weights in glm (PR#8720)

First, R-bugs is not for asking questions, only for reporting things you 
are *certain* are bugs: see the R FAQ.

Why are you using weights to omit cases?  If you had used subset, this 
would have worked.  The problem is the use of zero weights, which are not 
intended to be used in this way.  We can fix this up, but it is not the 
correct way to use glm.

On Tue, 28 Mar 2006, robert.pusz at wp.pl wrote:

Full_Name: Robert Pusz
Version: 2.2.1
OS: Windows
Submission from: (NULL) (157.25.9.126)


Hello,
In my opinion something is wrong with 'weights' option in glm.
My code is following:
###begin of the code####
cl <- c(5012, 106, 3410, 5655, 1092, 1513, 557, 1351, 3133, 2063, 3257,
4179, 5582, 5900, 8473, 4932, 3463, 5596, 2262, 0, 2638, 1111,
4881, 4211, 6271, 5257, 6926, 6165, 0, 0, 898, 5270, 2268, 5500,
6333, 1233, 1368, 0, 0, 0, 1734, 3116, 2594, 2159, 3786, 2917,
0, 0, 0, 0, 2642, 1817, 3479, 2658, 225, 0, 0, 0, 0, 0, 1828,
0, 649, 984, 0, 0, 0, 0, 0, 0, 599, 673, 603, 0, 0, 0, 0, 0,
0, 0, 54, 535, 0, 0, 0, 0, 0, 0, 0, 0, 172, 0, 0, 0, 0, 0, 0,
0, 0, 0)

w <- c(1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0,
1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1,
1, 1, 1, 1, 1, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 1, 0,
1, 1, 0, 0, 0, 0, 0, 0, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0,
0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0)

row <- c(1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10,
1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10,
1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10,
1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10,
1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10)

col <- c(1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 5,
5, 5, 5, 5, 5, 5, 5, 5, 5, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 7, 7,
7, 7, 7, 7, 7, 7, 7, 7, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 9, 9, 9,
9, 9, 9, 9, 9, 9, 9, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10)
Row <- ordered(as.factor(row))
Col <- ordered(as.factor(col))
fit <- glm(cl ~ Row + Col, family = quasipoisson, weights = w)
###end of the code####

When I have written summary(fit) I have got:
           Estimate Std. Error t value Pr(>|t|)
(Intercept)  7.64459         NA      NA       NA
Row2        -0.05467         NA      NA       NA
Row3         0.24541         NA      NA       NA
Row4         0.42035         NA      NA       NA
Row5         0.43961         NA      NA       NA
Row6         0.04532         NA      NA       NA
Row7        -0.04881         NA      NA       NA
Row8         0.25370         NA      NA       NA
Row9        -0.14976         NA      NA       NA
Row10       -0.01267         NA      NA       NA
Col2         0.69283         NA      NA       NA
Col3         0.62603         NA      NA       NA
Col4         0.27695         NA      NA       NA
Col5         0.06056         NA      NA       NA
Col6        -0.19582         NA      NA       NA
Col7        -0.83044         NA      NA       NA
Col8        -1.27914         NA      NA       NA
Col9        -1.93235         NA      NA       NA
Col10       -2.49709         NA      NA       NA

When I omited 'weights=w' above table was filled in with numbers, but the
results were wrong (because of taking zeros in regression).

Could you tell me what's wrong?
Kind regards,
Robert

-- 
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

I'm not even sure we want to fix this up. I recall some nasty issues
with DF that have no proper solution that way - an observation with a
tiny weight represents an observation with a large variance and
contributes 1DF to the residual, with weight zero it is not supposed
to contribute at all, so there's a discontinuity for weights
approaching zero.