The Central Idea
Suppose a model produces
\[ y=f(x;\theta), \]where \(x\) represents inputs and \(\theta\) represents parameters.
If \(x\) or \(\theta\) changes slightly, does \(y\) barely change, or does it change dramatically?
That is a sensitivity question.
If we are uncertain about the true values of \(x\) or \(\theta\), how uncertain should we be about \(y\)?
That is an uncertainty question.
A model is most useful when we understand both.
Sensitivity and Uncertainty Are Different
Sensitivity concerns responsiveness.
Uncertainty concerns lack of precise knowledge.
A parameter can be highly uncertain but have little effect on the output. Another parameter can be known precisely but have a very strong effect.
Therefore, high uncertainty and high sensitivity are not the same thing.
A Simple Example
Suppose profit is modeled by
\[ P=(p-c)q-F, \]where:
- \(p\) is unit price;
- \(c\) is unit cost;
- \(q\) is quantity sold;
- \(F\) is fixed cost.
If \(q\) is uncertain and profit changes sharply when \(q\) changes, then profit is sensitive to quantity.
Local Sensitivity
Local sensitivity examines how an output changes near a particular reference point.
If
\[ y=f(x), \]then the derivative
\[ \boxed{\frac{dy}{dx}=f'(x)} \]measures local sensitivity.
A large magnitude of \(f'(x)\) means small changes in \(x\) can produce relatively large changes in \(y\).
Example: Local Sensitivity
Suppose
\[ A(r)=\pi r^2. \]Then
\[ \frac{dA}{dr}=2\pi r. \]At \(r=10\),
\[ \frac{dA}{dr}=20\pi. \]Near \(r=10\), a one-unit change in radius changes area by approximately \(20\pi\) square units.
Sensitivity Depends on the Operating Point
In nonlinear models, sensitivity may change from one region to another.
For
\[ A(r)=\pi r^2, \]the sensitivity
\[ 2\pi r \]increases with \(r\).
The model is more sensitive to a one-unit change in radius when the radius is large than when it is small.
Finite-Difference Sensitivity
Derivatives are not always available or convenient.
A simple numerical sensitivity estimate is
\[ \boxed{ \frac{\Delta y}{\Delta x} = \frac{f(x+\Delta x)-f(x)}{\Delta x} }. \]This measures the observed change in output per unit change in the input.
One-at-a-Time Sensitivity Analysis
A basic practical method is to vary one input or parameter while holding the others fixed.
For example:
- compute the baseline output;
- increase one parameter by \(10\%\);
- recompute the output;
- decrease the same parameter by \(10\%\);
- compare the changes;
- repeat for other parameters.
This is often called one-at-a-time sensitivity analysis.
Baseline Values
Sensitivity analysis usually begins with a baseline scenario.
Suppose
\[ C=F+cq \]with
\[ F=1000,\qquad c=8,\qquad q=500. \]Then baseline cost is
\[ C=1000+8(500)=5000. \]Other scenarios can be compared with this baseline.
Example: Varying One Parameter
Using
\[ C=1000+c(500), \]if \(c\) increases from \(8\) to \(8.80\), then
\[ C=1000+8.80(500)=5400. \]A \(10\%\) increase in unit cost raises total cost from \(5000\) to \(5400\).
The output change is
\[ 400. \]Absolute Sensitivity
Absolute sensitivity measures output change in the original units.
If a parameter changes by
\[ \Delta x \]and the output changes by
\[ \Delta y, \]then the absolute sensitivity over that change is
\[ \frac{\Delta y}{\Delta x}. \]Relative Sensitivity
Absolute sensitivity can be difficult to compare across variables with different units.
Relative sensitivity compares percentage changes.
A common measure is elasticity:
\[ \boxed{ E = \frac{\Delta y/y}{\Delta x/x} }. \]In differential form,
\[ \boxed{ E = \frac{x}{y}\frac{dy}{dx} }. \]Interpreting Elasticity
If
\[ E=2, \]then a \(1\%\) increase in \(x\) produces approximately a \(2\%\) increase in \(y\), locally.
If
\[ E=-0.5, \]then a \(1\%\) increase in \(x\) produces approximately a \(0.5\%\) decrease in \(y\).
