Comparing partition and mixture models with akaike information criteria.

Area of Science:

• Computational Biology
• Phylogenetics
• Statistical Modeling

Background:

• Phylogenetic models frequently incorporate mixture and partition components.
• Information criteria may incorrectly favor partition models over mixture models, even when mixture models are misspecified.

Purpose of the Study:

• To investigate why information criteria favor partition models over mixture models in phylogenetic analysis.
• To explain the fundamental differences in probability calculations between mixture and partition models.
• To propose corrections for non-comparable AIC estimates.

Main Methods:

• Comparative analysis of mixture and partition model probability calculations.
• Examination of how differing calculations affect Akaike Information Criterion (AIC) estimates.
• Exploration of generalizable methods to rectify the issue.

Main Results:

• Partition and mixture models differ in calculating likelihood contributions: mixture models use marginal probabilities, while partition models use conditional probabilities.
• These distinct calculation methods lead to non-comparable AIC estimates.
• The study identifies and explores three generalizable correction strategies.

Conclusions:

• The discrepancy in probability calculations is the root cause of information criteria favoring partition models.
• Addressing these calculation differences is crucial for accurate model selection in phylogenetics.
• Proposed corrections aim to improve the reliability of phylogenetic model evaluation.