Statistical uncertainty in forest composition estimates obtained from fossil pollen spectra via the R-value model

Abstract

Estimates of past forest composition obtained from Late Quaternary pollen spectra via a calibration of modern pollen spectra in terms of species abundances are subject to various sources of error, whose combined effect requires statistical analysis. Two statistical procedures, the maximum likelihood method and an approach using series expansions, are used to estimate standard deviations associated with forest composition estimates obtained via the R-value method of calibration. The two approaches yield similar values. The series expansion method also allows one to allocate pollen counting effort between fossil and modern samples in such a way as to maximize the precision of the final estimates. M.B. Davis's original controversial estimates of early Holocene forest composition in Vermont, U.S.A., are shown to have been vitiated by statistical errors. The optimum allocation procedure here suggests increasing the relative effort put into the modern count. This change would have improved but not rescued the estimates; omission of Larix, however, led to a substantial reduction in the errors. Exceptionally poor pollen producers such as Larix should generally be excluded from quantitative calibration; the remaining taxa should be calibrated on the basis of large samples of pollen, the modern pollen being collected preferably from a network of surface sampling sites.