Example: Elasticity of a Power Model
Suppose
\[ y=ax^b. \]Then
\[ \frac{dy}{dx}=abx^{b-1}. \]Therefore,
\[ E = \frac{x}{ax^b}\left(abx^{b-1}\right) = b. \]The exponent \(b\) is the elasticity of the power model.
Parameter Sensitivity
If a model depends on several parameters,
\[ y=f(x;\theta_1,\theta_2,\ldots,\theta_k), \]sensitivity to parameter \(\theta_j\) can be measured locally by
\[ \boxed{ \frac{\partial y}{\partial \theta_j} }. \]Partial derivatives isolate the local effect of one parameter while the others are held fixed.
Example: Profit Sensitivity
For
\[ P=(p-c)q-F, \]the partial derivatives are
\[ \frac{\partial P}{\partial p}=q, \] \[ \frac{\partial P}{\partial c}=-q, \] \[ \frac{\partial P}{\partial q}=p-c, \] \[ \frac{\partial P}{\partial F}=-1. \]These expressions show how local profit sensitivity depends on the current operating conditions.
Interactions
Changing one parameter at a time can miss interactions.
Suppose
\[ y=xy. \]The effect of changing \(x\) depends on the current value of \(y\), and the effect of changing \(y\) depends on the current value of \(x\).
When parameters interact, simultaneous variation may be necessary.
Global Sensitivity
Global sensitivity analysis examines model behavior across a broader region of the input or parameter space.
Instead of asking what happens near one baseline point, it asks which inputs matter across the full range of plausible values.
This is especially important for nonlinear models.
Scenario Analysis
Scenario analysis evaluates a model under deliberately chosen sets of assumptions or parameter values.
Common scenarios include:
- baseline;
- optimistic;
- pessimistic;
- high-growth;
- low-growth;
- high-cost;
- low-demand.
Scenarios are not probabilities unless probabilities are explicitly assigned to them.
Best-Case and Worst-Case Analysis
If parameters are known only to lie in intervals, one approach is to examine extreme combinations.
Suppose
\[ P=(p-c)q-F \]with
\[ 20\leq p\leq 22, \] \[ 12\leq c\leq 14, \] \[ 400\leq q\leq 500. \]A best-case and worst-case calculation provides a simple range of possible outcomes under those bounds.
Threshold Analysis
Sometimes the important question is not the exact output but when a decision changes.
Suppose a project is acceptable only when
\[ P>0. \]Sensitivity analysis can identify the parameter values at which
\[ P=0. \]Those values are decision thresholds.
Example: Break-Even Sensitivity
Suppose
\[ P=(p-c)q-F. \]Break-even occurs when
\[ P=0. \]Therefore,
\[ (p-c)q=F \]and
\[ \boxed{ q=\frac{F}{p-c} }. \]This formula shows directly how the break-even quantity depends on fixed cost, price, and unit cost.
What Is Uncertainty?
Uncertainty means that the exact value, state, parameter, outcome, or model structure is not known with complete precision.
Uncertainty can arise from:
- measurement error;
- sampling variation;
- future events;
- random processes;
- limited data;
- unknown parameter values;
- model-form choices;
- simplifying assumptions.
Input Uncertainty
Inputs supplied to a model may be uncertain.
A future demand forecast might be
\[ D=1000\pm 100. \]A measured length might be
\[ L=20.0\pm 0.1. \]Uncertain inputs produce uncertainty in model outputs.
Parameter Uncertainty
Parameters estimated from data are rarely known exactly.
If a growth rate is estimated as
\[ r=0.04, \]the true rate may plausibly be somewhat higher or lower.
Parameter uncertainty should be propagated through the model when it could materially affect conclusions.
Structural Uncertainty
Sometimes the main uncertainty is not the parameter value but the model form itself.
Should growth be modeled as
\[ P(t)=P_0e^{rt} \]or with a logistic mechanism?
Should a relationship be linear or nonlinear?
Competing plausible structures create model uncertainty.
Scenario Uncertainty
Future conditions may depend on events that cannot be reduced to one parameter estimate.
A planning model may therefore analyze several scenarios rather than claim one precise forecast.
Scenario analysis is especially useful when uncertainty is driven by qualitatively different future regimes.
