HOW INTELLIGENCE WORKS · SENSITIVITY ANALYSIS · eduKateSG
How a Mind Finds Which Assumptions Matter Most to the Outcome
Sensitivity analysis is the intelligence process that deliberately varies assumptions, inputs or model choices to discover which changes meaningfully alter the conclusion and which leave it essentially intact.
Working model → uncertain assumptions → vary one or several inputs → observe output movement → rank influence → inspect interactions → protect critical assumptions → simplify low-impact ones.
This article belongs to the How Intelligence Works series. Evidence Weighting asks how strongly evidence should change belief; Trade-off Reasoning compares competing objectives. Sensitivity analysis owns a different question: which assumptions are actually carrying the conclusion?
The Same Answer May Rest on Unequal Assumptions
Two assumptions can appear side by side inside a model while playing very different roles. Change one slightly and nothing important happens. Change the other slightly and the decision reverses.
Sensitivity analysis turns that hidden dependency into something inspectable.
A conclusion is fragile when a small unseen change can move it a long way.
1. Start With the Uncertain Inputs
Sensitivity analysis begins by naming quantities, assumptions, classifications or relationships that are not perfectly known. The goal is not to distrust everything equally, but to identify plausible variation.
An input that is both uncertain and influential deserves more attention than one that is uncertain but inconsequential.
2. Vary the Input and Watch the Output
The simplest analysis changes one input while holding the rest stable enough to reveal direction and magnitude.
| Response | Interpretation |
|---|---|
| Little movement | The conclusion is relatively insensitive to that input |
| Large smooth movement | The input strongly influences the result |
| Threshold jump | A decision boundary lies nearby |
| Direction reversal | The relationship may be nonlinear or conditional |
| No stable pattern | Interactions or model instability require deeper inspection |
3. Sensitivity and Uncertainty Are Different
Uncertainty describes what is not known precisely. Sensitivity describes how much the outcome depends on that uncertainty.
A very uncertain variable can be low sensitivity. A tightly estimated variable can be high sensitivity if even a small change moves the decision across a boundary.
Uncertainty asks how wide the input could be. Sensitivity asks how far the answer moves when the input moves.
4. Interactions Can Hide the Real Driver
Inputs may matter only in combination. Temperature may matter more when pressure is high. Practice time may matter differently depending on sleep, prior knowledge or feedback quality.
Strong sensitivity reasoning therefore moves beyond one-at-a-time variation when interactions are plausible.
5. Sensitivity Analysis in Mathematics
Mathematical modelling becomes more informative when a solver asks how assumptions and parameters affect the result rather than reporting one number as though it were inevitable.
The canonical modelling route at How Mathematical Modelling Works owns the full reality-to-model process. Sensitivity analysis isolates the stress test inside that process.
6. Sensitivity Analysis in Learning
A learner’s performance may depend strongly on one weak prerequisite and only weakly on several visible surface features.
Teachers can vary support deliberately: remove the vocabulary difficulty, change time pressure, provide a representation, or isolate a prerequisite. The performance movement reveals which condition carries the most leverage.
This turns diagnosis into controlled comparison rather than impression.
7. Sensitivity Analysis in Planning
Plans often contain assumptions about time, demand, cost, availability and behaviour. Testing which assumption most threatens the plan helps allocate monitoring and contingency effort.
The most fragile assumption should not remain hidden in a footnote while low-impact details receive most of the discussion.
8. Sensitivity Helps Decide What to Measure Better
Once influence is known, information gathering becomes more efficient. High-sensitivity uncertain inputs are natural candidates for better measurement, research or monitoring.
This connects directly to Value of Information: sensitivity tells us which unknown can move the answer; value asks whether reducing that unknown is worth the cost.
9. Sensitivity-Analysis Failure Atlas
| Failure | What happens | Repair |
|---|---|---|
| One-number illusion | Single forecast is treated as certain | Vary plausible inputs |
| Wrong range | Inputs are varied unrealistically | Use evidence-based bounds |
| Interaction blindness | Combined effects are missed | Test linked assumptions together |
| Equal-attention bias | Low-impact inputs receive the same effort as critical ones | Rank influence |
| Threshold blindness | Decision flips near a boundary unnoticed | Probe around the threshold |
| Model lock | Only parameters change while model form stays unquestioned | Vary structural assumptions too |
| False precision | Sensitivity numbers hide uncertain evidence | Report bounds and interpretation |
10. Sensitivity Analysis and Counterfactual Simulation Are Different
Counterfactual Simulation explores alternative futures or pasts. Sensitivity analysis systematically changes inputs to measure how much the outcome depends on them.
Counterfactuals imagine alternatives. Sensitivity ranks leverage.
11. Teams Need a Shared Fragility Map
Teams often debate many assumptions without knowing which ones can change the decision. A shared sensitivity map focuses discussion on the assumptions that matter most.
The strongest meeting question may be: which assumption, if wrong, changes everything?
12. Institutions Need Assumption Registers
Long-lived decisions accumulate assumptions that may become invisible. Institutions can record which assumptions were critical, how sensitive the result was and what signal should trigger review.
This converts sensitivity analysis into maintenance infrastructure rather than a one-off calculation.
13. Artificial Intelligence and Sensitivity Analysis
AI-supported decisions may depend on prompts, retrieved evidence, tool outputs, thresholds and model settings. Reliable systems should know which of these changes materially alter the answer.
When small prompt or source changes produce large outcome changes, the system is fragile and should not present one output as uniquely determined.
Sensitivity testing can therefore reveal hidden dependence before automation scales it.
14. The Sensitivity Analysis Audit
- Outcome: Which conclusion or decision is being tested?
- Assumptions: Which inputs are uncertain?
- Bounds: What variation is plausible?
- Movement: How much does the output change?
- Thresholds: Where does the decision flip?
- Interactions: Which inputs matter together?
- Ranking: Which assumptions carry the most influence?
- Measurement: Which high-impact unknown deserves better evidence?
- Structure: Does model form itself need testing?
- Action: What contingency follows from the fragility map?
15. CivDJ Reading: Touch One Fader and Watch the Whole Mix
In the CivDJ frame, sensitivity analysis asks which channel, Master assumption or receiver estimate most strongly changes the final mix.
Low-impact channels can remain simple. High-sensitivity channels deserve better evidence, tighter controls and stronger monitoring.
Do not spend equal attention on every fader. Find the fader that changes the room.
16. Return to the Assumption Carrying the Weight
Sensitivity analysis gives intelligence a map of leverage.
It shows where uncertainty matters, where measurement matters and where a seemingly stable answer is balanced on one fragile assumption.
The strongest model is not the one whose assumptions are hidden. It is the one whose dependence on them can be inspected.