Suggestion : Add functions for PCA and CA that just call rda() and cca() in the background

[cmartin]

Hi,

I'm teaching every year an undergrad course where we use vegan for many useful things (nmds, hclust,kmeans, etc.).

Since this is an undergrad course, we don't go into canonical ordinations, but I like my students to have a constant interface when doing multivariate analysis so we use vegan also for pca and ca, especially since vegan is THE multivariate library to know in ecology.

Every year though, I see the confusion in their eyes that they use a function named for another statistical method because it can do both.

Would it be possible to add both a pca and ca method that just call rda and cca transparently in the background?

Thanks you for your time and great package!

[jarioksa]

You mean these functions?

ca <- function(x, ...) cca(x, ...)
pca <- function(x, scale = FALSE, ...) rda(x, scale = scale, ...)

[gavinsimpson]

@jarioksa Yes; I face the same questions every time I have to teach vegan to people. I know it's trivial to write those functions myself, but that doesn't help the students.

I was going to propose this in a PR but @cmartin has beaten me to the proposing.

Making this change gets my full support and I'll even do the PR if you are OK with it Jari?

[gavinsimpson]

I would also add pco() to the proposed list as a wrapper around wcmdscale()

[cmartin]

Yes @jarioksa , that's exactly what I meant.

Thanks @gavinsimpson for putting this in action.

I can't believe how fast you guys reacted! I should have posted this suggestion YEARS ago LOL

[jarioksa]

OK Gav, just go on!. Anyway it's a bit more complicated than my suggestion above: you need to adjust the (printed) call as well. There may be other bits where you need changes for support functions. For pco you probably mean making eig=TRUE default?

[gavinsimpson]

@jarioksa Will do (and yes, I realize the actual functions will need some messing with the call - things are never that simple :-)

For pco(), yes, I think so. At least that's how I was calling wcmdscale() when I wrote my slide on it but I don't use it except when I have to teach, so I need to refresh my memory on it.

[jarioksa]

See issue #186 which suggests having eigenvalues etc for default wcmdscale. Then I declined, because wcmdscale was designed as a drop-in replacement of stats::cmdscale, but pco() could solve this anomaly.

[jarioksa]

Another option is to make pco as a wrapper to dbrda (which only has formula interface currently).

Please note that it would be possible to call directly ordConstrained set of functions:

ca <- function(x, ....) ordConstrained(x, method="cca", ...)
pca <- function(x, scale = FALSE, ...) ordConstrained(x, method="rda", arg = scale, ...)
pco <- function(d, ...) ordConstrained(d, method = "dbrda", ...)

This will set header info and names as well (ordHead), and if you want to omit those, there are only essential analytic steps: init* and ordResid. For instance:

pco <- function(d, ...)
{
    d <- initDBRDA(d)
    ordResid(d)
}

This will only return numeric result without janitorial stuff of dbrda. However, I think you may want to have Lingoes and Cailliez distortions in pco and that goes more easily with wcmdscale (or then you need to edit dbrda to accept non-formula interface).

[jarioksa]

@gavinsimpson I think it is best to implement pco() as a wrapper to dbrda so that all these functions (pca, ca, pco) have similar output, support functions and user experience. You can also have Cailliez & Lingoes distortions with dbrda (but not with direct call to ordConstrained(..., method = "dbrda", ...)). The prototype would be:

pco <- function(d, ...) dbrda(d ~ 1, ...)

[gavinsimpson]

These are now implemented in the GitHub (development version).

@jarioksa, should these be cherry-picked to 2.6-6.2 so that they will be more immediately avaiable?

[jarioksa]

@gavinsimpson yes indeed, I think that these should be cherry-picked -x (use this switch) to cran-2.6-6 branch. Now I think that we should release 2.6-6.2 in mid-August before the (northern hemisphere) university teaching season (which seems to be called 'semester' in the English world, while 'semester' in Scandinavia means holiday). The fix is a (relatively) save wrapper and intended for teaching purposes, and we should have it in teaching release.

Please note that cran-2.6-6 has not moved inst/NEWS.md to the main directory. I did not see a need for this because there is no github.io page for the 2.6-6 branch. Currently the NEWS live in parallel universes: master lives like there will be no new CRAN releases, and cran-2.6-6 like there will be. After release the NEWS need synchronizing, but now they are separate.

[gavinsimpson]

@jarioksa Done; I think a mid-August release would be good