Related Experiment Videos
Methods for handling longitudinal outcome processes truncated by dropout and death.
Lan Wen1, Graciela Muniz Terrera2, Shaun R Seaman1
1MRC Biostatistics Unit, University of Cambridge, IPH Forvie Site, Robinson Way, Cambridge, UK.
Biostatistics (Oxford, England)
|October 14, 2017
Summary
This study compares methods like multiple imputation (MI) and inverse probability weighting (IPW) for handling missing data in aging cohort studies. It introduces new augmented IPW estimators for more meaningful, partly conditional inference.
Area of Science:
- Epidemiology
- Biostatistics
- Longitudinal Studies
Background:
- Cohort studies in aging often face incomplete data due to subject dropout or death.
- Standard missing data methods may lead to inference about hypothetical 'immortal cohorts', which may not reflect real-world aging processes.
- Partly conditional inference, focusing on the living cohort, offers a potentially more meaningful approach for aging studies.
Purpose of the Study:
- To clarify and compare the assumptions of multiple imputation (MI), linear increments (LI), and inverse probability weighting (IPW) for partly conditional inference on continuous outcomes in aging cohort studies.
- To propose novel augmented IPW estimators that enhance efficiency and robustness for partly conditional inference.
- To evaluate the performance of these missing data methods through simulation studies and real-world data application.
Main Methods:
- Adaptation of existing MI, LI, and IPW methods for partly conditional inference.
- Development and application of augmented IPW estimators.
- Simulation studies to assess bias and efficiency under various missing data scenarios.
- Illustration using data from the 'Origins of Variance in the Old-old' Twin study.
Main Results:
- The study clarifies the distinct assumptions required by MI, LI, and IPW methods for valid partly conditional inference.
- Augmented IPW estimators demonstrated improved efficiency and robustness to model misspecification compared to standard IPW.
- Simulation results indicated that the methods provide approximately unbiased estimates when assumptions are met, but can be biased if violated.
Conclusions:
- MI, LI, and augmented IPW are viable methods for partly conditional inference in aging cohort studies with missing data.
- The choice of method depends on the specific study assumptions and data characteristics.
- Careful consideration of missing data mechanisms and method assumptions is crucial for valid inference in longitudinal aging research.