Aleatory and Epistemic Uncertainty
A useful conceptual distinction separates two broad sources of uncertainty.
Aleatory uncertainty refers to inherent randomness or variability in a process.
Epistemic uncertainty refers to incomplete knowledge.
Random customer arrivals may be aleatory. An imprecisely estimated service rate may be epistemic.
Ranges as Uncertainty Representations
One simple way to represent uncertainty is with an interval:
\[ x\in[x_{\min},x_{\max}]. \]The model can then be evaluated across that range.
This is useful when probabilities are unavailable or unjustified.
Probability Distributions
When uncertainty can be described probabilistically, an input may be modeled as a random variable.
For example,
\[ D\sim N(1000,100^2) \]represents uncertain demand with mean \(1000\) and standard deviation \(100\), under a normal-model assumption.
The output then becomes a random quantity as well.
Uncertainty Propagation
Uncertainty propagation means carrying uncertainty in inputs and parameters through the model to determine uncertainty in the output.
If
\[ Y=f(X), \]and \(X\) is uncertain, then \(Y\) is uncertain because its value depends on \(X\).
Linear Approximation to Uncertainty
If \(Y=f(X)\) and uncertainty in \(X\) is small, then
\[ \Delta Y\approx f'(X)\Delta X. \]This is the same local derivative that appears in sensitivity analysis.
Sensitivity tells us how strongly uncertainty is transmitted.
Several Uncertain Inputs
For
\[ Y=f(X_1,X_2,\ldots,X_n), \]a first-order approximation is
\[ \Delta Y \approx \frac{\partial f}{\partial X_1}\Delta X_1 + \cdots + \frac{\partial f}{\partial X_n}\Delta X_n. \]Each uncertain input contributes through both its uncertainty and the model's sensitivity to it.
Monte Carlo Simulation
A widely used numerical method for uncertainty propagation is Monte Carlo simulation.
The basic workflow is:
- assign distributions or sampling rules to uncertain inputs;
- draw one set of input values;
- run the model;
- store the output;
- repeat many times;
- summarize the resulting output distribution.
What Monte Carlo Produces
Instead of one output, simulation produces many possible outputs:
\[ y_1,y_2,\ldots,y_N. \]These can be summarized by:
- mean;
- median;
- standard deviation;
- quantiles;
- probability of exceeding a threshold;
- probability of loss;
- minimum and maximum simulated outcomes.
Example: Uncertain Profit
Suppose
\[ P=(p-c)Q-F \]and demand \(Q\) is uncertain.
If
\[ Q\sim N(500,50^2), \]then profit is also uncertain.
Repeatedly sampling \(Q\) and computing \(P\) creates a simulated distribution of possible profit outcomes.
Expected Output
When a stochastic model is used, the expected output may be important.
If
\[ P=(p-c)Q-F \]and \(Q\) is random, then
\[ E[P] = (p-c)E[Q]-F. \]Expected value summarizes the average outcome, but it does not describe the full risk.
Expected Value Is Not Enough
Two models can have the same expected output and very different uncertainty.
One may produce outcomes tightly clustered near the mean.
Another may produce both large gains and large losses.
Decisions often depend on spread, tails, and thresholds as well as the average.
Confidence and Prediction Intervals
Statistical models often summarize uncertainty using intervals.
A confidence interval may quantify uncertainty about an estimated parameter or mean.
A prediction interval addresses uncertainty about a future observation.
These are different targets and should not be interpreted as interchangeable.
Correlation among Uncertain Inputs
Uncertain inputs are not always independent.
Suppose price and demand tend to move in opposite directions.
Sampling them independently may generate unrealistic combinations.
Dependence structure should be represented when it materially affects results.
Robustness
A conclusion is robust when it remains broadly unchanged across reasonable variations in inputs, parameters, assumptions, or model choices.
For example, if one investment remains preferred under every plausible scenario, the decision is robust to the modeled uncertainty.
Robust conclusions are usually more decision-worthy than conclusions that reverse under tiny changes.
Fragility
A conclusion is fragile when small changes in assumptions or parameters change the result substantially.
Suppose a project appears profitable only when demand exceeds exactly
\[ 1000 \]units, while plausible demand lies between \(900\) and \(1100\).
