Predicting dropout in intensive longitudinal data: Extending the joint model for autocorrelated data.

Area of Science:

• Clinical Research Methods
• Biostatistics
• Psychometrics

Background:

• Ecological momentary assessment (EMA) generates intensive longitudinal data valuable in clinical research.
• Participant dropout in EMA studies compromises statistical power, introduces bias, and negatively impacts treatment outcomes.
• Existing dropout prediction methods often overlook temporal data dynamics and precise dropout timing.

Purpose of the Study:

• To develop and validate an extended joint model (JM) for predicting dropout in EMA studies.
• To incorporate autoregressive components into JMs to capture temporal dependencies in EMA data.
• To dynamically update dropout risk predictions using both baseline and time-varying covariates.

Main Methods:

• Extended standard JM by adding an autoregressive submodel to account for EMA data autocorrelation.
• Validated the extended JM using simulation studies under various missingness mechanisms (MCAR, MAR, MNAR).
• Applied the extended JM to an empirical EMA dataset, analyzing baseline and time-varying predictors of dropout.

Main Results:

• The extended JM demonstrated good parameter recovery in simulations across different missingness scenarios.
• High accuracy was achieved in predicting dropout, outperforming a baseline-only survival model in the empirical data analysis.
• Sensitivity analysis indicated stable fixed-effect estimates but sensitive random-effect estimates for autocorrelation depending on missingness assumptions.

Conclusions:

• The extended JM effectively integrates temporal dependencies from EMA data for improved dropout prediction.
• This approach enhances the utility of JMs in clinical research for predicting outcomes and managing EMA data.
• Accurate dropout prediction is vital for mitigating negative impacts on statistical power and treatment efficacy in longitudinal studies.