Sensitivity Analysis

In decision scoring, sensitivity analysis examines whether plausible changes in the weights, the scoring anchors, or the measured inputs would change the ranking of options. The UK multi-criteria analysis manual lists it as the final step of a multi-criteria analysis, after the weights and scores have been combined and the results examined. Its output is not a new winner but a statement of how robust the base case is and what would overturn it.

Vary the judgments and the inputs separately

Use illustrative 0–100 scores for handling time, cost, and administration: A=(80,40,60) and B=(60,80,80). Under weights of 50/30/20, A totals 0.5 × 80 + 0.3 × 40 + 0.2 × 60 = 64 and B totals 0.5 × 60 + 0.3 × 80 + 0.2 × 80 = 70. Moving 20 percentage points from cost to time, to 70/10/20, gives A 72 and B 66: the order reverses. The boundary can be located exactly: at 60/20/20 both score 68, and any larger shift toward time favors A. With administration held at 20%, the choice hinges on whether handling time deserves more than 60% of the weight: at exactly 60% the options tie, and only above it does A lead.

For administration, use a linear scale from 0 points at 20 hours to 100 points at 0 hours: score = 100 − 5 × hours. B’s base score of 80 therefore represents 4 hours. If B’s administration effort is an estimate with a range of 3 to 6 hours, its total moves between 71 and 68 and stays above A’s 64, so that uncertainty does not matter for the order. If B’s quoted cost turns out to exclude a required connector, the corrected cost is entered and the totals recomputed. NASA’s decision-analysis guidance gives a ranking-focused test: within a given analysis, uncertainty matters only if reducing it could credibly change the ranking of the alternatives. That test concerns the order of options; uncertainty that leaves the order unchanged can still matter for other purposes, such as whether a mandatory gate is met.

What it is not

Sensitivity analysis is not permission to adjust weights after the results until a preferred option wins. The approved base case stays fixed and visible; scenarios are recorded beside it with the reason each was tried. It does not make a failed or unverified gate negotiable, and it does not convert a score difference into a probability of success. When a result is close or unstable, the useful next step is to name the evidence gap that would settle it, such as a narrowly scoped integration trial, and close it.

The same words are used elsewhere with different meanings: in machine learning, sensitivity can mean a model’s dependence on how features are represented, and in diagnostics it is the true-positive rate. Here it is a robustness check on a decision.

References

• UK Government: Multi-criteria analysis manual
• NASA: Decision analysis

Figures here are illustrative.