Some issues with the correlated-evolution method for testing causal hypotheses in comparative linguistics

While comparative grammar research in the 20th century based its universal claims on stratified sampling (e.g. Bell 1978, Bakker 2011), in the 21st century, some authors have emphasized that sampling does not solve the issue of non-independence because all languages probably derive from a common ancestor or ancestral bottleneck (e.g. Maslova 2000; Levinson et al. 2011: 512). They have therefore given preference to the correlated-evolution method (Felsenstein 1985; Mace & Pagel 1994) that is firmly established in biology. Some representatives of this trend are Dunn et al. (2011), Widmer et al. (2017), and Jäger (2019).

There is no doubt that this method can be used in linguistics as in biology, but how useful is it for advancing our understanding of worldwide patterns? The fact that Dunn et al. (2011) found no correlated-evolution evidence for the dependency length theory of word order correlations (Dryer 1992; Hawkins 1994; Temperley & Gildea 2018) is good reason to be skeptical of this particular paper’s approach, because the dependency length theory has great explanatory power, far beyond the small set of grammaticalized word order patterns that Dunn et al. examined. (See here for an accessible introduction to the paper by Dunn et al.) In particular, it explains a large number of word order variation effects, as documented in detail by Hawkins and others. Nobody has doubted that online word order variation in flexible constructions is largely due to processing efficiency, and the match between grammatically conditioned variation and online variation is so clear that this connection is also beyond reasonable doubt (for example, relative clause extraposition in English is grammatically constrained, but clearly linked to processing efficiency).

And once one examines a larger set of languages, the evidence for the Greenbergian word order correlations becomes overwhelming (Dryer 2011; Jäger 2019). Dunn et al. (2011) did not find evidence for correlated evolution, but they were clearly limited in their findings by their small sample size (only four families). Even though Levinson et al. (2011: 520, n. 5) emphasize that there were about 240 character state changes in their data, this is evidently not enough to detect the pattern. They think that their sample size is “large enough for any reasonably strong bias to shine through” (2011: 520), but what do they mean by “reasonably strong”? Authors such as Greenberg (1963), Nichols (1992) and Dryer (1992) were not motivated by the desire to prove any particular theory of word order correlations, but they pointed to existing patterns, even relatively weak ones, and looked for explanations. Tentatively, Dryer (1992) favours the dependency length theory, and so far, this is clearly the main contender and has little competition. (At some point, generative syntacticians thought they could contribute to explaining the correlations, e.g. Baker 2001, but this hope has been largely given up.) So even if the effect is not strong enough to show up in Dunn et al.’s results, it appears to be strong enough to result in worldwide distributions that can be explained (“for free”) by a theory that is independently needed and motivated.

Levinson et al. (2011: 513) admit that if there is no evidence for causal connections (presumably due to cognitive biases), “the distributional patterns would still need explanation” – but what kind of explanation? They merely say: “presumably in purely historical terms”. But “history” is not an explanation of general patterns, and the patterns have long been seen as strong enough that historical accident can be ruled out. Or would they want to suggest a historical explanation in terms of two ancient protolanguages, one showing VO&NGen order, and the other showing OV&GenN order, with most modern VO languages descending from the first, and most OV languages descending from the last? This is of course extremely implausible, but it is the only way in which I can make sense of the suggestion to explain the correlations “in purely historical terms”.

If one inspects the trees of Austronesian, Indo-European and Uto-Aztecan languages in Dunn et al. (2011), one immediately sees that almost all of the word order changes cluster geographically in such a way that they must be understood as contact-induced. For example, the  subtree figure copied here from the overview page shows OV/postpositional order in the east and VO/prepositional order in the west, with Iranic languages showing intermediate patterns (the prepositions in Iranic languages reflect the geographically intermediate position of this subfamily).

Apparently “spontaneous” word order changes are simply not frequent enough to be detected by looking at a handful of families.

So it seems that weak effects that show up only in large samples (but not necessarily in four families) cannot be detected by the correlated-evolution method. But maybe this method is more suited to phenomena that are due to a stronger causal factor?

