Probabilistic Risk Assessment (PRA)

Kaplan and Garrick defined risk as the answer to three questions — what can go wrong, how likely is it, and what are the consequences — formalized as a set of triplets, each pairing a scenario with its frequency and its consequence. This is central because it says a complete risk description is the whole set of such triplets, not a single average number, and it built uncertainty into the definition from the start. Everything PRA does is aimed at enumerating these triplets and quantifying them, and the familiar risk-exceedance curve is just the triplet set viewed across consequence magnitude.

HAZOP and similar methods identify hazards and deviations qualitatively but do not compute how likely or how severe they are. PRA goes further: it quantifies scenario frequencies using linked event trees and fault trees, models consequences, and propagates uncertainty to produce a numerical risk with confidence bounds. The two are complementary — qualitative hazard identification often supplies the initiating events and scenarios that a PRA then quantifies. PRA is reserved for the high-consequence systems where the cost of the detailed quantitative modeling is justified by the stakes of the decision.

Because the failure rates and event frequencies that drive the result are estimated from limited data and models, not known exactly, so a single risk number without error bars would be misleading. Following Kaplan and Garrick and the NASA guide, PRA represents each uncertain input as a probability distribution and propagates these — usually by Monte Carlo sampling — so the output risk is itself a distribution reported with means and percentiles. This makes the confidence in the estimate explicit and reveals which uncertain inputs dominate the spread, which is exactly the information a decision-maker needs to know whether to act or to gather more data.