Less tautology, more biology? A comment on “high-density” morphometrics

Abstract

In the context of geometric morphometric analyses of modularity and integration using Procrustes methods, some researchers have recently claimed that “high-density geometric morphometric data exceed the traditional landmark-based methods in the characterization of morphology and allow more nuanced comparisons across disparate taxa” and also that, using “high-density” data (i.e., with dozens or hundreds of semilandmarks), “potential issues [with tests of modularity and integration] are unlikely to obscure genuine biological signal”. I show that the first claim is invalidly tautological and, therefore, flawed, while the second one is a speculation. “High-density” geometric morphometrics is a potentially useful tool for the quantification of continuous morphological variation in evolutionary biology, but cannot be said to represent absolute accuracy, simply because more measurements increase information, but do not by default imply that this information is accurate. Semilandmarks are an analytical expedient to break the continuity of regions devoid of clearly corresponding landmarks, but the shape variables which they generate are a function of the specific choice of the placement and possible mathematical manipulation of these points. Not only there are infinite ways of splitting a curve or surface into discrete points, but also none of the methods to slide the semilandmarks increases the accuracy of their mapping onto the underlying biological homology: indeed, none of them is based on a biological model, and the assumption of universal equivalence between geometric and biological correspondence is unverified, if at all verifiable. Besides, in the specific context of modularity and integration using Procrustes geometric morphometrics, the limited number of scenarios simulated until now may provide interesting clues, but do not yet allow strong statements and clear generalizations. The Procrustes superimposition does alter the ‘true’ covariance structure of the data and sliding semilandmarks further contributes to this change. Although we hope that this might only add a negligible source of inaccuracy, and simulations using landmarks (but no semilandmarks yet) suggest that this might be the case, it is too early to confidently share the view, expressed by the promoters of high-density methods, that this is “Not-Really-a-Problem”. The evidence is very preliminary and the dichotomy may not be this simple, with the magnitude (from negligible to large) and direction (inflation of modularity, integration, or both) of a potential bias in the tests likely to vary in ways specific to the data being analysed. We need more studies that provide robust and generalizable evidence, without indulging in invalid tautology and over-interpretation. With both landmarks and semilandmarks, what is measured should be functional to the specific hypothesis and we should be clear on where the treatment of the data is pure mathematics and where there is a biological model that supports the maths.