Let us take the example of triangular vowel systems (i-a-u, or i-e-a-o-u), which are widely thought to be explained by optimal acoustic dispersion (e.g. Gordon 2016: 58). This seems to be a very strong causal factor, because there are (virtually) no languages with linear vowel systems (no language has only i-e-a, or only u-o-a, or only i-ɨ-u, or only æ-a-ɒ). Does correlated evolution offer any insights here? A generally skeptical attitude might lead us to suspect that the triangular form of vowel systems was actually inherited from Proto-World and has simply remained constant, rather than being recreated by each new generation in some way. Since (virtually) all languages have a triangular system, this is hard to disprove in a systematic way. But on the other hand, every historical linguist knows that the Proto-World explanation must be wrong. There are many attested vowel changes (e.g. ō > ē > ī in English feet, a > ō > u in German Mutter, u > o > a in English god, au > æa > ī in English beam), whose rationale is typically mysterious – they appear to be random changes with no particular motivation. But in the end, languages always end up with triangular vowel systems, strongly suggesting that there is an adaptive attractor state. Here one might want to demand that it must actually be shown that a change u > a actually happens at the same time as some other change X > u (so that it is clear that the rise of a new u is causally connected with the loss of the old u), but this cannot be done by means of the biology-inspired correlated-evolution method. In biology, all we can observe is synchronic states of DNA, and we infer the changes from the distributional patterns. In sound systems, by contrast, we sometimes know what must have happened to actual words, and typological comparison does not offer any additional insights here. (On the contrary, it can lead to further confusion.)

Another striking universal pattern is the relationship between length of words (and other expressions) and their frequency of use (“Zipf’s Law of Abbreviation”, e.g. Bentz & Ferrer-i-Cancho 2016). This is generally explained by information-theoretic optimization (e.g. Ferrer-i-Cancho 2018), and this explanation is again without a serious contender. Does correlated evolution throw any light on it? Is there a need to disprove an alternative possibility, namely that this pattern was inherited from Proto-World? Again, as in the case of triangular vowel systems, this is a theoretical possibility, but everyone knows that this cannot be true. New words arise all the time and old words die, and there is no hint that somehow the length of new words is constrained by the length of dying words. So clearly, Proto-World cannot play a role in explaining this pattern. Moreover, all languages have words with completely new meanings that did not exist in the past, and these words obey the same regularities. For example, the word for ‘windshield wiper’ is longer than the word for ‘tyre’ in most languages, for good Zipfian reasons (we talk much more about tyres than about windshield wipers because they are far more critical to the functioning of a car). This appears sufficient to prove that the explanation cannot be the ancestral language, and since there is no alternative to the information-theoretic (efficiency-based) explanation, we do not need a special biology-based method in addition to the traditional sampling method. Thus, it is not generally true that “universal typological dependencies should be detectable in the history of language families” (Levinson et al.’s 2011 paper title).

Jäger (2019) applied the correlated-evolution method to differential object marking (DOM). He did find a significant effect, thus confirming earlier claims (e.g. Bossong 1991; Aissen 2003; Schmidtke-Bode & Levshina 2018). But is his approach in any way an advance over the sampling-based methods? Sampling is intended to minimize not only genealogical non-independence, but also geographical non-independence. If one bases one’s inferences about past changes exclusively on typological data, it seems that one can go quite wrong in one’s reconstructions. In DOM, there are some striking areal effects, as it is particularly widespread in central and southern Asia – it is found throughout Indic, Dravidian, and Trans-Himalayan (apart from Sinitic). Apparently the correlated-evolution method will give us a great likelihood of reconstructing DOM to the protolanguages here, but an alternative possibility is that the DOM pattern spread much more recently throughout the region. LaPolla (1994) showed that DOM across Trans-Himalayan is quite similar typologically, but all these languages use different markers – so we really don’t have any evidence that the pattern goes back to the protolanguage. Maybe these contact-induced effects do not pose a specific problem to Jäger’s approach (I may not  have understood it well enough), but since the contact-induced word order changes had a serious effect on Dunn et al.’s (non-)results, I have to mention them here.

In a personal exchange, Gerhard Jäger told me that he thought that “which method is best to correct for the sampling problem in typology cannot be answered by pure reasoning; it is an empirical question.” If so, I would like to see the empirical reasons for favouring the correlated-evolution approach for SOME phenomena – the mere fact that there is a consensus around them in biology is not sufficient for claiming that they are needed in linguistics, too.

One additional argument that is often given in favour of the correlated-evolution approach is that it allows us to make use of all the data, and doesn’t force us to “throw away data”. But this is an argument in favour of these approaches only if the high-quality data of the relevant kind already exists. If such data do not yet exist, then the fact that these methods need a very large amount of data is a strong argument against them. Linguists are often interested in fine-grained questions, for which the collection of data from hundreds of languages is impossible or prohibitively expensive. So of course, they need to base their research on a sample of languages that is as representative as possible, probably involving only one language per family.

