Analysis of instability of ordinations

Tausch et al. (Journal of Vegetation Science, 6, 897-902; 1995) found out that some programs for Correspondence Analysis were dependent on the input order of species and sites. Peter Minchin and me have studied the performance of CANOCO. We explain the main results of our studies in a an article in Journal of Vegetation Science 8, 447-454; 1997 (you can have a look at its abstract). The main results are:

• CANOCO is unstable as well, but not as badly as the software studied by Tausch et al.
• Too lax convergence criteria are the main source of instability in orthogonal CA.
• An order dependent bug in rescaling is the main source of instability detrended CA.

What to do?

You will find most detailed information in our article in Journal of Vegetation Science 8, 447-454, 1997. In this web page you can find some software to help to solve the problems of instability. You may take the following steps in analysing the stability of your application:

• Those users who have the CANOCO source files and a FORTRAN compiler can make the changes themselves following the guidelines in our article (i.e., using stricter convergence criteria and correcting the rescaling bug).
• Users should check whether the current version of CANOCO is sufficiently stable for their data sets by getting the programs CEPSHUFL and SOLCOMP from this web server. For many data sets the results may be adequate. If the results are deemed unstable, users may take one of the following steps.
• For CA, use our program CASVD.
• For DCA, use our debugged, strict version of DECORANA.

Upgraded software

• Bruce McCune informed us that the new release of their PC-ORD software uses now "superstrict" convergence criteria, and DCA bug is corrected. More on their web pages. [Note added 29/8/97]
• Cajo ter Braak informed us that the current version 3.15 of CANOCO uses "strict" convergence criteria and the DCA bug is corrected. More detailed description can be found through official CANOCO v3.15 web pages. These pages include some further discussion on CANOCO instability, and challenge some of the theses. Here is our response to Cajo ter Braak on rescaling bug [Note added 5/11/1997, updated 9/3/98]
• Canoco for Windows 4.0 is now available. Check CfW homepage [Note added 28/9/98]

Go to the frontpage of Jari Oksanen