I’ve now had a chance to take a closer look at Caleb Everett’s recent PLoS ONE article on the correlation between ejective consonants and altitude, and I think the ejective

distribution is explainable on purely classical grounds, that is,

common inheritance and areal diffusion. That is, it is NOT necessary to resort to

an explanation in terms of altitude. The classical explanation is preferable

since the mechanisms of inheritance and diffusion are observable in

plenty, whereas noone has yet observed the hypothesized mechanism whereby

the speakers start using more ejectives in response to moving to higher

altitudes.

Here’s why: The UPSID database via WALS contains 527 languages coded

for ejectiveness, with 87 with ejectives (having WALS value 2, 5, 6,

or 8) and 440 without ejectives. The 87 languages with ejectives

belong to 54 families (ejectives_families). The 54 families are not necessarily

independent because some of them are geographically contingent. I

grouped the 87 languages into areas such that the languages within an

area could plausibly be connected by adjacency or via a path of

adjacent languages also known to have ejectives. There is some amount

of leeway here but if there really is a causal link between altitude

and ejectives it should survive nevertheless. The 87 languages divide

into 19 areal groups (listed in ejectives_areal).

Now if we pick languages such that no two of them belong to the same areal

group or the same family, then there are arguably genealogically and areally

independent. Since there are 19 areal groups (and 54 families) one cannot

pick more than 19 independent languages, but it turns out that it is

possible to pick 19 languages which are indeed both areally and

genealogically independent. I picked a random such set, shown below,

together with their altitude (as per the Shuttle Radar Topography Map at

their centre coordinate):

Hadza [hts] 1049

Yapese [yap] 0

Hamtai [hmt] 1175

Hausa [hau] 594

Yuchi [yuc] 263

Trumai [tpy] 312

Southern Nambikuara [nab] 640

Zuni [zun] 2009

Itelmen [itl] 35

Itonama [ito] 156

Michoacan Mazahua [mmc] 3399

Ona [ona] 117

Tsimshian [tsi] 177

Soqotri [sqt] 364

Korean [kor] 158

Shua [shg] 1057

Wichi Lhamtes Guisnay [mzh] 208

Abkhaz [abk] 0

Central Aymara [ayr] 4298

The mean elevation of these languages is 842.6 m and the median elevation 312 m. Is

this higher than expected for “random” languages? I took 1000 samples of

19 languages from different families and checked their mean and median

altitude. 56 of the 1000 random samples had a higher mean altitude than

842.6 m and 296 of the 1000 random samples had a higher median than 312 m,

giving p-values of 0.056 and 0.296 respectively. In other words, NOT

SIGNIFICANT by conventional significance limits.

So why does Everett claim to have controlled for areality and genealogy

and still reports a significant correlation (in his Table 3 and associated

text)? (The remainder of the significance results reported in his paper

treats languages as independent, not controlling for areality/genealogy.)

1. Everett doesn’t measure the direct altitude of a language but the

proximity to a “major region” of high elevation. What counts as a

major region is never objectively defined, and clearly there is some

cheating going on when using the distinction “major/non-major” because

the proposed theory (which has to do with air pressure etc) is

oblivious as to whether the speakers are in a “major” or “minor”

region of altitude. The difference is significant because it turns over

many datapoints, so that, e.g., the New Guinea Highlands is not counted

as a high-altitude region!

2. The division of the non-ejective languages into clusters “by assuming

clusters contain ten languages each, in keeping with the approximate size

of the clusters of languages with ejectives” is underspecified. This is

a highly non-trivial step (there are *many* ways to divide non-ejective

languages into clusters containing ten each) and the fact that it is not

explained means that the experiment is not reproducible. If I was the

reviewer I would never have let that through.

3. There is arbitrariness with measures such as 200 km and 500 km from a

major region, and some comments in the paper suggest that data dredging

has taken place. That is, comments of the kind “only marginally further

away from high elevation at 380 km”, might reflect that different thresholds

have been tried and 200/500 km are included in the paper because they

happened to give the best results. If so, all p-values should be corrected

for multiple testing. Because of the irreproducibility (as of point 2 above)

I cannot re-do the test with other thresholds than 200/500 km to check.

One should perhaps also not forget that the UPSID database in the first

place is not a random sample of the languages (or families or regions) of

the world and therefore it is not legitimate to draw statistical

conclusions based on it. It would be better to re-check it on Steven Moran’s or

Mark Donohue’s databases.

One should not forget that the location points for the languages in WALS are another source of potential arrors. The location points are misleading because first they reduce a language’s distribution to one point while in most cases it ought rather be an area (which might very well span quite various altitudes) and second because the points were arranged to allow maximum visibility / minimum point overlap.