In a recent paper about similarities between biology and linguistics, biologist Lindell Bromham formulates the problem of non-independence as follows:

“Adding language family as a factor in the analysis does not remove the problem, because phylogenetic non-independence occurs at all level of the genealogical hierarchy: closely related languages will tend to be more similar to each other than to their more distant relatives within the same family, and more closely related language families will tend to share more in common with each other than more distant families … For the same reason, balanced sampling (choosing one representative of each group) will also not make the problem of non-independence go away: there will still tend to be more similarities between representatives from closely related groups than to more distant groups. The paucity of practical solutions to Galton’s problem in linguistics is odd given the amount of energy expended on solving the same problem in evolutionary biology.” (Bromham 2017)

In this passage, Bromham clearly presupposes that a dataset of comparative linguistics contains more than one language per family. If there is only one language per family, these considerations play no role. The difference between the usual practice in linguistics and the practice in biology is not odd at all if we take into account the enormous costs of creating worldwide datasets in linguistics. Biologists have engaged in serious worldwide comparisons for centuries, but linguists have been doing this only since the 1960s, and how to compare languages is very contentious in the discipline (which results in great difficulties in securing research funding). We have nothing like the large-scale molecular datasets that are now routinely used by biological evolutionists.

So in practice, it appears that “phylogenetic comparative methods” (= correlated-evolution methods) are applicable only in a small subset of situations, and it is not clear to me that they are superior to the traditional sampling method. Where they give different results from other methods, it is unclear what can be concluded from this. So I see the burden of proof on those who think that these methods are superior to the traditional, much simpler and much cheaper methods.


Aissen, Judith. 2003. Differential object marking: Iconicity vs. economy. Natural Language & Linguistic Theory 21(3). 435–483.

Baker, Mark C. 2001. The atoms of language. New York: Basic Books.

Bakker, Dik. 2011. Language sampling. In Jae Jung Song (ed.), The Oxford handbook of linguistic typology, 100–127. Oxford: Oxford University Press.

Bell, Alan. 1978. Language samples. In Joseph H. Greenberg (ed.), Universals of human language, vol. 1, 123–156. Stanford: Stanford University Press.

Bentz, Christian & Ramon Ferrer-i-Cancho. 2016. Zipf’s law of abbreviation as a language universal. In Christian Bentz, Gerhard Jäger & Igor Yanovich (eds.), Proceedings of the Leiden Workshop on Capturing Phylogenetic Algorithms for Linguistics. Tübingen: University of Tübingen. https://doi.org

Bossong, Georg. 1991. Differential object marking in Romance and beyond. In Douglas Kibbee & Dieter Wanner (eds.), New analyses in Romance linguistics, 143–170. Amsterdam: Benjamins.

Bromham, Lindell. 2017. Curiously the same: swapping tools between linguistics and evolutionary biology. Biology & Philosophy 32(6). 855–886.

Dryer, Matthew S. 1992. The Greenbergian word order correlations. Language 68(1). 81–138.

Dunn, Michael, Simon J. Greenhill, Stephen C. Levinson & Russell D. Gray. 2011. Evolved structure of language shows lineage-specific trends in word-order universals. Nature 473(7345). 79–82. https://doi.org10.1038/nature09923.

Felsenstein, Joseph. 1985. Phylogenies and the comparative method. The American Naturalist. The University of Chicago Press 125(1). 1–15. https://doi.org10.1086/284325.

Ferrer-i-Cancho, Ramon. 2018. Optimization models of natural communication. Journal of Quantitative Linguistics. Routledge 25(3). 207–237. https://doi.org10.1080/09296174.2017.1366095.

Gordon, Matthew Kelly. 2016. Phonological typology. Oxford: Oxford University Press.

Greenberg, Joseph H. 1963. Some universals of grammar with particular reference to the order of meaningful elements. In Joseph H. Greenberg (ed.), Universals of language, 73–113. Cambridge, MA: MIT Press.

Hawkins, John A. 1994. A performance theory of order and constituency. Cambridge: Cambridge University Press.

Jäger, Gerhard. 2019. Referential scales and differential case marking: A study using hierarchical models in Bayesian phylogenetics. In. Pavia.

LaPolla, Randy J. 1994. Parallel grammaticalizations in Tibeto-Burman languages: Evidence of Sapir’s ‘Drift.’ Linguistics of the Tibeto-Burman Area 17(1). 61–80.

Levinson, Stephen C., Simon J. Greenhill, Russell D. Gray & Michael Dunn. 2011. Universal typological dependencies should be detectable in the history of language families. Linguistic Typology 15. 509–534.

