Modeling Behavioral Dynamics in a Complex World: From Informative Missingness to Predictive Forecasting

Digital technologies—smartphones and wearable devices—have transformed how scientists observe human behavior, enabling the study of fine-grained fluctuations in daily life. However, intensive longitudinal data (ILD) collected in real-world contexts introduce fundamental methodological challenges: systematic missingness, intricate temporal dependencies, and substantial individual heterogeneity. Addressing these challenges is essential for drawing valid inference and for translating high-frequency behavioral data into actionable insight.

In this talk, I present research organized around three complementary aims. First, I introduce a framework that treats missing data in ILD not as a random artifact, but as an informative, time-varying behavioral process, allowing for joint inference on participant compliance and substantive outcomes. Second, I illustrate how dynamic social processes—such as couple synchrony and co-regulation—can be modeled using multilevel and time-series approaches applied to daily diary and wearable data. Third, I present recent work on behavioral health forecasting, showing how deep learning models can complement statistical models to forecast behavioral change before it occurs.

Together, these projects demonstrate how advanced statistical models and modern computational methods can help us better understand and forecast behavior in a complex world, with implications for personalized intervention and prevention in behavioral and health research.