Science does it again

A new paper out in Science magazine by Quentin Atkinson entitled "Phonemic Diversity Supports a Serial Founder Effect Model of Language Expansion from Africa" goes a long way to show that what counts most in science these days is a good yarn. The abstract is here, and the article by the ever-enthusiastic Nicholas Wade in the New York Times is here:

I will merely place below for public consumption my review of Atkinson's paper for Science. Evidently my arguments were not compelling enough to get in the way of a pretty story.

I have to admit that I'm not sure why I was asked to review this paper. I am not an expert on historical linguistics or particularly on language evolution. I certainly have been an outspoken critic of the reviewing practices of Science; perhaps it was for that reason that I was asked. But if that's the case, then it did nothing to improve my perception of Science's reviewing practices.


This paper is a revision of an earlier paper, which addresses the previous reviewers' criticisms. Atkinson's basic thesis is that a series of migrations from the original African homeland over the course of the last 50-70K years, explains an observed reduction in the average size of phoneme inventories as one moves away from Africa. The smallest phoneme inventories, on average, are found in South America and Oceania, the last places to be colonized by humans.

The thesis depends upon another result, due to Hay and Bauer, but also replicated by Atkinson, that there is a positive correlation between population size and phoneme diversity. Indeed both population size, and distance from Africa are, jointly, strong predictors of phoneme diversity.

The conclusion is that this is strong evidence that language probably evolved in Africa, and moved outward as humans migrated 50-70K years BP.

Atkinson is careful in his analysis, and covers many aspects, including an analysis of 2560 possible origin points, the results of which show that the best models are consistent with an origin somewhere in Central and Southern Africa. To test the possibility of polygenesis, he considers models with a second point of origin. That analysis posits South America as a second point of origin, but this implausible result is argued to be an artefact. He also argues that more recent expansions since the last glacial maximum cannot account for the global cline in phoneme inventory size.

The paper's claims are striking and, if correct, of utmost importance in our understanding of the evolution of human language, and indeed of humankind itself.

But such striking claims require striking support. And, while I cannot truthfully offer an alternative analysis of the data Atkinson presents, I remain seriously concerned about a number of issues.


Population versus size of phoneme inventory

The analysis depends crucially on the relation between population size and phoneme inventory size, discovered first by Hay and Bauer, and replicated in this study. Without that relationship, the story would not work.

In Atkinson's data, the correlation between log population and his normalized phoneme diversity has an r value of 0.385. While the r value is not huge, it is highly significant. But the language population ranges (and this also applies to Hay and Bauer's study) range from a few tens of speakers all the way up to languages with speakers numbering in the hundreds of millions.

Consider these numbers in the context of the time periods Atkinson is talking about, well more than 10,000 BP. Estimates of world populations from that period are obviously uncertain, but a common estimate for 10,000 BP has about 1 million humans on the entire planet. 50-70K BP would presumably have even smaller numbers.

How large could any single population of humans speaking the same language have been prior to 10K BP? If we are very generous and assume that the distribution of languages was similar to today where roughly 1/10 of the world's population speaks Mandarin, then maybe 100K speakers. This seems awfully unlikely --- see below --- but if one grants that there were languages that were that big that long ago, a more appropriate range to consider is languages that have populations between a few tens and about 100,000 speakers. I have not run the numbers, but just looking at the plots in Figure S1, the correlation for that range looks as if it would be much lower than the r=0.385 for the entire range.

But who seriously believes that there were any groups of 100K speakers of a single language at such an early date? Prior to 10K BP, humans lived in hunter-gatherer societies, not the larger settled communities that formed after agriculture, let alone what one saw after the foundation of cities. For such societies very small groups are the norm. Would we expect to see any groups that had populations of more than a few hundred or a few thousand? Precolonial Australia is probably a more reasonable model for the Paleolithic, in which case we are talking about groups of a few thousand max. In that case, if the norm for most languages was indeed a few thousand speakers at a maximum, then all languages in the earliest periods would be tiny.

Demonstrate that there's a strong correlation between phoneme diversity and population size, computed over population ranges plausible for the Paleolithic, and we might have some basis for a discussion. Since there is no natural law that dictates the correlation that Hay and Bauer discovered, and the correlation is lacking in any currently understood explanation, such a demonstration seems a bare minimum requirement.

