Dated ancestral trees from binary trait data and its application to the diversification of languages

Binary trait data record the presence or absence of distinguishing traits in individuals. We treat the problem of estimating ancestral trees with time depth from binary trait data. Simple analysis of such data is problematic. Each homology class of traits has a unique birth event on the tree, and the birth event of a trait visible at the leaves is biased towards the leaves. We propose a model-based analysis of such data, and present an MCMC algorithm that can sample from the resulting posterior distribution. Our model is based on using a birth-death process for the evolution of the elements of sets of traits. Our analysis correctly accounts for the removal of singleton traits, which are commonly discarded in real data sets. We illustrate Bayesian inference for two binary-trait data sets which arise in historical linguistics. The Bayesian approach allows for the incorporation of information from ancestral languages. The marginal prior distribution of the root time is uniform. We present a thorough analysis of the robustness of our results to model mispecification, through analysis of predictive distributions for external data, and fitting data simulated under alternative observation models. The reconstructed ages of tree nodes are relatively robust, whilst posterior probabilities for topology are not reliable.

💡 Research Summary

The paper addresses the challenging problem of reconstructing dated ancestral trees from binary trait data—datasets that record the presence (1) or absence (0) of discrete features across a set of taxa, such as languages or species. Traditional approaches often ignore two crucial aspects: (1) each trait originates from a single birth event somewhere on the tree, and (2) the observed traits are biased toward the tips because traits that arose deep in the tree are more likely to have been lost before reaching the present. Moreover, empirical datasets routinely discard “singleton” traits—those observed in only one taxon—further skewing the observable pattern.

To overcome these limitations, the authors develop a fully probabilistic, model‑based framework. They treat the evolution of the set of traits as a birth‑death process operating on the tree: each trait is born at a single point, then may die (be lost) independently along each lineage. The model explicitly incorporates the observation process that removes singletons, thereby correcting the bias that would otherwise pull the inferred root toward the leaves. A uniform prior is placed on the root age, ensuring that the posterior distribution of the tree’s time depth is driven primarily by the data rather than by arbitrary prior assumptions.

Inference is performed within a Bayesian paradigm using a Metropolis‑Hastings Markov chain Monte Carlo (MCMC) algorithm. The sampler jointly updates tree topology, node ages, and the birth‑death parameters governing trait dynamics. Crucially, the proposal mechanisms respect the combinatorial constraints imposed by the binary data and the singleton‑removal process, allowing efficient exploration of the high‑dimensional posterior space.

The methodology is applied to two historical‑linguistics case studies: one involving Indo‑European languages and another involving Austronesian languages. Both datasets consist of binary lexical or phonological features after the standard preprocessing step of singleton removal. The analysis yields posterior distributions for node ages and for the tree topology. The authors find that estimated ages of internal nodes (i.e., the timing of common ancestors) are relatively robust across a range of model misspecifications and alternative observation models. In contrast, posterior probabilities for specific topological arrangements are low and highly sensitive to the details of the observation model, indicating that the binary data alone provide limited information for resolving the exact branching order.

To assess robustness, the authors conduct posterior predictive checks and simulate data under alternative observation schemes (e.g., models where traits evolve continuously rather than as discrete births and deaths). Predictive distributions for external validation data (such as known dates of ancient languages) confirm that age estimates remain accurate, while the ability to recover the true topology deteriorates markedly under misspecified observation models. This pattern underscores the asymmetry between temporal and topological information content in binary trait matrices.

The paper concludes that while binary trait data, when analyzed with a properly calibrated birth‑death model and an explicit treatment of singleton removal, can yield reliable estimates of divergence times, they are insufficient for confident reconstruction of tree topology. The authors recommend augmenting the Bayesian framework with additional sources of information—such as ancient language attestations, geographic constraints, or morphological character matrices—to improve topological inference. They also highlight the broader applicability of their approach to other fields where binary presence/absence data are common, including paleobiology (fossil trait matrices) and epidemiology (presence/absence of genetic markers in pathogen lineages).

Overall, the study provides a rigorous statistical foundation for dating phylogenies from binary traits, demonstrates the importance of modeling observation bias, and offers practical guidance on the limits of what binary data can reveal about evolutionary history.