-
Notifications
You must be signed in to change notification settings - Fork 96
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Misleading summary in capscale (db-RDA)? #636
Comments
Please do not call I really do not know what is right and wrong in handling negative eigenvalues in I'll make the |
First a WARNING: I made a wrong assumption about weights in Then to I have now gone through handling the imaginary components in various vegan functions. We seem to have three different ways of handling those, but I am not sure which of these are misleading (or at least two may be misleading, but perhaps nobody wants to have the third one). We have two related functions: Here a comparison of printed output of unconstrained analysis of Dune meadow data with
The total inertia – or the sum of all eigenvalues – is estimated directly from the input dissimilarities prior to ordination, and it is the same in both analyses. This inertia is the same as the total variation in the data. In
The first default It is clear that we have been inconsistent, and this inconsistency must be cleared. We have used three different strategies:
Something must be done, but I'm not sure what. Option one is to follow the majority of methods that compare the proportions to real components and ignores imaginary. The only change needed is to make the brief result look like:
Option 2 is to change other functions to compare proportions to the original total inertia, which means that people must learn to live with >100% proportions. Obviously this concerns In both cases those functions that refuse to handle negative eigenvalues could be made consistent with this decision. Finally, I'm still more attracted to deprecating Opinions are appreciated. |
Thanks a lot for taking care of this topic so quickly and for the detailed analysis and explanation! I am not expert enough to have a clear opinion, but it seems to me that Option 2 is very unintuitive and explained proportions seem overestimated and not meaningful. I have remade my analysis with |
Earlier we used the observed total from input dissimilarities that could have negative eigenvalues. Analysis discards these negative axes and now we assess proportions with the Euclidean mapping discarding negative eigenvalues and hence higher total.
I think I have now changed most things to be consistent following option 1: ignore imaginary components and only base proportions on real components:
The corresponding part in
$r.squared
[1] 0.2031098
$adj.r.squared
[1] 0.133984 Many other functions changed on the way. See commit history if you want to see all changes. Some of these changes do not directly relate to these Some bigger associated changes are:
|
Dear devs,
I recently calculated db-RDA with capscale and realised one thing in the model outputs:
The reported results differ when calling the model object directly with
model
and when callingsummary(model)
. When calling the model object negative Eigenvalues are reported as seperate item of sum of negative eigenvalues (Imaginary; as stated in the documentary). When callingsummary(model)
the item of sum of negative eigenvalues is entirely missing and all proportions are calculated only based on the axes with non-negative Eigenvalues. This may be very misleading, especially when one is only usingsummary()
to call the full results?Thanks for all and all the best,
Friedemann
Here an example output:
The text was updated successfully, but these errors were encountered: