Estimating multiparameter partial expected value of perfect information from a probabilistic sensitivity analysis...

Area of Science:

• Decision Analysis
• Computational Statistics
• Health Economics

Background:

• Partial expected value of perfect information (EVPI) is crucial for quantifying the value of information in decision models.
• Current estimation methods, like 2-level Monte Carlo, are computationally intensive and face challenges with correlated parameters.
• Efficient estimation of EVPI is vital for informed decision-making in complex models.

Purpose of the Study:

• To introduce a novel, computationally efficient nonparametric regression-based method for estimating partial EVPI.
• To demonstrate the applicability and advantages of this new method across various model complexities and parameter distributions.
• To provide an alternative to the computationally demanding 2-level Monte Carlo procedure.

Main Methods:

• Developed a nonparametric regression approach utilizing existing probabilistic sensitivity analysis (PSA) samples.
• Implemented the method using Generalized Additive Models (GAM) and Gaussian processes.
• Compared the efficiency and performance against the traditional 2-level Monte Carlo method in case studies.

Main Results:

• The regression-based method requires only PSA samples, simplifying the estimation process.
• Demonstrated superior computational efficiency compared to the 2-level Monte Carlo method in two case studies.
• The method is versatile, applicable to models of any complexity and parameter distribution specifications.

Conclusions:

• The novel regression-based method offers a more efficient and practical approach to estimating partial EVPI.
• This method reduces computational burden and circumvents sampling difficulties associated with correlated parameters.
• Availability of R code facilitates the adoption of this advanced technique in decision modeling.