Ilmu Sosial editorial
Open AccessOA2024
Biases of Principal Component Analysis (PCA) in Physical Anthropology Studies Require a Reevaluation of Evolutionary Insights
A new study evaluates the standard two-step approach of Generalized Procrustes Analysis (GPA) followed by Principal Component Analysis (PCA) in physical anthropology, finding that PCA results are not reliable, robust, or reproducible. The authors introduce MORPHIX, a Python package that shows supervised machine learning classifiers outperform PCA for classification and detecting new taxa.
Nima Mohseni; E. Elhaikยท bioRxivยท 2024ยท DOI 10.1101/2023.12.06.570437
The core problem
Physical anthropologists, including palaeoanthropologists, anatomists, and ontogenists, rely on geometric morphometrics to quantitatively analyze physical forms such as skull morphology. The standard workflow involves two steps: Generalized Procrustes Analysis (GPA) to superimpose landmark coordinates, followed by Principal Component Analysis (PCA) to project the data onto uncorrelated variables for visualization and interpretation. PCA outcomes are commonly used to draw phenetic, evolutionary, and ontogenetic conclusions. However, recent challenges to PCA's use in genetic studies have raised concerns about its central role in morphometrics. This study aims to evaluate the accuracy, robustness, and reproducibility of PCA in physical anthropology, particularly in light of contentious cases like Homo Nesher Ramla, an archaic hominin with questionable taxonomy. The authors developed MORPHIX, a Python package for processing superimposed landmark data with classifier and outlier detection methods, to test PCA against benchmark data of crania from five papionin genera.
Innovation
The authors employed benchmark data consisting of crania from five papionin genera. They developed MORPHIX, a Python package that processes superimposed landmark data using classifier and outlier detection methods, with visualization capabilities. The study compares PCA outcomes with supervised machine learning classifiers for classification tasks and detection of new taxa. The methodology involves applying GPA to obtain superimposed coordinates, then performing PCA to project data onto principal components. The accuracy, robustness, and reproducibility of PCA are assessed by comparing its outcomes with those from machine learning classifiers. The authors also examine the use of PCA in phylogenetic inference, using the case of Homo Nesher Ramla as an example. The analysis focuses on whether PCA results are artefacts of the input data and whether alternative methods provide more reliable classifications.
Introduction
Physical anthropologists, including palaeoanthropologists, anatomists, and ontogenists, rely on geometric morphometrics to quantitatively analyze physical forms such as skull morphology. The standard workflow involves two steps: Generalized Procrustes Analysis (GPA) to superimpose landmark coordinates, followed by Principal Component Analysis (PCA) to project the data onto uncorrelated variables for visualization and interpretation. PCA outcomes are commonly used to draw phenetic, evolutionary, and ontogenetic conclusions. However, recent challenges to PCA's use in genetic studies have raised concerns about its central role in morphometrics. This study aims to evaluate the accuracy, robustness, and reproducibility of PCA in physical anthropology, particularly in light of contentious cases like Homo Nesher Ramla, an archaic hominin with questionable taxonomy. The authors developed MORPHIX, a Python package for processing superimposed landmark data with classifier and outlier detection methods, to test PCA against benchmark data of crania from five papionin genera.
Why it matters
The findings challenge the widespread use of PCA in physical anthropology and phylogenetic inference. The authors argue that PCA's limitations stem from its sensitivity to input data, which can lead to misleading interpretations of evolutionary relationships. They propose that supervised machine learning classifiers offer a more accurate alternative for classification and outlier detection. The study's implications extend to thousands of published studies that have relied on PCA, suggesting a need for reevaluation of prior claims concerning the origins and relatedness of inter- and intra-species. The authors emphasize that their work can be used to improve phylogenetic and taxonomic reconstructions by providing a more reliable methodological framework. They also note that the use of PCA in genetic studies has been previously challenged, and their findings extend these concerns to morphometrics. The development of MORPHIX provides a tool for researchers to process landmark data with advanced classifiers and outlier detection, facilitating more robust analyses.
Who should read this
Pembaca Research Digest dan praktisi rumpun ini.
Opening member contentโฆ