A Comparison of Nonmetric Multidimensional Scaling, Principal Components and Reciprocal Averaging for the Ordination of Simulated Coenoclines, and Coenoplanes

Abstract

The use of nonmetric multidimensional scalings as an ordination method has been studied by the use of simulated coenoclines and coenoplanes. It was found that the method always produced better ordinations than principal components analysis and in most cases better than reciprocal averaging, providing the multidimensional scaling was calculated using a space of the same dimensions as the simulated data. For real data of which the dimensionality is unknown, the minimum spanning tree can provide a useful means of estimating this dimensionality. Nonmetric multidemensional scaling was also less susceptible to distortion of the ordination by single gradients of high beta diversity and two—gradient situations when each gradient is of a different beta diversity. A number of different similarity measures were evaluated for use in conjunction with nonmetric scaling and the cos theta coefficient was found to generally give good results. The use of a log transformation of the species values has a marginal effect while the use of a `Manhattan' distance metric in ordination space produced inferior results.