Phylogenetic methods for cultural traits: reading history off a language tree

Here is a question that sounds unanswerable. Where did newly married couples live in the western Pacific five thousand years ago? Houses rot, and a rule about whose family the couple joins leaves nothing for an archaeologist to dig up. Yet a 2009 paper by Fiona Jordan, Russell Gray, Simon Greenhill and Ruth Mace gave an answer with stated confidence, using present-day ethnographic descriptions and a family tree of languages.

The trick is to treat a language tree as a rough history of the societies that speak those languages, and then ask which history of a custom fits that tree best.

Why a tree at all

Our explainer on Galton's problem showed what happens when related societies are counted as independent cases: one old custom, inherited by twenty descendant societies, looks like twenty pieces of evidence. Ruth Mace and Mark Pagel argued in 1994 that anthropology should answer this the way evolutionary biology does. Species also share traits because they share ancestors, and biologists had learned to make their comparisons along a phylogeny, a tree of descent, instead of across a flat list.

For human societies the best available tree is usually linguistic. Languages split as communities move apart, and the shared vocabulary of related languages records those splits. Gray, Alexei Drummond and Greenhill (2009) built a tree of 400 Austronesian languages from lexical data with Bayesian methods. It places the family's origin in Taiwan about 5,230 years ago and shows a series of settlement pauses and expansion pulses across the Pacific. On a tree like that, each living society sits at a tip, and every internal branch point stands for an ancestral community whose customs nobody wrote down.

The working assumption is that customs such as descent rules or residence after marriage pass down the same branches as words, at least most of the time. That assumption is the method's strength and its weak point.

Reconstructing an ancestor by hand

Take the smallest tree that makes the point. Three societies, A, B and C. A and B speak closely related languages and share a recent ancestor, N. C is a more distant relative, and its line meets N's at the root, R.

        R
       / \
      N   C
     / \
    A   B

Code each society for one trait: 1 if couples usually settle with the wife's family (matrilocal), 0 if they usually settle with the husband's (patrilocal). A and B are matrilocal; C is patrilocal. What was R?

First pass: count the changes

The oldest approach, parsimony, picks the history with the fewest changes. Work up from the tips. A and B agree on 1, so N is 1 with no change needed. At R the two branches disagree, N saying 1 and C saying 0, so whichever state R had, one change happened below it. Parsimony shrugs: R could be 1 with a switch on the branch to C, or 0 with a switch on the branch to N. One change either way, no preference.

Second pass: put probabilities on the branches

Probabilistic methods replace "fewest changes" with a model of how often a trait flips along a branch. To keep the numbers friendly, suppose every branch on this tree is long enough that a society keeps its state with probability 0.8 and switches with probability 0.2.

(In the standard two-state model with switching rate q, the chance of keeping the state along a branch of length t is 1/2 + 1/2 e^(-2qt), so 0.8 corresponds to e^(-2qt) = 0.6. As branches get longer, both chances drift toward 1/2. That is the model's way of saying that information about very deep ancestors fades.)

Now compute, for each possible state of a node, the probability of everything observed below it. Biologists call this bookkeeping pruning.

At N, if N was 1, both A and B kept the state:

L_N(1) = 0.8 × 0.8 = 0.64

If N was 0, both switched:

L_N(0) = 0.2 × 0.2 = 0.04

At R, combine the branch down to N (where N might have been either state) with the branch down to C, which we know ended at 0:

L_R(1) = (0.8 × 0.64 + 0.2 × 0.04) × 0.2 = 0.520 × 0.2 = 0.104

L_R(0) = (0.2 × 0.64 + 0.8 × 0.04) × 0.8 = 0.160 × 0.8 = 0.128

Give the two root states equal weight before looking at the data, and the probability that R was matrilocal is

0.104 / (0.104 + 0.128) = 0.104 / 0.232 ≈ 0.45

Two of the three societies are matrilocal, yet the root leans slightly the other way. That is the tree doing its job. A and B are close kin, so their agreement mostly tells us about N, not two separate things about R. C, alone on its branch, weighs about as much as the pair. Change C to matrilocal and the same arithmetic gives about 0.93 for a matrilocal root.

Real analyses run this over hundreds of tips, estimate the switching rates from the data instead of assuming them, and repeat the calculation across many plausible trees so that doubt about the tree becomes doubt about the answer. The arithmetic at each node is the arithmetic above.

Two traits that travel together: cattle and descent

Reconstruction is half the toolkit. The other half asks whether two traits change together along the tree, and which tends to change first.

Matrilineal descent, in which group membership passes through women, is rare among societies that keep large livestock. Clare Holden and Ruth Mace (2003) asked whether that pattern reflects a real link or only shared history. Their sample was 68 Bantu- and Bantoid-speaking populations of sub-Saharan Africa, and their scaffold was a maximum-parsimony tree of Bantu languages. They tested whether cattle-keeping and descent evolved together, and whether cattle tended to arrive before matriliny was lost. The results supported the hypothesis that acquiring cattle led formerly matrilineal societies to change to patrilineal or mixed descent.

Tests of this kind compare two models of change. In the first, each trait flips at its own rates, ignoring the other. In the second, the rate at which descent changes may depend on whether a society keeps cattle, and the rate at which cattle are gained or lost may depend on descent. If the second model fits the tree much better, the traits are evolving together, and the estimated rates show which transitions are common. If the route "gain cattle, then lose matriliny" has fast rates and the route "lose matriliny, then gain cattle" is slow, that is evidence about order, not just association.

