Hi, Jovana. Geiger’s parameters aren’t quite what you described.
SE=NA: estimate what measurement error is, not assume it’s zero. This is then a free parameter, penalized appropriately in the AIC calculations. SE=0 (or run geiger with defaults): force the assumption that measurement error is zero SE=42, SE=c(42, 10, 20, 30): assume measurement error is 42 for all taxa, or assume it’s 42 for taxon 1, 10 for taxon 2, 20 for taxon 3, and 30 for taxon 4. Jeremy Beaulieu and I have a paper in review (and a preprint) on measurement error that might be of interest (and links out to other literature on it): https://doi.org/10.1101/2024.08.19.608647. TL;DR: we think defaults in most programs should be to estimate measurement error (which we have relabeled as “tip fog” for clarity, since what the parameter fits is broader than just imprecision in how one measures). Best, Brian _________________________________________ Brian O’Meara He/Him Professor, Dept. of Ecology & Evolutionary Biology University of Tennessee, Knoxville From: R-sig-phylo <r-sig-phylo-boun...@r-project.org> on behalf of Jovana Malikovic <jovana.maliko...@anatomy.uzh.ch> Date: Thursday, October 3, 2024 at 5:00 AM To: r-sig-phylo@r-project.org <r-sig-phylo@r-project.org> Subject: [R-sig-phylo] Including measurement error in the estimation_ fitContinuous_geiger package [You don't often get email from jovana.maliko...@anatomy.uzh.ch. Learn why this is important at https://aka.ms/LearnAboutSenderIdentification ] Dear all, I would like to clear up some confusion regarding the SE argument within the geiger package. Namely, there can be 3 scenarios: Scenario 1: SE = NA (no measurement error modeled/ignore SE) By setting SE = NA, we tell the function to ignore any potential measurement error in the trait data. The model assumes the trait values are measured perfectly, meaning all observed variation in trait values is due to evolutionary processes. Scenario 2: no SE argument (implicitly assuming SE = 0) When we omit the SE argument, the function assumes that there is no uncertainty explicitly stated for the traits. However, the model still treats the trait data as exact but does not impose the same assumption of perfect measurement as SE = NA. Essentially, it acts as if the measurement error is not considered, which may implicitly include the idea that some noise might still exist, although it isn't modeled. Scenario 3: SE = real calculation When we provide a vector of standard error, the function adjusts the likelihood to account for measurement uncertainty in the trait data. The trait values are considered as estimates of true trait values with some noise due to measurement error. In practice, not providing the SE argument can result in model fits that are somewhat similar to those obtained by including estimated SEs, particularly because both scenarios lead to the model being less simplistic than when using SE = NA. My question is: what is the difference between Scenario 1 and Scenario 2? The model does exactly the same in Scenario 1 and 2, yet the outcomes are different, just because in scenario 1 we are making a clear decision to ignore errors and in the scenario 2 the model makes that assumption on its own because no errors were mentioned? My understanding from the literature (Silvestri 10.1111/2041-210X.12337) is that the whole reason for considering the error is because more complex models may by mistake be preferred if the error is not accounted for. Can someone explain, please. Best wishes, Jovana [[alternative HTML version deleted]] _______________________________________________ R-sig-phylo mailing list - R-sig-phylo@r-project.org https://stat.ethz.ch/mailman/listinfo/r-sig-phylo Searchable archive at http://www.mail-archive.com/r-sig-phylo@r-project.org/ [[alternative HTML version deleted]] _______________________________________________ R-sig-phylo mailing list - R-sig-phylo@r-project.org https://stat.ethz.ch/mailman/listinfo/r-sig-phylo Searchable archive at http://www.mail-archive.com/r-sig-phylo@r-project.org/
