It seems that the standardized deviance residulas, that one gets on
plots of a glm.object x with plot(x) are calculated as
r <- residuals(x)
s <- sqrt(deviance(x)/df.residual(x))
w <- weights(x)
hii <- lm.influence(x)$hat
r.w <- if (is.null(w)) r else (sqrt(w) * r)
rs <- r.w/(s * sqrt(1 - hii))
This implies that, for example, for binomial B(ni,pi) data the devaince
residials (which are just r) are weighted not only with sqrt(1-hii), but
also with 1/sqrt(ni) and s, leading to absurd values. As a result all
leverage/outlier diagnostics is absolutly wrong.
Am I right and this should be reported as a bug?
Many thanks,
Tatyana
--
Tatyana Krivobokova
Bielefeld University
Department of Economics and Business Administration
PF 100131, D-33501 Bielefeld
+49 (0)521 106 4876
http://www.wiwi.uni-bielefeld.de/~krivobokova
