Using State-Space Model with Regime Switching to Represent the Dynamics of Facial EMG Data

Facial electromyography (EMG) is a useful physiological measure for detecting subtle affective changes in real time. It can differentiate valence of emotion, capture transient and covert affective response, and provide an approximate continuous-time measure of emotion change. In this thesis, facial EMG data is analyzed using time series analysis methods. By allowing certain parameters to switch between several discrete stages (regimes), regime-switching models can be used to describe heterogeneous transition patterns and capture time-varying association between EMG signals and other covariates. The main purpose of this thesis is to construct and propose different regime-switching state-space models suited for representing the time-varying dynamics of facial EMG data and its relationship with self-reported affect intensity. The Kim filter, which is an extension of the Kalman filter, is proposed to estimate latent states and the Gaussian maximum likelihood method is used for parameter estimation. Indices for diagnostic checks and model fit evaluations will be discussed for model comparison purposes. Results based on empirical EMG data indicate that regime-switching model with autoregressive regression slope (the RS-AR model) is most appropriate to represent the EMG dynamics. Monte Carlo simulation results also indicate that parameters in the proposed model can be accurately recovered under different conditions.