Nova Academy · Foundations of Inquiry · Lesson 029 · November 18, 2026
What you will learn
By the end, you should be able to define sensitivity analysis; change one model input while holding others fixed; compute and compare outcomes; distinguish local sensitivity from a wider scenario test; and explain why sensitivity does not establish a model’s correctness.
The central idea
A model transforms inputs and assumptions into outputs. Sensitivity analysis asks: if a plausible input changes, how much does the result change? This follows Lesson 028 on propagation of measurement uncertainty, but is not the same thing. Uncertainty propagation quantifies how uncertain inputs contribute to uncertain outputs under specified assumptions; sensitivity analysis explores the output’s response to chosen input changes. Neither substitutes for validation against observations.
Worked example: a simple water-use model
Suppose a classroom estimates daily water use as V = n × q, where n is the number of people and q is liters used per person per day. For an invented class of n = 20 and assumed q = 12 L/person/day, V = 20 × 12 = 240 L/day. Hold n = 20 fixed and vary q: at 10 L/person/day, V = 200 L/day; at 12, V = 240 L/day; at 14, V = 280 L/day. A ±2 change in q around 12 (±16.7%) produces a ±40 L/day change in V (±16.7%). The local derivative ∂V/∂q = n = 20 people, so each 1 L/person/day change in q changes V by 20 L/day. The arithmetic checks: 20 × 14 − 20 × 12 = 40 L/day.
A second parameter and interactions
Hold q = 12 fixed and vary n from 18 to 22 people: V becomes 216 to 264 L/day. When both parameters change, interactions matter because V = n × q. For example, n = 22 and q = 14 gives 308 L/day, rather than the 240 + 24 + 40 = 304 L/day obtained by simply adding separate baseline changes. The extra 4 L/day comes from the cross-product 2 × 2. This is a small but important reason to test combinations, not only one parameter at a time.
Activity: make a sensitivity table
Choose a simple model such as cost = price × quantity. Write a baseline price of $5 and quantity of 8 units, giving $40. Hold quantity at 8 and test prices $4, $5, $6: costs $32, $40, $48. Then hold price at $5 and test quantities 6, 8, 10: costs $30, $40, $50. Record which input change produces the larger absolute difference for these chosen ranges, and explain why different ranges could reverse that conclusion. Use fictional values, not personal spending records.
Interpreting the results
State the baseline, units, parameter ranges, and which inputs were held fixed. An output that changes sharply across a plausible input range is sensitive to that assumption. An output that barely changes is comparatively robust over the tested range. Neither observation proves that the assumptions, structure, or data are valid. Sensitivity results depend on the selected range, scale, and model.
Limitations and responsible practice
One-at-a-time tests can miss interactions, nonlinear thresholds, and correlated inputs. Global sensitivity analysis varies multiple parameters across defined ranges or distributions and can reveal effects that local tests miss. Do not assign probability to scenarios without a defensible probability model. If the original equation omits an important mechanism, a beautifully precise sensitivity table can still mislead. Always pair model exploration with empirical validation where possible.
Reflection
Which assumption in the example would you most want to measure rather than guess? What evidence would help you choose a plausible range? How might a model be insensitive to one parameter yet still wrong?
Knowledge check
1. What does sensitivity analysis vary? 2. If n = 20 and q rises from 12 to 13 L/person/day, what happens to V? 3. Why can simultaneous parameter changes differ from the sum of separate effects? 4. Does low sensitivity prove a model is correct?
Answer key
1. One or more model inputs or assumptions, to see how outputs respond. 2. V rises from 240 to 260 L/day, a 20 L/day increase. 3. Nonlinear models and interactions, including the n × q cross-product, can create additional effects. 4. No; low sensitivity only describes response within the tested ranges and model.
Sources and further reading
NIST/SEMATECH e-Handbook of Statistical Methods (modeling, experimental design, and uncertainty); JCGM Guide to the Expression of Uncertainty in Measurement (measurement uncertainty); US EPA — Sensitivity Analysis (model sensitivity concepts). The classroom numbers here are illustrative, not research findings.
Continue learning
Review Lesson 028 — Propagation of Measurement Uncertainty · Nova Academy index
