Nonparametric inference in hidden Markov models using P-splines.

Area of Science:

• Statistics
• Ecology
• Bioacoustics

Background:

• Hidden Markov models (HMMs) are statistical tools for analyzing time series data with underlying unobserved states.
• Parametric assumptions for state-dependent distributions in HMMs can lead to model complexity and hinder interpretation.
• Accurate modeling is crucial for inferring ecological behaviors from movement data.

Purpose of the Study:

• To investigate the practical challenges of parametric Hidden Markov Models in ecological time series analysis.
• To introduce and evaluate a nonparametric approach for estimating state-dependent distributions in HMMs.
• To improve the parsimony and interpretability of HMMs for biological inference.

Main Methods:

• Application of Hidden Markov Models (HMMs) to vertical dive speed data of beaked whales.
• Comparison of parametric HMMs with a novel nonparametric estimation method.
• Nonparametric estimation using B-spline basis functions with a smoothness penalty.

Main Results:

• Parametric HMMs resulted in overly complex state processes, impeding biological interpretation of whale dive data.
• Nonparametrically estimated HMMs achieved comparable data fit with fewer states and enhanced interpretability.
• The nonparametric approach demonstrated greater parsimony and facilitated more meaningful biological inference.

Conclusions:

• Nonparametric estimation of state-dependent distributions in HMMs provides a more flexible and interpretable alternative to parametric approaches.
• This method enhances the utility of HMMs for ecological studies, particularly in analyzing animal movement and behavior.
• The B-spline based nonparametric HMM offers a robust framework for balancing model fit and smoothness in complex time series data.