medical Center
10 North Greene Street
GRECC (BT/18/GR)
Baltimore, MD 21201-1524
(Phone) 410-605-7119
(Fax) 410-605-7913 (Please call phone number above prior to faxing)>>>
Peter Dalgaard <p.dalgaard at biostat.ku.dk> 1/29/2010 6:27 PM
>>>
John Sorkin wrote:
> windows XP
> R 2.10
> 
> When computing the variance of a linear combination of the
coefficients from a Poisson regression (i.e. glm with log link and
offset) should one use the scaled or unscaled covariance matrix? For a
simple linear regression (i.e. lm), I believe we use the unscaled
matrix; for Poisson regression I believe we use the scaled
matrix.
> 
> Questions: (1) Am I correct about the use of the scaled matrix for
> the
Poisson regression?
> (2) What is the difference between the scaled and unscaled matrices?
> Thanks,
> John
The scaled covariance matrix is the unscaled one multiplied by the 
dispersion, which is 1 for the Poisson and binomial families. I.e. they

are one and the same in this case. For the Gaussian case, the
dispersion 
is obtained from the the residual SS. In either case, the scaled
version 
  is the one to use, although you might use it unscaled in Poisson.

One little mystery is why vcov.glm (which you might as well have used),

does this:

{
     so <- summary.glm(object, corr = FALSE, ...)
     so$dispersion * so$cov.unscaled
}

rather than just return cov.scaled? (Historical reasons, I suspect).
> 
> 
> 
> John David Sorkin M.D., Ph.D.
> Chief, Biostatistics and Informatics
> University of Maryland School of Medicine Division of Gerontology
> Baltimore VA Medical Center
> 10 North Greene Street
> GRECC (BT/18/GR)
> Baltimore, MD 21201-1524
> (Phone) 410-605-7119
> (Fax) 410-605-7913 (Please call phone number above prior to faxing)
> 
> Confidentiality Statement:
> This email message, including any attachments, is for\...{{dropped:26}}