Obtaining accurate and precise environmental reconstructions from the modern analog technique and North American surface pollen dataset

Using a recently compiled surface-sample pollen dataset for North America, we test methods to improve the skillfulness (i.e. accuracy and precision) of the modern analog technique (MAT) for continental- to sub-continental-scale paleoenvironmental reconstructions. Interregional floristic differences challenge such reconstructions, because with widening spatial extent each pollen type represents an increasing number of species, each with its unique niche, thus blurring pollen-climate relationships. We conduct a series of cross-validation experiments in which we first explore (1) whether increasing the spatial extent of the pool of potential analogs improves or worsens MAT skill, (2) whether MAT skill is improved by increasing the number of pollen taxa and splitting pollen taxa into regional groups, and (3) the differences in MAT skill among environmental variables. Regional splits are guided by the range boundaries of parent species. Results are analyzed for multiple environmental variables. We then systematically explore options for (a) the value of the no-analog/analog threshold (T), (b) the maximum number of modern analogs (N) allowed for a target fossil sample, and (c) whether the environmental average constructed from the modern analogs should be weighted by their compositional dissimilarity (D) to the target sample. We demonstrate that substantial interregional differences in pollen-climate relationships, particularly between eastern and western North America, degrade MAT precision and accuracy, but these adverse effects can be overcome by expanding the list of taxa used and/or splitting pollen types into regional groups. MAT precision was best when pollen types were regionally split and better when more taxa were used, although increasing the taxon list from 64 to 135 types did not substantially increase performance. Temperature-related variables were reconstructed more precisely than hydrological variables, and there was little difference in MAT skill between climatic and bioclimatic variables (e.g. mean July temperatures versus growing degree days). T scales with the number of taxa analyzed, and there is a tradeoff between skill (best when T is low) and utility (if T is low, few samples can receive environmental inferences). For N, there is a tradeoff between precision and accuracy, such that retaining just the single best analog had the worst precision and best accuracy. Strongly weighting by taxonomic dissimilarity (1/D2) consistently reduced precision, but MAT precision was similar for unweighted and inverse-distance weighted averages. Given the above tradeoffs, we recommend using the 64-taxon list with taxa split by region, 0.20⩽T⩽0.30, 3⩽N⩽7, and either no weighting or an inverse-distance weighting, for North American applications of the MAT.