Stochastic Sensitivity Analysis — Quantifying Output Uncertainty via Probabilistic Input Sampling

Deterministic analysis varies one parameter at a time while holding others fixed, producing a tornado diagram. Stochastic SA varies all parameters simultaneously according to their distributions, yielding a full output distribution and interaction-aware sensitivity indices — far more informative when parameters are correlated or multiple thresholds are at risk.

For stable mean and median estimates, N = 1,000 is usually sufficient. For reliable tail probabilities (e.g., 95th percentile) or Sobol indices, N = 10,000 or more is recommended. Latin Hypercube Sampling can achieve similar precision at roughly one-fifth the draws of pure random sampling.

EVPI is a decision-theoretic output of PSA that quantifies the maximum worth of eliminating all parameter uncertainty. It is computed directly from the PSA output distribution as the difference between the expected net benefit under perfect information and under the current decision. A high EVPI signals that further research is economically warranted.

No. PSA only propagates uncertainty in the parameters of the chosen model structure. Structural uncertainty — whether the model itself is the right representation of reality — requires separate analysis such as model comparison, cross-validation, or expert elicitation about competing structures.

Spearman or partial rank correlation coefficients are quick to compute and work well for monotonic models. Sobol variance-based indices are model-free and capture non-linear and interaction effects, making them preferable for complex or non-monotonic models, at the cost of requiring larger N.