Non-metric Multidimensional Scaling (NMDS) is an ordination method that has been shown to be robust and effective for community analysis. NMDS is based on non-parametric model, which assumes only that the community dissimilarity between sampling units increases monotonically with ecological distance (degree of difference in ecological factors). With increases in computing power and the availability of software, NMDS has become widely-used by ecologists. Nevertheless, there are several common misunderstandings about the method and some recent papers have suggested that it produces misleading results and should not be used. I present insights into NMDS gained from 40 years of experience in programming and using the method to analyze community data. Using simulated community data with known structure, I address myths and misconceptions about NMDS and demonstrate its effectiveness in summarizing community data and revealing underlying ecological gradients. In the process, I clarify some technical aspects that are important for its effective use, and describe some innovations that make the technique even more useful.
Results/Conclusions
Though NMDS uses only the ranks of the dissimilarities, I show that NMDS ordination space is Euclidean, so it is valid to interpret directions and distances in the ordination space. There are several ways to compute ordination scores for species and the ordination can be scaled so that distances are measures of beta diversity. Bray-Curtis and related indices are the best dissimilarity measures to use with NMDS. An important option is how to handle tied dissimilarities. I show why the primary or weak approach should be preferred, otherwise arch distortion occurs when high beta diversity is high. I present evidence that shortest path pre-processing of dissimilarities is neither necessary nor beneficial to NMDS performance. I demonstrate that rules of thumb about stress values are not useful, since stress is influenced by the noise level in the data: an ordination with high stress can be ecologically informative if the data are noisy. Assigning proportions of "variance explained" to NMDS axes is uninformative because the method does not seek to explain variance. I defend NMDS from recent criticisms, which are based on poor option choices and the use of extreme examples, which are unlike the kinds of data usually collected by ecologists.