Random walk as a null model for high-dimensional morphometrics of fossil series: geometrical considerations

Abstract

Over the past quarter-century there has been considerable innovation in methods for assessing the tempo and mode of evolution in paleobiological data sets. The current literature of these methods centers on three competing hypotheses—stasis, random walk, and directional trend—corresponding to an increasing scaling of variance with time interval (unchanging, for stasis; linear, for random walk; quadratic, for trend). For applications to a single trait there are powerful methods for discriminating among these hypotheses; but for multivariate data sets, especially the very high-dimensional multivariate data arising in image-feature-based and morphometric studies, current statistical approaches appear to be of less help. This paper proves that in the limiting case of high-dimensional morphospaces, the principal component or principal coordinate ordination of every sufficiently lengthy isotropic random walk tends to the same geometrical shape, which is not that of an ellipsoid and for which the principal components or coordinates are not independent even though they are uncorrelated. Specifically, the “scatter” of PC1 against PC2 is just a parabolic curve. The quantitative characteristics of this specific shape are not described appropriately by the corresponding “covariance structure” or Gaussian model, and the discrepancy may be pertinent to much of the existing literature of methods for differentiating among those three models of evolutionary multivariate time series. From a close examination of this common geometry of the ideal random walk model as seen in its principal components, I suggest a test for stasis, along with a mixed model illustrated by a reanalysis of some data of Gunz et al., and a related test for directional trend. These comments are intended to apply to all high-dimensional morphospaces, not just those arising in geometric morphometrics. Applications of principal components in this context distort high-dimensional data in ways that have a tendency to mislead; but these distortions can be intercepted so that studies of tempo and mode can nevertheless proceed.