Reply to Nash: Color terms are lost, despite missing data.

Haynie and Bowern (H&B) (1) use promising computational phylogenetic methods to test the standard view of color terminology structure, epitomized in the World Color Survey (WCS) (2). H&Bmatched Bayesian phylogenies for 189 Pama–Nyungan languages with the color terms in each vocabulary. Their inferred ancestral state reconstructions found the expected “general support for the WCS model of color term development....” However, H&B unexpectedly claimed to find “extensive evidence for the loss (as well as gain) of color terms” and “exceptions to their [WCS] predicted patterns, such as the loss of color terms in multiple subgroups.” Their inference is invalidated by the way the data were selected, their patchiness, and the assumption that every unknown or missing datum is a true absence. The study was built on the useful Chirila database (3), to which I was kindly granted access. Chirila consists of doculects (ref. 3, p. 14) which are not always the best consolidation of available lexical information. Investigation of other vocabulary sources has supplied 1 “black,” 7 “white,” and about 10 “red” terms overlooked by H&B, eliminated a few spurious “green” terms, and shown that sources lacking terms for black, white, or red are most likely deficient (not evidence for true absence). For example, in H&B’s corrected data table, seven languages are coded as containing a green term and no red term, but four (Dhangu, Djinang, Kukatj, and Kurnu) do have a red term and two (Mbabaram and Wargamay) have no green term, the purported green word denoting only “unripe, raw.” In this light, H&B have not demonstrated loss of these color terms or of evolutionary pathways contrary to those advanced by the WCS. The crucial problem is that H&B’s method does not distinguish between absence and missing: Whenever the sources for a language do not provide a particular color term, that color was coded as unlexified in the language. However, the checklists used over 50 y ago to elicit words of Australian languages contained few, if any, color meanings: black, white, maybe red, and, uncommonly, yellow or green. H&B observe that “the general trend suggested by ancestral node reconstruction probabilities is consistent with WCS evolutionary pathways” (ref. 1, p. 13670), and indeed the nonterminal nodes of their figure 3 show hardly any loss between pairs. Loss is overwhelmingly apparent in their figure 3 in the last step of descent to terminal nodes, which is indicative of gaps in the terminal node data. In general, it may well be difficult to incorporate in the analysis the distinction between “absent” and “missing” data, and the reconstruction of some ancestral states could be indeterminate. H&B do not discuss this or the possible application of computational phylogenetic techniques which do allow for missing data. A recent survey of the impact of missing data on species tree estimation underscores “the importance of understanding the influence of missing data in the phylogenomics era” (4). Details are available at hdl.handle.net/1885/123084.