Two Filtering Methods of Forecasting Linear and Nonlinear Dynamics of Intensive Longitudinal Data

Area of Science:

• Behavioral Science
• Data Science
• Computational Statistics

Background:

• Intensive longitudinal data (ILD) collection is increasing due to new technologies.
• Existing analytical methods for ILD are not yet widely applied for behavioral forecasting.
• Forecasting applications for behavioral ILD remain scarce.

Purpose of the Study:

• To establish a general modeling framework for ILD.
• To extend this framework to Kalman and ensemble forecasting methods.
• To apply and compare these forecasting methods to daily drug and alcohol use data.

Main Methods:

• Developed a general framework for modeling intensive longitudinal data (ILD).
• Extended the framework to implement Kalman and ensemble forecasting methods.
• Applied these methods to daily drug and alcohol use data, creating a nonlinear dynamical system model.

Main Results:

• Implemented Kalman and ensemble forecasting in open-source software.
• Illustrated key differences between Kalman and ensemble forecasting methods using behavioral data.
• Compared these methods against simpler forecasting approaches.

Conclusions:

• Ensemble forecasts may be more suitable than Kalman forecasts for nonlinear dynamical systems in behavioral research.
• Further development and application of forecasting evaluation methods are necessary for ILD.
• This work provides a foundation for improved short- and long-term predictions from behavioral ILD.