Does anybody know what are the attributes of a glm fit object that will allow the "predict call" to produce an se.fit? I am deleting most of the attributes as the size of the final object is 5Gb and I want to reduce it to under 20Mb, but that causes as error when I ask for an se.fit . mod.b$fitted.values <- 1:10 mod.b$prior.weights <- 1:10 mod.b$data <-mod.b$data[1:10,] mod.b$residuals <- 1:10 mod.b$linear.predictors <- 1:10 mod.b$qr$qr <- mod.b$qr$qr[1:10,] mod.b$effects <- mod.b$effects[1:100] mod.b$weights <- mod.b$weights[1:100] mod.b$model <- mod.b$model[1:10,] mod.b$y <- mod.b$y[1:10] p1 <- predict(mod.b,new=newdata,type="link",se.fit=T) Error in Qr$qr[p1, p1, drop = FALSE] : subscript out of bounds I believe the covariance matrix of the coefficients is all that should be needed and that is quite small. However, the covariance matrix is not an attribute of the model object. Thanks everybody. [[alternative HTML version deleted]]
> Does anybody know what are the attributes of a glm fit object that will allow > the "predict call" to produce an se.fit? > >.... > > I believe the covariance matrix of the coefficients is all that should be needed > and that is quite small. However, the covariance matrix is not an attribute of > the model object.Try saving the summary object? That includes the covariances (scaled and unscaled). See ?summary.glm. It might also leave you with a smaller object, though I'm not sure about that. Also note that on this list, when prediction from glm comes up, there's a strong likelihood that someone will point out that the covariances are not necessarily sufficient for reliable confidence intervals on prediction (and look! that just happened). You might want to hunt around for more authoritative comment on that if the intervals/standard errors are critical for something . S Ellison ******************************************************************* This email and any attachments are confidential. Any use...{{dropped:8}}