Another issue is that Atkinson's model assumes that we are talking about loss of inventory when populations get smaller, and that therefore there is an implicit directionality to the "law": if not, then why would migration result in loss of inventory and why wouldn't languages regain lost diversity when the populations regrew? But there's no evidence of a directionality here: Hay and Bauer certainly do not propose any such directionality. The closest thing they are willing to offer as an possible line of explanation of their observation is the work of de Boer on emerging systems. Putting aside the question of whether de Boer's simulations are really relevant to the transmission of already functioning languages to new generations, the model is neutral on the issue of loss in small populations versus gain in larger ones: it merely falls out that one gets fewer phonemes (vowels in de Boer's case) in smaller populations, and more in larger ones.

Indeed, the best explanation I can think of for why such a correlation might obtain is that in large languages one finds dialectal variation, and dialectal variation might plausibly be a source of borrowing of new phonemes into other varieties. But that mechanism would mean that large languages would gain phonemic diversity not that small languages would lose it.

The world population density for the time period we are talking about also raises another issue: surely there would have been ample time for "founder effects" to manifest themselves within Africa. Africa is after all a very large continent, and with such a sparse population groups might have been driven out, or for other reasons sought new lands to colonize without leaving the continent. On that assumption, might one not expect as great a range within Africa as across the entire world?

Furthermore there are all sorts of reasons for small populations besides migration: war, famine and disease would all have taken their toll on prehistoric populations, and caused conditions that might foster a decrease in phoneme inventory --- assuming again that the correlation is robust over small population ranges to begin with. Again this would suggest that one should find equal diversity within Africa as across the rest of the world.

The founder effect

While I understand Atkinson's claim, I must admit to not fully understanding the mechanism by which migration produces smaller phoneme diversity. This relates to the point about the directionality of the "law" discussed above. Putting aside the critiques in the preceding section and accepting that the phoneme diversity/population correlation works for Paleolithic population sizes, and that there is the implicit directionality, how exactly is the story supposed to work? Clearly phoneme diversity reduction must take some time. The Plymouth Bay Colony did not all of a sudden speak a variety of English with fewer phonemes merely because they migrated to the New World. Had they remained an isolated and small colony for many generations, perhaps they might have ended up that way. The question then is how long one would expect that to take. Most small languages today are spoken by groups of people who live in fairly close proximity to groups speaking other small languages: Papua New Guinea is a good example. In such cases, the populations do not have much room to grow, short of conquering their neighbors. Presumably in Paleolithic times there would have been no neighbors close by for any migratory group, so might one not expect that group to grow in size once they settled in a new area? Then if that happens, wouldn't one expect there to be less reduction in the phoneme inventory?

I think such an astounding claim as Atkinson's (even if it is in line with the commonly believed out-of-Africa story involving migrations starting around 70-50Ky BP) requires a carefully articulated model, and I don't believe Atkinson presents any mechanism by which these putative serial reductions in inventory size were accomplished.

The signal

Atkinson presents a number of arguments in response to previous reviewers in support of the idea that one can expect to see a "signal" over such an immense period of time.

Nonetheless, I remain skeptical. Certainly there are cases of fairly massive reductions in phoneme inventory sizes within just a few thousand years. A clear example comes from Austronesian. Blust's reconstruction of Proto-Austronesian has on the order of 30 phonemes (more if we count diphthongs). If we assume something like 5,000 years between the origin of Austronesian in Taiwan, and modern languages, one has seen a massive reduction in the number from around 30 (which is also approximately correct for modern languages like Malay) down to 13 for Hawaiian or 11 for Northwest Mekeo. And that's in 5,000 years, which is less than 10% of the time period that Atkinson is talking about.

True, this particular example is in keeping with the hypothesis of serial founder effects, since phoneme inventory sizes have decreased. But it's not in line with the idea that there is any kind of stability within language families over such time periods, let alone the time periods that Atkinson is talking about. Such variation might be expected to add a lot of noise to the signal.

Other issues

A technical point that needs to be clarified is that the data on vowel, consonant and tone inventories that Ian Maddieson did for WALS does not give actual counts, but count ranges. For example, for consonants we are only given info on whether the language has a small, moderately small, average, moderately large, or large inventory, interpreted thus:

Small:       6-14
Moderately Small: 15-18
Average:	19-25
Moderately Large: 26-33
Large: 34 or more
A similar situation holds for vowels and tones. Does Atkinson have access to data that I'm not aware of? If so, then it would be nice to know more details on that data. But if Atkinson is using the data that anyone can download from the WALS website, it would be good to clarify that he doesn't have the actual phoneme counts, just approximate ranges. Did I somehow miss such a clarification?

I am not sure this will affect the results, but it might.


Atkinson's thesis is striking, but as I said above such striking conclusions require striking support, and I believe that the paper in its current form does not provide enough support.

See here for a Slovenian translation.

See here for a Portuguese translation.

See here for a Russian translation.

Last modified: Sat Aug 11 12:36:27 EDT 2018