I am following up on an old post. Please, comment: it appears that
predict(glm.model,type="response",se.fit=T) will do all the conversions and give se on the scale of the response. This only takes into account the error in parameter estimation. what a "prediction" interval is meant to be usually means it has to capture the error due to both parameter estimation and sampling variation ie it encompasses the actual realizations. for a given parameter there is sampling variation and that is not included in the output of predict. the discreteness of models makes it quite difficult to estimate a percentile interval, though. for binary outcomes, I think it does not make sense. for Poisson and binomial (grouped binary) I think it is possible to get approximations at least and this is what the original poster needed I think. so, let's say we have plow and pup for an observation from predict. if size=100 for that obs. predlow=qbinom(.025,100,plow) predup=qbinom(.975,100,pup) will give the prediction bounds. this I think partly ignores possible overdispersion. Please, suggest a better way taking overdispersion into account (in the qbinom part). Thanks everybody. Stephen B. -- View this message in context: http://r.789695.n4.nabble.com/Prediction-confidence-intervals-for-a-Poisson-GLM-tp841577p2533070.html Sent from the R help mailing list archive at Nabble.com. ______________________________________________ R-help@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide http://www.R-project.org/posting-guide.html and provide commented, minimal, self-contained, reproducible code.
