Research paperA comparison of novel and traditional numerical methods for the analysis of modern pollen assemblages from major vegetation-landform types

- Comparison of several novel and traditional numerical methods
- Pollen distinguishes vegetation-landform types well independently of method used.
- The novel methods are not superior to the traditional numerical methods.
- The novel methods are more robust and rigorous statistically.
- There now exists an expanding tool-box for numerical analysis of pollen data.

In recent decades there has been a major development of numerical and statistical methods with considerable potential for the handling, analysis, and summarisation of complex multivariate pollen data. We compare 'traditional' ordination and hierarchical clustering methods with recently-developed methods such as principal curves, non-hierarchical agglomerative clustering, multivariate classification trees, and random forests by investigating modern pollen-vegetation relationships using the well-known surface pollen data-set of Lichti-Federovich and Ritchie (1968) from the Western Interior of Canada.Our results show that the modern pollen data display well the major vegetation-landform types from which the surface samples were collected independent of which numerical method is used. Ordination methods partly capture the geographical position of the lakes from which the samples were collected. Different ordination methods produce similar results. Principal curves are not superior to the results from conventional ordination methods. Ward's hierarchical clustering method performs as well as other agglomerative clustering techniques. Multivariate classification trees and random forests produce similar results, but the latter has smaller prediction errors.In general, the novel methods do not provide any new insights into the data, nor do they have any impact on previous conclusions. However, these novel methods have several advantages over traditional techniques because they make fewer assumptions about the underlying character or structure of the data, and they include effective ways of displaying the quantitative importance of different variables. As it is not possible to recommend one method over another, our conclusion is that there is now a large number of robust numerical methods that can be used to detect the underlying patterns within modern pollen assemblages.