Hi everyone,
let me clarify this with an example: Lets say, I have an animal model with two read-out parameters, like for example neutrophil and monocyte infiltration. If mouse A is identified as an outlier for the monocyte infiltration (by Grubbs-testing), but not for the neutrophil number, do I have to exclude the value of this mouse A for the neutrophil-readout as well?
In the first place, it sounds logical to do it that way. However... the more parameters I would measure (or include by choice), the more I can use this test to basically do cherry-picking. Or the more I measure, the less animals are left.
I hope it is understandable what I want to ask ;) Thank you very much for your input!
Cheers,
Peter
https://www.rdocumentation.org/packages/outliers/versions/0.14/topics/grubbs.test
There's no way to say this politely. Once you're disqualifying outliers from your data, after you know what the data is, then you are by definition cherry-picking. You can do more of it, or less of it. How much is OK? There's no firm cut-off. The more you do it, the more likely your result ends up false, your paper unpublishable, and your academic integrity in peril. It is best not to go there, period.
Thank you very much for your replies.
In my animal-model I have about 20 readouts. In the meantime, since I started this thread, I analyzed it both ways - completely 'consequent use' of Grubbs or no use at all. In the end I had to use the 'no-Grubbs' solution, since there are simply no animals left to analyze in the 'consequent use'-way. Meaning even if you try to use Grubbs completely correctly, it does not work.
Together with Johns statement, my summary is: you shouldn't use the Grubbs-test for a) feasibility reasons when done correctly and b) for integrity reasons of your research.
- Badges
- Science topic
- Similar topics
- Medicine
- Disease
- Disease Models
- Animal Models