Improving the accuracy of demographic and molecular clock model comparison while accommodating phylogenetic...

Area of Science:

• Bayesian phylogenetics and molecular evolution
• Computational evolutionary biology
• Statistical model selection

Background:

• The harmonic mean estimator (HME) for marginal likelihood estimation in Bayesian phylogenetics is known to perform poorly.
• Newer methods like path sampling (PS) and stepping-stone (SS) sampling show promise but have not been widely adopted due to computational demands and lack of implementation in common software.
• The performance of these advanced methods on complex, real-world evolutionary and population genetic models remains largely unknown.

Purpose of the Study:

• To investigate the performance and computational feasibility of PS and SS sampling for complex phylogenetic models.
• To compare PS and SS sampling against HME and a posterior simulation-based analogue of Akaike's information criterion (AICM) using synthetic and real-world data.
• To re-evaluate previous phylogenetic analyses that relied on HME, particularly for models of demographic change and relaxed molecular clocks.

Main Methods:

• Application of path sampling (PS) and stepping-stone (SS) sampling to compare models of demographic change and relaxed molecular clocks.
• Utilized synthetic datasets and three real-world phylogenetic datasets where HME yielded unexpected results.
• Evaluated a posterior simulation-based analogue of Akaike's information criterion through Markov chain Monte Carlo (AICM) as a computationally less intensive alternative.

Main Results:

• Confirmed that the harmonic mean estimator (HME) systematically overestimates marginal likelihoods and leads to unreliable model selection.
• AICM showed improvement over HME but remained partially unreliable for model choice.
• Path sampling (PS) and stepping-stone (SS) sampling substantially outperformed HME and AICM, leading to adjusted conclusions for the reanalyzed real-world datasets.

Conclusions:

• PS and SS sampling are reliable and superior methods for marginal likelihood estimation and model selection in complex Bayesian phylogenetic analyses.
• Previous inferences based on HME for demographic and relaxed clock models may require revision.
• The investigated methods, including PS and SS sampling, are now implemented in the BEAST software package, facilitating their wider application.