Mace, Ruth & Mark Pagel. 1994. The comparative method in anthropology. Current Anthropology 35(5). 549–557.

Maslova, Elena. 2000. A dynamic approach to the verification of distributional universals. Linguistic Typology 4. 307 – 333.

Nichols, Johanna. 1992. Linguistic diversity in space and time. Chicago: The University of Chicago Press.

Schmidtke-Bode, Karsten & Natalia Levshina. 2018. Assessing scale effects on differential case marking: Methodological, conceptual and theoretical issues in the quest for a universal. In Ilja A. Seržant & Alena Witzlack-Makarevich (eds.), Diachrony of differential argument marking (Studies in Diversity Linguistics 19), 463–489. Berlin: Language Science Press.

Temperley, David & Daniel Gildea. 2018. Minimizing syntactic dependency lengths: Typological/cognitive universal? Annual Review of Linguistics 4(1). 67–80. https://doi.org10.1146/annurev-linguistics-011817-045617.

Widmer, Manuel, Sandra Auderset, Johanna Nichols, Paul Widmer & Balthasar Bickel. 2017. NP recursion over time: Evidence from Indo-European. Language 93(4). 799–826. https://doi.org10.1353/lan.2017.0058.

2 thoughts on “Some issues with the correlated-evolution method for testing causal hypotheses in comparative linguistics

  1. Gerhard Jäger hat mir zu diesem Blogpost per E-mail geschrieben, “dass ich die statistischen/phylogenetischen Methoden nicht für “superior” halte, sondern für eine Ergänzung zu den klassischen Methoden. Vielleicht ein Punkt, wo die modernen Methoden etwas mehr liefern als manuelle Arbeit: Ich habe (leider noch nicht aufgeschrieben) abgeschätzt (z.B. hier mit phylogenetischen Methoden Evidenz gefunden, dass diachron jede Lineage langfristig zu ca. 2/3 der Zeit im Zustand SOV verbringt. SVO-Sprachen sind gegenwärtig wegen der schnellen Ausbreitung von Bantu, Austronesisch und West-IE massiv überrepräsentiert. Sowas wurde zwar in der typologischen Literatur schon öfter vermutet, aber die diachrone Generalisierung bekommt man “zu Fuß” nicht so ohne weiteres hin. – Außerdem möchte ich noch anmerken, dass die statistischen Methoden halt die methodischen Vorzüge der naturwissenschaftlichen Methode haben, also die Prämissen einer Analyse und die Inferenzmethoden liegen offen zutage und können im Prinzip von jeder interessierten Person repliziert werden.”

  2. I do not think we know triangular vowel system inheritance from PW to be wrong. At least if by “triangular” we really mean “not linear” as your examples seem to suggest (= including also “square” vowel systems such as /i (e) æ ɑ (o) u/), then languages do not “end up” as nonlinear: they simply have been, in every known language family, always nonlinear as far back as we can reconstruct. There are simply no examples of a language starting out with a linear vowel system and proceeding to evolve into having a nonlinear one. So I actually do not see much evidence against the hypothesis that linear vowel systems would be also stable, but they regardless do not occur due to PW having had a nonlinear vowel system. It seems plausible this could be a similar bistable attractor as the presence of clicks (languages that have clicks are very likely to continue maintaining them, while languages that do not have clicks are very likely to continue not introducing them; this suggests that at least Proto-Exo-African rather probably did not have clicks).

    The more specific version where by “triangular” we mean “without front/back contrasts in open vowels” (common, but far from universal) is not very clear either really. Triangular / rectangular seems to me like a fairly stable parameter across language families’ histories: e.g. Dravidian is strongly triangular-maintaining, Indo-European and Austronesian are mostly triangular-maintaining, Uralic is mostly rectangular-maintaining; and any transitions between these states tend to mark off major subgroups. I wouldn’t conclude that triangular systems being the more common ones today is due to Proto-World specifically, but then mere commonness doesn’t need quite that much anyway: it could result just from some key deep nodes (e.g. Proto-Australian, Proto-Amerind) having had triangular systems.

    A much better example of a typological observation on vowel systems that can be sufficiently explained via adaptation is perhaps the back/front and rounded/unrounded correlation. In this case we DO know e.g. that whenever nonback vowels evolve into back vowels, it is highly likely for this to also introduce rounding (a > o is a very common change while the seemingly simpler a > ɤ is rather rare). This fact would be difficult to explain without recourse to acoustic dispersion. Inversely, too, among front vowels unconditional y > i or ø > e are common, unconditional i > y or e > ø are just about unattested.

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.