student.uva.nl
What is your study programme?
UvA Logo
What is your study programme?
Colloquium credits

Presentation Master's thesis - Zeno Glastra van Loon - Brain & Cognition

Colloquium credits

Presentation Master's thesis - Zeno Glastra van Loon - Brain & Cognition

Last modified on 08-05-2025 15:42
Emotions in Motion: Hidden Markov Models Outperform Vector Autoregressive Models when Predicting Emotions from Intensive Longitudinal Data
Show information for your study programme
What is your study programme?
or
event-summary.start-date
16-06-2025 10:30
event-summary.end-date
16-06-2025 11:30
event-summary.location

Roeterseilandcampus, Gebouw: G, Straat: Nieuwe Achtergracht 129-B, ruimte GS.08

Due to limited room capacity, attendance is on a first-come, first-served basis. Teachers must adhere to this.

Psychological research increasingly prioritizes within-person processes over between-person differences. These processes are often explored by analysing intensive longitudinal data with Vector Autoregressive (VAR) models. Hidden Markov Models (HMMs), which allow for latent state-switching dynamics, may offer a more appropriate alternative. HMMs might be especially suited for dynamic phenomena like emotions. This study empirically compares the predictive performance of VAR and HMMs across 39 open-access datasets. Model performance was assessed by out-of-sample Root Mean Square Error (RMSE). We found that HMMs generally outperformed VAR models. Across all studies, HMMs yielded lower RMSEs for 78.5% of participants. The average RMSE for HMMs was 0.130 (SD = 0.058), compared to 0.158 (SD = 0.086) for VAR models, with a mean ΔRMSE of 0.029 (SD = 0.056), indicating that VAR models had 21.5% higher prediction errors. We investigated the influence of multiple study and participant characteristics on the ΔRMSE using multilevel models. This showed that a higher number of variables was associated with a better fit for HMMs while more observations per participant was associated with a better fit for VAR models. Taken together, our findings suggest that HMMs often provide superior out-of-sample prediction accuracy compared to VAR models.