Recognizing species diversity among large-bodied hominoids: a simulation test using missing data finite mixture...

A persistent problem in paleoanthropology is the recognition of intra- vs. inter-specific differences within fossil samples. Exacerbating this situation is the often fragmentary nature of the fossils themselves, thus precluding rote applications of many multivariate approaches designed for complete case analyses. In this paper we apply finite mixture analysis to samples of large-bodied hominoids to test this procedure's efficacy in clustering individuals by species without a priori knowledge of group membership. In addition, we stochastically remove individual specimens and measurements, simulating small, incomplete fossil samples, and re-apply the finite mixture procedure to test how often it correctly assigns these "fragmentary" specimens. Finite mixture analysis can be highly accurate, even when confronted with small sample sizes and missing data. For example, a combination of 124 chimpanzees and humans are correctly identified in one analysis, and the accuracy drops only 2% to become 98% when the total sample size is reduced to 16 and missing data patterns are applied. In comparisons to better known methods that have been used to recognize groups in the fossil record, such as k-means, the benefits of finite mixture analysis are readily apparent. First, k-means is unable to accommodate missing data, an obvious deficiency when investigating the fossil record. Second, in direct comparisons of their ability to accurately assign "unknowns" to taxa, finite mixture performed at least as well as, and often better than, k-means in our analyses. A potential test that can be used to identify species in the fossil record, derived from comparisons of results generated from a general vs. a restricted (isometry-corrected) finite mixture analysis, is presented.