Measuring complexity for hierarchical models using effective degrees of freedom

Area of Science:

• Ecological modeling
• Statistical ecology
• Quantitative biology

Background:

• Hierarchical models are crucial for ecological dynamics, incorporating fixed and random effects.
• Measuring model complexity via effective degrees of freedom (EDF) is vital for accurate model selection and understanding.
• Estimating EDF requires assessing the shrinkage of random effects towards a shared mean.

Purpose of the Study:

• To introduce and validate a method for estimating effective degrees of freedom (EDF) in ecological models.
• To demonstrate the utility of the conditional Akaike Information Criterion (cAIC) for EDF estimation.
• To showcase the application of this EDF estimation method across diverse ecological case studies.

Main Methods:

• Applied the conditional Akaike Information Criterion (cAIC) for EDF estimation.
• Utilized a finite-difference approximation to the gradient of model predictions.
• Validated the method against established Bayesian criteria.

Main Results:

• The cAIC method for EDF estimation demonstrated behavior similar to widely used Bayesian criteria.
• Case studies revealed ecological insights, such as favoring time-varying parameters in survival models and identifying differential complexity in phylogenetic and species distribution models.
• The method successfully identified species requiring greater model complexity in distribution modeling.

Conclusions:

• The proposed cAIC-based EDF estimation provides valuable ecological and statistical insights.
• Comparing EDF across experimental units, models, and data partitions enhances understanding.
• This approach is broadly applicable to nonlinear ecological models.