Within- and Between-Individual Compliance in Mobile Health: Joint Modeling Approach to Nonrandom Missingness in an...

Area of Science:

• Health Informatics
• Biostatistics
• Behavioral Science

Background:

• Missing data are common in mobile health (mHealth) and ubiquitous health (uHealth) research.
• Nonresponse is influenced by within- and between-person factors, complicating data analysis.
• Existing methods often fail to distinguish these factors, especially for non-random missingness.

Purpose of the Study:

• Demonstrate joint modeling for mHealth/uHealth data analysis.
• Show how accounting for behavior dynamics and missingness improves validity of health behavior inferences.
• Illustrate joint modeling for non-ignorable missingness in ecological momentary assessment and wearable device studies.

Main Methods:

• Applied joint modeling to 1 year of daily smartphone-based ecological momentary assessment (affect, energy) and smartwatch-tracked physical activity (PA) data.
• Combined multilevel vector autoregressive models for behavior dynamics and multilevel probit models for missingness.
• Handled missingness during model fitting, unlike traditional imputation methods, and conducted sensitivity analyses and simulations.

Main Results:

• Joint modeling detected cross-regressive effects missed by other methods; higher energy predicted higher PA the next day.
• Revealed both missing not at random (lower PA predicted PA missingness) and missing at random (employment status predicted device-PA missingness) mechanisms.
• Simulations confirmed joint modeling improves estimate accuracy and identifies non-ignorable missingness.

Conclusions:

• Recommend joint modeling with multilevel decomposition for non-ignorable missingness in mHealth/uHealth intensive longitudinal data.
• Advocate for using missing data models to understand mechanisms and guide data collection.
• Emphasize the importance of accounting for distinct missingness sources for robust health behavior research.