Numerical analysis of pollen samples from central Canada: A comparison of methods

The compilation of large and complex sets of modern pollen data has stimulated the use of new methods, including the use of multivariate statistical techniques, for summarizing and presenting these data. This paper compares several of these methods by applying them to Lichti-Federovich and Ritchie's (1968) 131 sediment samples of modern pollen from central Canada. Maps of the major pollen types are presented, and the data are analyzed by canonical variates analysis, principal components analysis, principal coordinates analysis, and minimum-variance cluster analysis.The maps show the geographical distribution of the principal pollen types and reveal that steep gradients in the percentages of eight of the nineteen pollen types used in this study separate the samples in the southwest from the remaining samples. Excluding the southwestern samples, the maps show the frequencies of the other pollen types to be aligned north to south with high values of sedge, birch, and heath pollen in the north, high values of pine in the south, and high values of spruce and alder in between. This same general structure is evident in the results of the four multivariate analyses. The samples are distributed in a closely similar manner along the first two axes derived from canonical varietes analysis, principal components analysis, and principal coordinates analysis. The first axis of each analysis separates the samples in the southwest from the rest of the samples, and the second axis shows these latter samples to be spread fairly evenly along a north—south gradient from the tundra samples to the mixed coniferous—deciduous forest samples. Minimum-variance cluster analysis also shows these divisions by clustering the samples into three major groups: the southwestern samples from the prairie, aspen parkland, and deciduous forest; the northern samples from the tundra and forest—tundra; and the intermediate samples from the mixed forest and closed coniferous forest regions. Further division by the clustering technique yields fourteen groups, and these show the pollen samples to cluster slightly differently from their classification based on their location within the vegetational units. For example, the samples of the forest—tundra and the open coniferous forest are grouped together, but the samples of the upland mixed forest are too heterogeneous in pollen composition to be placed in one group.These results indicate the power of these numerical methods that use prescribed mathematical steps to analyze all samples and major pollen types simultaneously and thereby reveal the basic structure in the data based on numerical criteria alone. These summaries aid an investigator in visualizing the important trends and divisions in a data set and in finding those samples needed for a particular comparison.