Uncertainty, Risk & Sensitivity Analysis

Characterising what is uncertain, how much it matters, and where.

Inputs ranked by how far each one moves the answer A vertical line marks a central estimate. Six horizontal bars cross it, each reaching a different distance either side, with a marker at each end: the bars are ordered so the widest sits at the top and the shortest at the bottom. The longest bar is drawn in the accent colour, the one input the conclusion is most exposed to.

At a glance

Produces
A classification of the uncertainty present, and a measure of how far the conclusions depend on it.
Depends on
Global rather than one-at-a-time testing, and a willingness to publish where the analysis is fragile.
Fails when
Probabilities are attached to deep uncertainty so that a single number can be quoted.

This work distinguishes between the kinds of uncertainty present in a problem - measurable variability, incomplete knowledge, model uncertainty, and deep ambiguity - and tests how far conclusions depend on particular assumptions, data or modelling choices. It answers the question of how much confidence an analysis has earned.

How we approach it

Sensitivity analysis is often treated as a final check. We run it as an ongoing question: what would have to be different for this to be wrong?

  • Uncertainty is classified before it is quantified. Applying probabilities to deep uncertainty produces false precision rather than useful information.
  • Global sensitivity methods are preferred to one-at-a-time testing, because interactions between inputs are usually where the fragility lives.
  • Value of information is assessed explicitly. Some uncertainties are not worth resolving, and saying so frees resources for those that are.
  • Results are reported as conditional on assumptions. A single headline number without its sensitivity structure is treated as an incomplete result.