Thanks for making this package available!
The standard error estimates from "cpglm" look a bit odd. I think that they're
using the estimate of dispersion (phi) from glm (which is a method of moments
estimator, I think?) rather than the maximum likelihood estimate which cpglm
reports. I might be confused on some details of GLM estimation, though, sorry!
This follows the behavior of glm on other models, but since cpglm reports a
maximum likelihood estimate of phi, I expected the standard errors would use
that estimate as well. An example is attached showing the difference in
standard errors.
I think a fix might be as simple as changing the code in cpglm.profile that
extracts the variance-covariance matrix from:
summary.glm(fit)$cov.scaled
to: summary.glm(fit, dispersion=phi.max)$cov.scaled
Original issue reported on code.google.com by danieljs...@gmail.com on 25 Mar 2014 at 7:28
Original issue reported on code.google.com by
danieljs...@gmail.comon 25 Mar 2014 at 7:28Attachments: