Partitioning morphological phylogenetic matrices by character-wise likelihood: a new criterion that improves model fit and the reliability of inference
Morphological characters are heterogeneous in their rates of change and levels of homoplasy, and partitioning the matrix to accommodate this can affect estimates of branch lengths and divergence times. Here I introduce a likelihood-based (LnL) criterion for partitioning morphological data, in which characters are grouped by their character-wise log-likelihood under the best-fitting model, with partitions defined through rank-based binning (equal-frequency n-tiles and sample quantiles). I compare it against homoplasy-based criteria (extra steps, the retention index, and the relative homoplasy index), a character-type (binary versus multistate) criterion, statistical clustering (EvoPhylo), and unpartitioned analysis across eight non-simulated datasets (five empirical, two domesticated, and one artificial) and twelve simulated ones, evaluating model fit, Bayes factors, and the accuracy, precision, and coverage of branch-length and divergence-time estimates. The LnL criterion was the best-fitting, selected under the AIC, AICc, and BIC and, decisively, the non-circular stepping-stone Bayes factor in most datasets, while the unpartitioned model never won. Its trees were the most conservative, resolving fewer clades but making the fewest confident false claims, and no further from the reference tree (nRF) than any other criterion. The choice of criterion had no significant effect on the accuracy or precision of divergence times and affected branch lengths only in the smallest matrices. Likelihood-based partitioning improves model fit and yields more conservative, better-supported trees, and is simple to apply, needing only a single IQ-TREE 3 run with the -wsl option followed by binning of the character log-likelihoods. I recommend it as an accessible default for morphological phylogenetics.
Authors
- Felipe Barbosa
Institutions
- Museu Paraense Emílio Goeldi (BR)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-21
- DOI
- https://doi.org/10.5281/zenodo.24408739
- Primary Topic
- Genomics and Phylogenetic Studies
- Type
- preprint