CAP - BiodiversityR versus Primer

[Salvador Herrando-Perez]

Dear colleagues.

As most of you might know, CAP (as originally described by Anderson & Willis 2003*) is a discriminant analysis (or Canonical Correlation Analyses) on PCoA space. I am referring to the discriminant option here**.

I have found that the CAP menu in the Primer software (by Marti Anderson) and the function CAPdiscrim in BiodiversityR (by Roeland Kindt) produce different results in both 'best m' and classification rates, though they result in the same PCoA space (identical multivariate representations, different scales of PCoA axes) and identical Pillar trace statistic. Using the dune dataset from BiodiversityR, m = 9 in BiodiversityR versus m = 6 in Primer, while at any given m both platforms differ in classification rates - the mismatch persists when I have done the same with a range of my own datasets.

I wonder whether you might have encountered this problem and figure out where the solution lies. I have discussed the problem with Roeland but we could not find out why this is happening.

Many thanks in advance for your thoughts.

Salva

* Anderson, M.J., and Willis, T.J. (2003). Canonical analysis of principal coordinates: A useful method of constrained ordination for ecology. Ecology 84, 511-525.

** My query should be relevant to CAP users fully relying on R for multivariate analyses, since (to my knowledge) BiodiversityR seems to be the only package in R that does CAP as described by Anderson & Willis (2003), i.e., estimating m through a leave-one-out procedure - unlike capscale in Vegan.

[Pier Luigi Buttigieg]

Hi Salvador,

Unfortunately, I don't use Primer so can't address the main question in any detail; my guess is that they employ different thresholding criteria to decide group membership and optimal number of axes. Looking at the code (if it's open), documentation (if it exists in that detail), or contacting the Primer developers would be your best bet to get an authoritative answer. 

Have you tried other classification techniques such as those from the machine learning community? Many report success with random forest approaches or neural networks. Many of these are attractive in that they avoid the use of a dissimilarity matrix (and the interdependence that creates) and can often cope with unruly data.

PS: If you'd like to create a GUSTA ME page (micro-credited with your ORCID of course) on the CAP approach, please let me know!

[Salvador Herrando-Perez]

Hi Pier Luigi.
I have just received the open-source code from Roeland Kindt for CAPdiscrim and I will be contacting Marti Anderson about her code. If you thought this was relevant to the Forum, I could post Marti's reply here.
Cheers.
Salva
* I will be delighted to write a GUSTA ME page for CAP - could do it next month: would that be ok?

[Pier Luigi Buttigieg]

Hi Salvador,

I have just received the open-source code from Roeland Kindt for CAPdiscrim and I will be contacting Marti Anderson about her code. If you thought this was relevant to the Forum, I could post Marti's reply here.

 Absolutely! It would be great to hear what you find.

 

* I will be delighted to write a GUSTA ME page for CAP - could do it next month: would that be ok?

Great! Let me know when you're ready to move forward and I'll brief you on the (very easy) procedure.

[Salvador Herrando-Perez]

Will do