A Bayesian Penalized Hidden Markov Model for Ant Interactions

TL;DR

This paper introduces a Bayesian penalized hidden Markov model to analyze ant interactions, effectively smoothing state transitions to better reflect biological reality in social behavior data.

Contribution

It proposes a novel Bayesian ridge prior for HMMs that incorporates biological covariates, improving the biological plausibility of state transition estimates.

Findings

The model successfully captures biologically reasonable switching behavior.

Incorporation of covariates improves model interpretability.

Bayesian inference via MCMC enables robust estimation.

Abstract

Interactions between social animals provide insights into the exchange and flow of nutrients, disease, and social contacts. We consider a chamber level analysis of trophallaxis interactions between carpenter ants (\textit{Camponotus pennsylvanicus}) over 4 hours of second-by-second observations. The data show clear switches between fast and slow modes of trophallaxis. However, fitting a standard hidden Markov model (HMM) results in an estimated hidden state process that is overfit to this high resolution data, as the state process fluctuates an order of magnitude more quickly than is biologically reasonable. We propose a novel approach for penalized estimation of HMMs through a Bayesian ridge prior on the state transition rates while also incorporating biologically motivated covariates. This penalty induces smoothing, limiting the rate of state switching that combines with appropriate covariates within the colony to ensure more biologically feasible results. We develop a Markov chain Monte Carlo algorithm to perform Bayesian inference based on discretized observations of the contact network.