LoRaD: Marginal likelihood estimation with haste (but no waste)

Area of Science:

• Computational Biology
• Statistical Modeling
• Evolutionary Biology

Background:

• Bayesian model selection relies on accurate marginal likelihood estimation.
• Current methods like Steppingstone/Thermodynamic Integration require complex power posterior sampling.
• Faster methods like Generalized Harmonic Mean (GHM) may lack accuracy.

Purpose of the Study:

• Introduce the Lowest Radial Distance (LoRaD) method for marginal likelihood estimation.
• Compare LoRaD's performance against existing Bayesian phylogenetics methods.
• Assess LoRaD's applicability in both fixed and variable tree topology analyses.

Main Methods:

• LoRaD modifies the Partition-Weighted Kernel method.
• It requires sampling only from the posterior distribution.
• The method was tested on a fixed-topology molecular phylogenetics example with 180 parameters.

Main Results:

• LoRaD demonstrates improved accuracy compared to Generalized Harmonic Mean (GHM).
• Performance is comparable to the Generalized Steppingstone method in fixed-topology analyses.
• Accurate marginal likelihood estimation is possible in variable-topology cases under specific conditions.

Conclusions:

• LoRaD offers a more efficient and accurate approach to marginal likelihood estimation in Bayesian phylogenetics.
• The method simplifies computational requirements by avoiding power posterior sampling.
• LoRaD shows promise for both fixed and variable tree topology analyses in evolutionary studies.