Probabilistic sensitivity analysis: be a Bayesian.

Area of Science:

• Decision Analysis
• Biostatistics
• Health Economics

Background:

• Probabilistic sensitivity analysis (PSA) is crucial for evaluating uncertainty in model inputs.
• Defining appropriate probability distributions for model inputs is essential for robust PSA.
• Commonly, input distributions are based on data likelihood, but this can lead to issues with sparse data.

Purpose of the Study:

• To provide guidance on defining probability distributions for model inputs in PSA from a Bayesian perspective.
• To critically evaluate the common approach of using data likelihood for input distributions.
• To propose more appropriate prior distributions for PSA inputs.

Main Methods:

• Examined the common approach of using data likelihood for input distributions from a Bayesian viewpoint.
• Derived implicit prior distributions in examples involving proportions and relative risks.
• Compared these implicit priors with alternative prior distributions.

Main Results:

• Commonly used methods can yield unexpected results, especially with sparse data.
• Implicit prior distributions in standard approaches are often not as uninformative as assumed.
• More sensible and easily applicable prior distributions were proposed for specific examples.

Conclusions:

• Input probability distributions for PSA should not solely rely on data likelihood.
• Distributions should be based on the Bayesian posterior distribution.
• Explicitly stating prior distributions is critical for transparent and reliable PSA.