Penalized Estimation and Forecasting of Multiple Subject Intensive Longitudinal Data.

Area of Science:

• Social and Behavioral Sciences
• Psychometrics
• Data Science

Background:

• Intensive longitudinal data (ILD) offers rich insights into dynamic processes.
• Existing methods inadequately leverage ILD for individual-level forecasting.
• Forecasting individual dynamic processes remains a challenge in behavioral research.

Purpose of the Study:

• To present a novel methodological framework, multi-VAR, for analyzing and forecasting ILD from multiple individuals.
• To enable simultaneous estimation of dynamic models across individuals, accounting for heterogeneity.
• To improve the accuracy of individual-level dynamic process forecasting.

Main Methods:

• Developed the multi-VAR framework for penalized estimation of ILD.
• Proposed a proximal gradient descent algorithm for solving the multi-VAR problem.
• Proved the consistency of recovered transition matrices in the multi-VAR models.

Main Results:

• The multi-VAR framework successfully estimates dynamic models simultaneously for multiple individuals.
• The method adaptively adjusts to varying levels of heterogeneity in individual processes.
• Demonstrated superior forecasting performance compared to benchmark methods in an emotional experience dataset.

Conclusions:

• The multi-VAR framework provides a powerful new tool for individual-level forecasting with ILD.
• This approach enhances understanding of dynamic behavioral patterns by modeling heterogeneity.
• Future research can apply multi-VAR to diverse fields utilizing intensive longitudinal data.