[Rd] Fitted values with small weights in lm.wfit (PR#11979)

[Prof Brian Ripley]

There is nothing to reproduce here.

Small weights per se are not necessarily a problem, but a very large range 
in weights might be, e.g. when computing weighted residuals.  We need a 
repoducible example for this 'bug' 'report' to be of any use (and we asked 
for one in several places, including the R FAQ).

Note that 'predict' does not give residuals, nor does it use lm.wfit ....

E.g.

set.seed(1)
x <- 1:100
y <- rnorm(100)
w <- rep(1e-100, 100)
fit <- lm(y ~ x, weights=w)
> range(predict(fit) - fitted(fit))
[1] -1.804112e-16  7.077672e-16


On Thu, 7 Aug 2008, ablocker at gmail.com wrote:

> Full_Name: Alexander Blocker
> Version: 2.7.1
> OS: Ubuntu 8.04 / Windows XP
> Submission from: (NULL) (76.119.235.225)
>
>
> When running lm(modeleq, weights=wt, data=dataset) with small weights (<1e-10),
> I have encountered an odd phenomenon with fitted values. Due to numerical
> precision issues, the fitted values and residuals returned by lm.wfit (from its
> .Fortran call to dqrls) can differ greatly from those returned by running
> predict on the resulting lm object. This is completely attributable to the
> numerical precision passed to the given function, but I wonder if a warning
> message for weights below as certain threshold may be in order.

-- 
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