Dynamic prediction by landmarking with data from cohort subsampling designs

Area of Science:

• Biostatistics
• Epidemiology
• Health Informatics

Background:

• Longitudinal data from cohort studies and electronic health records can enhance health event prediction.
• Landmarking is a key approach for dynamic prediction using such data.
• Full cohort data collection is often resource-intensive, necessitating alternative strategies.

Purpose of the Study:

• To develop and evaluate statistical methods for dynamic prediction using subsampled cohort data.
• To adapt landmarking techniques for efficient analysis of limited data.
• To compare the performance of new methods against traditional full cohort analysis.

Main Methods:

• Conditional likelihood and inverse-probability weighting for landmarking with subsampled data.
• Simulation studies to assess method applicability and predictive performance.
• Application to nested case-control data from the Prostate, Lung, Colorectal and Ovarian (PLCO) Cancer Screening Trial.

Main Results:

• Developed methods provide accurate dynamic prediction using only a fraction of the full cohort data.
• The proposed techniques achieve predictive performance comparable to full cohort analyses.
• Demonstrated the utility of these methods on real-world clinical trial data.

Conclusions:

• Subsampling designs combined with novel statistical methods offer an efficient alternative for dynamic prediction in cohort studies.
• These approaches reduce data collection burdens without compromising predictive accuracy.
• The methods are applicable in settings with limited resources, enhancing the feasibility of longitudinal data analysis.