The 82nd Annual Meeting of the American Association of Physical Anthropologists (2013)

Crania, coordinates, and clusters: testing a finite mixture modeling approach for the detection of population structure in modern America using high-dimensional data


Department of Anthropology, University of Tennessee

Friday Morning, 301D Add to calendar

Finite mixture model-based clustering (FM-MBC) methods are showing promise for revealing structured human variation and teasing-out patterned information useful for craniometric-based classification without the need for reference samples or a priori group-identifiers. FM-MBC studies have revealed that it is possible to detect latent population structure within a mixture of craniometric distributions, that craniometric-based structure is hierarchical, that statistically inferred clusters correspond closely to predefined populations and that the best results are produced using a geographically diverse sample. What remains to be resolved is if FM-MBC analyses can be improved for finer levels of subdivision. Preliminary tests of substructure in contemporary America suggest that the relationships between the individual cluster allocations and true group memberships is complicated by admixture, secular change, and variability in self-identification. As 3-D coordinate data better captures the shape of the cranium, these problems may be overcome by incorporating geometric morphometric methods. This project applies these two powerful resources to the detection of finerā€grained, within-population structure using a dataset of digitized landmarks representing the craniofacial morphology of 600 self-declared American Blacks, Whites and Hispanics from the Forensic Anthropology Databank. This study assesses the cluster-solutions generated using the coordinate dataset and compares these to the cluster-patterns and misclassifications produced using inter-landmark distances for 900 individuals. It asks if FM-MBC can be effectively applied to high-dimensional data, if 3-D data offers improved classifications, and if the potential for improvement warrants the increased analytical complexity. Recommendations are offered for the combined application of these methods and use of open-source software.

This study was funded in part by a National Science Foundation Dissertation Improvement Grant, BCS-676917.

comments powered by Disqus