Skip to content

The TWINSPAN program

6 messages · Diogo B. Provete, Gavin Simpson, Jari Oksanen +1 more

#
On Wed, 2011-04-13 at 09:25 -0300, Diogo B. Provete wrote:
I'm sorry, (I don't like TWINSPAN...) but to claim TWINSPAN is not used
because it has been superseded by the IndVal approach is totally
incorrect.

TWINSPAN and IndVal do **very** different things; the former produces a
cluster analysis that happens to churn out [a form of] indicator species
values, whilst the latter **only** computes [a form of] indicator values
- you have to supply the clustering.

G

  
    
#
On 13/04/11 15:34 PM, "Gavin Simpson" <gavin.simpson at ucl.ac.uk> wrote:

            
Howdy all,

Gavin is absolutely correct here (and I am not a TWINSPAN fan either).

Various clustering methods are the closest thing to Twinspan in base R.
However, they don't provide you species clustering which makes Twinspan
unique. Twinspan works on the original community matrix and produces a
simultaneous classification for plots and species. I don't use
classification but casually, and I don't know if there are such simultaneous
two-way classification problems in R. Indval and friends for quite a
different problem, like Gavin wrote (twice).

As far as I know, Twinspan is not available in R. Two persons have contacted
me and proposed to port Twinspan to R, and I have provided them the basic
files and promised to help them in the work, but I haven't heard anything of
the project after the initial contact.

I do think that Twinspan is a suboptimal choice for classification problems,
but I won't go into details. I urge you to study its behaviour yourself if
get your hands on Twinspan.

Cheers, Jari Oksanen
#
On Wed, 2011-04-13 at 15:01 +0200, Andres Mellado Diaz wrote:
???

TWINSPAN /provides/ indicator values, but it is not its raison d'etre.
It *was* designed to *cluster* vegetation data in the two-way manner
Jari mentions and provides the indicator values as one of extra outputs.
In the past, one would have to use TWINSPAN to get indicator values
because there weren't many (any?) other options for computing them, but
if you wanted indicator values then you had to accept the TWINSPAN
clustering too - there was no either/or.

IndVal changed that so you *could* compute good indicator values along
the same lines as TWINSPAN but without having to use it esoteric
clustering algorithm. Of course Dufrene and Pierre cite the TWINSPAN
paper a lot; they were producing a new tool that at the grossest level
did something (one part) that TWINSPAN did and therefore could be
compared against.

Your entire email is focussed on one aspect of TWINSPAN and the
similarities between it and IndVal - you aren't seeing the woods for the
trees. TWINSPAN and IndVal are different beasts.

To your argument I might offer the repost: "post hoc ergo propter
hoc" (in a bastardised way: TWINSPAN and IndVal give me indicator
values, therefore TWINSPAN and IndVal are the same. ;-)

G