Timeline for (When) Does GLMM provide better predictions than logistic regression?
Current License: CC BY-SA 4.0
10 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| May 1 at 7:13 | comment | added | Roger V. | @Dave the closest example I van think of is detection of trisomy: e.g., in France the first obligatory ultrasound exam is at 11 weeks (just below the legal abortion threshold) and one is provided with a probability that the foetus is trisomic... but there is a hard set threshold for this probability, above which it mandates actual genetic test, by taking a sample of amniotic liquid (an invasive and risky procedure.) | |
| May 1 at 7:07 | comment | added | Roger V. | @Dave the doctor follows protocols, developed via statistical studies - prescribing him/her binary decisions. I would need a more detailed example to judge, whether this talk of probability has more than academic value (no offense intended - I spent decades in academia , and I know that the gap between academia and real applications is large...) | |
| Apr 30 at 16:50 | comment | added | Dave | @RogerV. Sure, the doctor either prescribes a treatment or not, but your regression does not. | |
| Apr 30 at 12:52 | answer | added | seanv507 | timeline score: 1 | |
| Apr 30 at 12:49 | comment | added | Roger V. | @seanv507 This is a point of view valid for a researcher, even a medical researcher... but not in real medical setting: I am not sure how a medical doctor could prescribe a treatment with probability 12% or a cancer test with probability 70% - they either do it or not. What is more important for me: this binary logic underlies the FDA and other equivalent medical authorities requirements for medications, tests, etc. | |
| Apr 30 at 12:44 | comment | added | Stephan Kolassa | .... and even if you do need to threshold your predictions: you are starting with probabilistic ones, so it makes sense to separate this modeling step and assess it on its own merits, and only subsequently consider which threshold(s) are optimal. 0.5 is usually not the best answer. | |
| Apr 30 at 12:37 | comment | added | seanv507 | @RogerV. that's no excuse. see fharrell.com/post/classification (which was in the answers linked in the previous comment): "A frequent argument from data users, e.g., physicians, is that ultimately they need to make a binary decision, so binary classification is needed. This is simply not true. First of all, it is often the case that the best decision is “no decision; get more data” when the probability of disease is in the middle. In many other cases, the decision is revocable, e.g., the physician starts the patient on a drug at a lower dose ..." | |
| Apr 30 at 11:52 | comment | added | Roger V. | @StephanKolassa I am working in a medical setting, where one has to make decisions healthy/sick. | |
| Apr 30 at 10:23 | comment | added | Stephan Kolassa | Since you are already working on probabilistic predictions, I would stick with those and not assess thresholded "hard" predictions. | |
| Apr 30 at 9:56 | history | asked | Roger V. | CC BY-SA 4.0 |