Dear Jacques,
Did you ever figure that error out? I have recently
hit it too. In my case it comes from glm() when
I replace family = binomial with family = cauchit
for some home brewed cauchy binary regression.
Things are ok with binomial but not with Cauchy.
The data set has some missing values. I tried replacing
the x by multiple x's and by a matrix. I suspect that
the intrinsic binomial regression is doing something
smarter with the missing data.
The code worked with an older R (1.7.1 or 1.8.1 I think)
on other data sets.
-Art
here's my cauchy regression hack:
cauchit = quasibinomial()
cauchit$family = "binomial"
cauchit$link = "cauchit"
cauchit$linkfun = function(mu){ qcauchy(mu) }
cauchit$linkinv = function(eta){ pcauchy(eta) }
cauchit$mu.eta = function(eta){ dcauchy(eta) }
then I use glm( .... family = cauchit )