Metamodeling for Policy Simulations with Multivariate Outcomes.

Area of Science:

• Health policy analysis
• Computational modeling
• Statistical learning

Background:

• Metamodels approximate complex models but face challenges with multiple correlated outputs in policy analysis.
• Developing effective metamodeling strategies is crucial for accurate policy simulations with multivariate outcomes.

Purpose of the Study:

• To develop and evaluate a framework for metamodeling policy simulations with multivariate outcomes.
• To accommodate correlated outputs in policy analysis using advanced algorithm adaptation techniques.

Main Methods:

• Combined multitarget stacking and regression chain with base learners (LR, EE, GPR, RFs, neural networks).
• Optimized models using variable selection and hyperparameter tuning.
• Applied the framework to emulate a hepatitis C testing and treatment microsimulation model.

Main Results:

• Output variables were significantly correlated (average ρ = 0.58).
• Multioutput adaptation and optimization methods modestly improved model fit (R²).
• Simpler models (LR, EE, RF) offered efficiency and interpretability advantages.

Conclusions:

• Base learner choice significantly impacts metamodel accuracy in policy analysis.
• The proposed framework is broadly applicable to health and medical decision analysis with multivariate outcomes.