Holden and Mace also discussed why matriliny goes with horticulture and is rare among herders, linking it to parental investment biased toward daughters and drawing on data on wealth and reproductive success from the matrilineal Chewa of Malawi and the patrilineal Gabbra of Kenya. The phylogenetic test establishes the pattern; the explanation is a further argument.

Residence after marriage in the Austronesian world

Jordan and her colleagues (2009) used Austronesian language trees covering 135 Pacific societies and estimated the type of residence at each ancestral node. Earlier scholars had proposed that proto-Oceanic society, around 3,500 years ago, was matrilocal. The phylogenetic analysis found strong evidence that matrilocality was the main pattern earlier still, around 5,000 to 4,500 years ago, at the root of the family and on its early branches. The authors noted that this helps make sense of the different patterns in mitochondrial DNA, inherited through mothers, and the Y chromosome, inherited through fathers, across the Pacific.

Laura Fortunato and Jordan (2010) then compared two families. They reconstructed Proto-Indo-European as virilocal (couples living with the husband's kin) and Proto-Malayo-Polynesian as uxorilocal (living with the wife's kin). The rates told a shared story: residence with the wife's kin was unstable, while residence with the husband's kin, once gained, was rarely lost.

Be clear about what kind of claim this is. Nobody observed an ancestral wedding. The result is a probability, conditional on the tree, on how present-day ethnographies were coded, and on the model of change. Each of those conditions is a place where a reader can push.

Where the tree breaks

Societies are not species. People marry across language boundaries, trade, migrate and copy their neighbors, and each of these moves customs sideways between branches. Biologists call this horizontal transmission; anthropologists more often say borrowing or diffusion. The copying rules in our piece on transmission biases operate between neighboring groups as well as within them.

How much does sideways movement matter? Greenhill, Thomas Currie and Gray (2009) simulated borrowing between languages and found that tree shapes inferred with Bayesian methods held up under realistic levels of it, while estimated dates were less reliable and tended to come out too young. That result concerns the language tree. Whether a particular custom moved between neighbors more easily than vocabulary does is a separate question for each study.

Some histories are not tree-shaped at all. Gray, David Bryant and Greenhill (2010) argued that "is culture tree-like?" should not be treated as a yes-or-no dispute. The shape of history (how tree-like it is) and its fabric (whether customs travel as a bundle with language) vary by degree, and the empirical job is to measure where a given trait sits. Network methods such as NeighborNet help here: where a tree would force conflicting signals into one branching pattern, a network draws the conflict as boxes you can see.

Rules of change may not travel either. Jiří Moravec and colleagues (2018) compared post-marital residence across five large families (Austronesian, Bantu, Indo-European, Pama-Nyungan and Uto-Aztecan), 371 groups in all, and found that each family had its own best-fitting model. A transition that is common in one family can be rare in another, so rates estimated in one region should not be exported to the rest of the world.

Reading one of these studies

The three-society tree above is worth keeping in your pocket for that last row. The swing from 0.45 to 0.93 came from one society on one branch. When a published reconstruction is strong, it is worth asking which branches carry it, and for the field's wider setting, our research overview places this work among the other lines of mathematical anthropology.

^† Links to books are affiliate links. As an Amazon Associate I earn from qualifying purchases.

Sources and further reading

• Mace, Ruth, and Mark Pagel. 1994. "The comparative method in anthropology." Current Anthropology 35 (5): 549-564. https://doi.org/10.1086/204317
• Holden, Clare Janaki, and Ruth Mace. 2003. "Spread of cattle led to the loss of matrilineal descent in Africa: a coevolutionary analysis." Proceedings of the Royal Society B 270 (1532): 2425-2433. https://doi.org/10.1098/rspb.2003.2535
• Gray, Russell D., Alexei J. Drummond, and Simon J. Greenhill. 2009. "Language phylogenies reveal expansion pulses and pauses in Pacific settlement." Science 323 (5913): 479-483. https://doi.org/10.1126/science.1166858
• Jordan, Fiona M., Russell D. Gray, Simon J. Greenhill, and Ruth Mace. 2009. "Matrilocal residence is ancestral in Austronesian societies." Proceedings of the Royal Society B 276 (1664): 1957-1964. https://doi.org/10.1098/rspb.2009.0088
• Fortunato, Laura, and Fiona Jordan. 2010. "Your place or mine? A phylogenetic comparative analysis of marital residence in Indo-European and Austronesian societies." Philosophical Transactions of the Royal Society B 365 (1559): 3913-3922. https://doi.org/10.1098/rstb.2010.0017
• Greenhill, Simon J., Thomas E. Currie, and Russell D. Gray. 2009. "Does horizontal transmission invalidate cultural phylogenies?" Proceedings of the Royal Society B 276 (1665): 2299-2306. https://doi.org/10.1098/rspb.2008.1944
• Gray, Russell D., David Bryant, and Simon J. Greenhill. 2010. "On the shape and fabric of human history." Philosophical Transactions of the Royal Society B 365 (1559): 3923-3933. https://doi.org/10.1098/rstb.2010.0162
• Moravec, Jiří C., Quentin Atkinson, Claire Bowern, Simon J. Greenhill, Fiona M. Jordan, Robert M. Ross, Russell Gray, Stephen Marsland, and Murray P. Cox. 2018. "Post-marital residence patterns show lineage-specific evolution." Evolution and Human Behavior 39 (6): 594-601. https://doi.org/10.1016/j.evolhumbehav.2018.06.002