Dear S+ users,
I have Poisson data collected in a two-factor experiment with
repetitions. The model E(Yijk) = Mu * Ai * Bj * (A:B)ij is of
application. The levels of the factors A and B have been chosen randomly
for this experiment and I am interested in the CV (coefficient of
variation) of the mean when one of the factors is kept at a fixed level.
Using "varcomp(log(y) ~ a*b, ...)" the variance components associated to
the factors and their interaction on a log scale can be computed. These
components can be interpreted in terms of CV² in the multiplicative
model (linear approximation). However, the "varcomp" approach is only
possible when the ranges of variation of the factors are limited and the
mean result not too near to 0, implying approximate constancy of the
variance of the transformed variables.
The same kind of data can also be analysed with much less restrictive
conditions using glm(y ~ a*b, family="poisson", ...). This produces
coefficients for the linear predictor log(E(Yij))=log(Mu) + log(Ai) +
log(Bj) + log((A:B)ij) which are nearly equal to the ones produced by
the "lm" approach, when restricted conditions apply. How can I estimate
(even roughly) the variances of the linear predictor terms ( {log(Ai)},
{log(Bj)} and {log((A:B)ij)} ) in the absence of a "varcomp" equivalent?
Thank you for your attention.
------------------------------------------------------------------------
--------
Pierre Delfosse Address: CENTEXBEL
pierre.delfosse@centexbel.be avenue du parc, 38
Telephone: 32-87-322443 B-4650 Chaineux BELGIUM
Fax: 32-87-340518
-----------------------------------------------------------------------
This message was distributed by s-news@wubios.wustl.edu. To unsubscribe
send e-mail to s-news-request@wubios.wustl.edu with the BODY of the
message: unsubscribe s-news
|