ľ¹Ï¸£ÀûÓ°ÊÓ

Cellphones, wearable devices, and the Internet of Things now make it possible to record social, psychological, and behavioural processes as they unfold in daily life. These technological advances have led to a surge of intensive longitudinal data (ILD), offering unprecedented opportunities to study the temporal dynamics of human behaviour, emotion, and physiology in naturalistic settings.

However, traditional statistical methods often aggregate data over time or assume gradual, homogeneous changes. Such approaches struggle to capture the abrupt transitions and complex, multimodal patterns that characterize many real-world processes, such as rapid shifts in mood, sudden changes in neural activity, or behavioural switching. As a result, critical information about the timing, frequency, and sequence of these dynamics is easily lost.

This dissertation introduces and develops multilevel Hidden Markov Models (mHMMs) as a powerful alternative for analysing ILD. Hidden Markov Models capture discrete latent states and the transitions between them, making them well suited to processes with abrupt changes. Extending this framework to a multilevel setting allows the incorporation of hierarchical structures and individual variability, which are inherent to social and behavioural data.

The dissertation has three aims: (1) to evaluate the performance of mHMMs through large-scale simulations, (2) to extend the methodological framework, including the R package mHMMbayes, and (3) to demonstrate applications across psychology, behavioural science, and neuroscience. Empirical studies highlight their value for modelling nonverbal communication, mood dynamics in bipolar disorder, animal behaviour, and neural activity. Together, these contributions establish mHMMs as an accessible and robust tool for uncovering latent dynamics in ILD.