Problem with gls

[Wim Kimmerer <kimmerer@sfsu.edu>]

Hello S-Plussers.


v. 6.2, Windows XP
I am trying to use gls to compare the output of a hydrodynamic model with
data for tidal height and flow.  The main issue for this particular
comparison is how close the slope of a regression is to 1.
  I have long time series (thousands of points) but of course lots
of autocorrelation.  I tried gls as follows:

z <- lm(data ~ model.lag,
data="">
va <- acf(z$resid, plot=F)$acf[2]
zg <- gls (data ~ model.lag, data="" correlation=corAR1(va))
------------------
This works OK, but the results are (snipped):
Linear regression:
              
Value Std. Error  t value Pr(>|t|) 
(Intercept) 115.4604  14.1993    
8.1314   0.0000
  model.lag   0.9823   0.0022  
438.0841   0.0000

Residual standard error: 776.5 on 2991 degrees of freedom
----------
gls:
Correlation Structure: AR(1)
 Formula:  ~ 1 
 Parameter estimate(s):
       Phi 
 0.9368339

Coefficients:
              
Value Std.Error  t-value p-value 
(Intercept) 96.05244  91.10014   1.0544  0.2918
  model.lag  0.91049   0.00573 158.7623 
<.0001

Residual standard error: 904.4512 

-----------
The slope is much lower with gls, and the residuals are larger.  Why
should this be so?  All I am trying to do is to account for the
autocorrelation in the original regression z, in the most robust
manner.  The regression coefficient in z is not in question but the
standard error is - so why do lm and gls not estimate the same
thing?

I have also tried gls in another situation where I had both
autocorrelation and increasing error variance with the mean.  In a
couple of cases gls converged but gave a slope that was clearly miles
away from the actual relationship.

I would appreciate any help from anybody.... the manuals etc. are not
much help at all.



Wim

======================
Dr. Wim Kimmerer
Research Professor of Biology
Romberg Tiburon Center
San Francisco State University
3152 Paradise Drive
Tiburon CA 94920
Ph. (415) 435-7143
Fax (415) 435-7120
http://online.sfsu.edu/~kimmerer/