The AIC’s Blind Spot: Why Comparing Phylogenetic Models Is Harder Than It Looks
A new study in *Systematic Biology* reveals a critical flaw in using standard information criteria like the Akaike Information Criterion (AIC) to compare complex phylogenetic models. Researchers demonstrate that partition models and mixture models, which are fundamental to modern phylogenetics, calculate likelihoods in fundamentally different ways. Mixture models average over possible parameter vectors for each site, while partition models condition on a fixed assignment. This statistical mismatch means AIC scores for these models are not directly comparable, potentially leading researchers to favor a poorly performing partition model over a correctly specified mixture model. The paper proposes three general methods to correct this issue, aiming to improve the reliability of model selection in evolutionary tree inference.
Why it might matter to you: For any researcher using model selection to infer evolutionary relationships or test hypotheses about molecular evolution, this work is a crucial methodological alert. It directly impacts the accuracy of your phylogenetic analyses and the robustness of conclusions about speciation, common ancestry, and evolutionary rates. Adopting the proposed corrections can help ensure your chosen model truly reflects the underlying evolutionary processes, strengthening the foundation of comparative genomics and macroevolutionary studies.
Source →Stay curious. Stay informed — with Science Briefing.
Always double check the original article for accuracy.