The robustness of Correspondence, Detrended Correspondence, and TWINSPAN Analysis

Abstract. Simulated vegetation data of known structure and varying complexity were analyzed with Correspondence Analysis (CA), Detrended Correspondence Analysis (DCA), and a polythetic divisive method of classification as incorporated in the computer program TWINSPAN. The CA and DCA axes were statistically compared with the original gradients. In the case of TWINSPAN, the results were visually evaluated. The analyses showed that both CA and DCA will recover the first vegetation gradient in an acceptable manner only if the variation in the vegetation due to second or higher order gradients is small. Second gradients were never recovered in a meaningful manner. The sample points along these axes were so far displaced from their original positions along the gradients that the axes could not be used to represent the gradients. In particular, when the two gradients have the same length, CA and DCA gave very poor results for both the first and the second axes. This is thought to be one of the main reasons why TWINSPAN does not perform well, especially after the first division.