[R] Confidence intervals for gls models?

[Prof Brian Ripley]

On Sun, 18 May 2003, Spencer Graves wrote:

> What about the obvious:
> 
> tstDf <- data.frame(x=1:9, y=rnorm(9), w=1:9)
> fit <- lm(y~x, tstDf, weights=w)
> pred <- predict(fit, se.fit=T)
> pred$fit + outer(pred$se.fit, c(-2, 2))
> 
> "predict.lm" might need weights for interval="prediction" with newdata, 
> but not with interval="confidence" ... or am I missing something?

That's weighted least squares, not generalized least squares.

predict.gls does not have an `se.fit' argument.

Howver, lm.gls in package MASS will do the trick at the existing data 
points.  (To predict at newdata you would need to have a model for the 
covariance matrix, and once you have that you are doing time series or 
kriging or ... and there are many other possibilities.)

> 
> hth.  spencer graves
> 
> Brown, David wrote:
> > Is there an easy way to compute confidence intervals (or prediction
> > intervals) for gls models?
> > 
> > E.g. for standard linear models, with the predict.lm function, we can set
> > interval="confidence" , level = 0.95 and type="response".

-- 
Brian D. Ripley,                  ripley at stats.ox.ac.uk
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