For Box-Cox transformations there are R tools that handle this automatically in the "tidyverts" packages. Check out the subsection "Bias adjustments" in section 5.6 "Forecasting Using Transformations" in the book: "Forecasting Principles and Practice (3rd ed)" which is freely available online at
https://otexts.com/fpp3/ HTH, Eric On Thu, Mar 19, 2026 at 6:07 PM Ben Bolker <[email protected]> wrote: > > https://cran.r-project.org/web/packages/emmeans/vignettes/transformations.html#bias-adj > is probably the easiest place to start. That machinery is assuming > that the transformation is stated *explicitly* in the model > formulation; the example used in the vignette is > > pigs.lm <- lm(log(conc) ~ source + factor(percent), data = pigs) > > I think it wouldn't work if the transformation was done upstream (i.e. > if your response variable was `log_conc`), and possibly not if you had > an unusual transformation. > > Looking at the function in emdbook, I don't think it's directly > useful for what you want. car::deltaMethod looks more useful. However, > it's designed for specific nonlinear functions of *parameters*, I > don't know if it can easily do bias correction on predictions. > > On Thu, Mar 19, 2026 at 11:42 AM Christofer Bogaso > <[email protected]> wrote: > > > > Thanks Ben. > > > > Could you please help pointing out names of specific functions on Bias > > correction? I searched with like ls('package:emdbook') etc. however > > failed to identify relevant functions. > > > > On Thu, Mar 19, 2026 at 6:55 PM Ben Bolker <[email protected]> wrote: > > > > > > There are functions in the emdbook, metafor, and car packages that > > > do some version of the delta method (although people use "delta > > > method" to refer both to adjusting E[f(y)] using a second-order > > > correction [since the first-order term disappears] and to adjusting > > > V[f(y)] using a first-order correction ...) > > > > > > emmeans also has such capabilities, search the vignettes for "bias > correction" > > > > > > cheers > > > Ben Bolker > > > > > > On Thu, Mar 19, 2026 at 8:57 AM Christofer Bogaso > > > <[email protected]> wrote: > > > > > > > > Hi, > > > > > > > > In many case, we need to transform the dependent variable before > > > > fitting a regression equation, to make it "well-behaved" like close > to > > > > normal curve etc. > > > > > > > > like, > > > > > > > > f(y) = alpha + beta1 X1 + beta2 X2 + ... + epsilon > > > > > > > > Now for prediction, R will typically calculate E[f(y)] based on the > > > > fitted coefficients. However, in real scenario, we actually need to > > > > find E[y]. > > > > > > > > Typically, we perform reverse transformation like on fitted E[f(y)] > directly. > > > > > > > > However, I believe that in this process, we also need to make some > > > > additional correction for non-linearity in the f() to correctly > > > > calculate E[y]. Onr possible way to do it, may be using Taylors > > > > approximation. > > > > > > > > My question is there any R function that would directly do that based > > > > on the shape of f()? > > > > > > > > Thanks for your time. > > > > > > > > ______________________________________________ > > > > [email protected] mailing list -- To UNSUBSCRIBE and more, see > > > > https://stat.ethz.ch/mailman/listinfo/r-help > > > > PLEASE do read the posting guide > https://www.R-project.org/posting-guide.html > > > > and provide commented, minimal, self-contained, reproducible code. > > ______________________________________________ > [email protected] mailing list -- To UNSUBSCRIBE and more, see > https://stat.ethz.ch/mailman/listinfo/r-help > PLEASE do read the posting guide > https://www.R-project.org/posting-guide.html > and provide commented, minimal, self-contained, reproducible code. > [[alternative HTML version deleted]] ______________________________________________ [email protected] mailing list -- To UNSUBSCRIBE and more, see https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide https://www.R-project.org/posting-guide.html and provide commented, minimal, self-contained, reproducible code.
