DCA

[Tim Dickinson]

On Tue, 6 Aug 1996, Utteridge Timothy Michael Arthur wrote:

> Dear Taxacomers,
>
> Having read Parnell & Waldren's paper about Detrended correspondence
> analysis (Taxon 45(1): 71-84. 1996) I was wondering if anyone has any
> experience of using DCA with taxonomic data. Parnell & Waldren list only
> advantages, there must be a few disadvantages or problems with using DCA.
> Comments on 'friendliness' of computer programs would be appreciated.
>
> Tim Utteridge
> Dept. of Ecology & Biodiversity
> University of Hong Kong
>
> Tel: +852 2857 9912
> Fax: +852 2559 5984

the parnell & waldren (p&w) paper struck me as an unfair put-down of
principal coordinates analysis, given that resemblance functions
like gower's coefficient for mixed data exist, and that biplots of
otus and characters can be obtained for any ordination, either by
formula or by using a statistics or matrix algebra package to
calculate the correlations between 2 matrices (i.e. of data and of
scores).

all of the common ordination methods (pca, pcoa, cva, cca), and dca
as well, start from a resemblance matrix of some kind. contrary to
what p&w suggest on p.73, the choice of resemblance function (i.e.
similarities, distances, etc.) does not have to be (and should not
be) arbitrary; guidelines are provided in legendre & legendre's
numerical ecology book (tables 6.3 - 6.5), as well as in a paper by
legendre & gower (or vice-versa) in journal of classification, and
elsewhere.

these points account for three of the four supposed advantages of
dca over pca and pcoa listed on p.73 (i don't know much about nmds),
so that i can't see rushing out to buy a dca machine just because
one has multistate or mixed multistate and continuous data.

james lyons-weiler also responded to this posting, as follows (in
part):

>...So far as what the [ordination] results mean, Tausch et al. (1995)
>J. Veg. Sci 6:897-902 have shown that any software package that conducts
>ordination analyses like PCA, DCA, and classification (cluster) analyses
>such as UPGMA or TWINSPAN have a MAJOR problem: their results depend on
>data entry order (which species comes first in the matrix, or which
>character is character #1).  That means that the software may settle on
>a different answer depending on a completelty arbitrary characteristic of
>the data.  The problem is due to the fact that multidimensional analyses
>are computationally complex, so the programmers have to use various
>estimation steps for eigenvalues.  If they didn't, the programs would take
>much longer to give an answer.

well, i haven't seen the rausch et al. paper yet (i just sent off
for a reprint, and haven't yet looked in the library for j. veg.
sci.), but i did write an s-plus function to jumble the rows and
columns of a symmetric resemblance matrix and find its
eigenstructure each time. as far as i can tell it does what i said
it does, and there is no difference in either the eigenvalues or the
eigenvectors regardless of the sequence of rows and columns. s-plus
is certainly a well-crafted piece of software, so maybe the problem
identified by rausch et al. only crops up in less carefully
programmed eigenanalysis routines. i can't see how the sequence of
rows or columns in a data matrix could affect calculation of the
resemblances themselves. if this really is a problem perhaps
someone can enlighten me.

as far as the "friendliness" of computer pgms goes, that varies a
lot. both canoco and s-plus are very powerful pgms in their (very)
different ways; neither one, in my opinion, is for the fainthearted.
ntsys-pc and syn-tax each do a lot (comparable to what a version of
canoco that's 3-5 y old does, but much less than what s-plus does)
and _are_ pretty user-friendly:  they are driven by well designed,
consistent menus. you still have to read the manual, or be shown
how to do particular things, but their learning curves are nothing
like those of the other pgms mentioned.

cheers, ---tad.