Estimating individualized treatment effects using an individual participant data meta-analysis.

Area of Science:

• Biostatistics
• Personalized Medicine
• Epidemiology

Background:

• Identifying individuals who benefit from specific interventions is key in personalized medicine.
• Estimating individualized treatment effects (ITE) often requires large datasets, as single trials are underpowered.
• Individual participant data meta-analyses (IPD-MA) offer a solution for robust ITE estimation, but combining them with prediction models is underexplored.

Approach:

• Compared five one-stage models (naive, random intercept, stratified intercept, rank-1, fully stratified) using S-learner and T-learner strategies.
• Evaluated model performance using c-statistic for benefit, prediction calibration, and mean squared error.
• Utilized simulated data and the INDANA IPD-MA with binary and time-to-event outcomes for a real-world comparison.

Key Points:

• The S-learner strategy demonstrated superior performance for ITE estimation across both binary and time-to-event outcomes in simulations.
• No single risk prediction model consistently outperformed others across all scenarios.
• For the INDANA dataset (binary outcome), naive and random intercept models showed the best performance.

Conclusions:

• The S-learner strategy, incorporating treatment interactions, is recommended for improved ITE estimation.
• No specific risk prediction method demonstrated a significant advantage over others in this IPD-MA context.