The decision is sensitive to demand uncertainty and should be presented as fragile rather than certain.
Tornado-Style Ranking
In practice, one-at-a-time sensitivity results are often ranked by the magnitude of their effect on the output.
The largest output changes identify the parameters that deserve the most attention.
This ranking can help prioritize:
- better measurement;
- better calibration;
- additional data collection;
- risk controls;
- model refinement.
Sensitivity Can Guide Data Collection
If the model is almost insensitive to a parameter, measuring that parameter more precisely may provide little benefit.
If the model is highly sensitive to another poorly known parameter, better measurement may substantially reduce output uncertainty.
Sensitivity analysis therefore helps allocate analytical effort.
Sensitivity Can Guide Model Simplification
A parameter with negligible influence across the relevant domain may sometimes be fixed or removed from a simplified model.
This should be done cautiously, because a parameter that is unimportant in one region may matter elsewhere.
Global sensitivity analysis can help test whether simplification is safe.
Uncertainty Does Not Mean Ignorance
Saying that an input is uncertain does not mean nothing is known.
We may know:
- a plausible range;
- a mean and standard deviation;
- a probability distribution;
- a set of scenarios;
- historical variability;
- physical bounds.
Good uncertainty analysis represents what is known without pretending to know more.
Common Error: Varying Implausible Values
Sensitivity analysis is only meaningful over relevant ranges.
If a probability parameter must satisfy
\[ 0\leq p\leq 1, \]testing \(p=1.4\) is outside the model's admissible domain.
Sensitivity ranges should respect constraints and meaning.
Common Error: Changing One Parameter and Calling It Uncertainty Analysis
One-at-a-time sensitivity analysis shows responsiveness.
It does not automatically quantify the probability or range of the model output.
Uncertainty analysis requires a representation of how uncertain the inputs or parameters actually are.
Common Error: Assuming Independence
If uncertain inputs are related, independent variation can exaggerate or understate output uncertainty.
Correlations and logical constraints should be preserved where they matter.
Common Error: Reporting Only a Baseline
A single baseline output can create false precision.
If important inputs are uncertain, model communication should often include ranges, scenarios, intervals, or probabilities in addition to the central estimate.
Common Error: Hiding Structural Uncertainty
A narrow parameter interval does not make the entire model certain.
If two plausible model forms produce very different results, structural uncertainty may dominate parameter uncertainty.
Competing model structures should be compared when the choice matters.
Common Error: Confusing Robustness with Accuracy
A model can be robust but wrong.
If every plausible parameter setting within an incorrect model structure produces the same wrong answer, the result is robust to parameter changes but not accurate.
Robustness complements validation; it does not replace it.
A Reliable Sensitivity Procedure
- Define the output or decision of interest.
- Choose baseline parameter and input values.
- Identify plausible ranges for quantities that may vary.
- Vary one quantity at a time for an initial screen.
- Record absolute and relative output changes.
- Use derivatives or elasticities when useful.
- Examine interactions among important parameters.
- Expand to global sensitivity if local results are insufficient.
- Rank the quantities that most strongly influence the result.
- Check whether the model conclusion remains stable.
A Reliable Uncertainty Procedure
- Identify uncertain inputs, parameters, assumptions, and model choices.
- Separate inherent variability from incomplete knowledge when useful.
- Represent uncertainty using ranges, scenarios, or probability distributions.
- Preserve important dependencies among uncertain quantities.
- Propagate uncertainty through the model.
- Summarize the resulting output distribution or range.
- Examine threshold and tail outcomes, not only the mean.
- Compare uncertainty with decision tolerances.
- Identify which uncertainties contribute most to decision risk.
- Communicate the conclusion together with its uncertainty.
Practice 1: Sensitivity or Uncertainty?
A modeler increases a parameter by \(5\%\) to see how much the output changes.
Is this primarily sensitivity analysis or uncertainty analysis?
Practice 2: Local Sensitivity
Suppose
\[ y=x^2. \]Find the local sensitivity
\[ \frac{dy}{dx} \]at \(x=4\).
Practice 3: Elasticity
For
\[ y=5x^3, \]find the elasticity of \(y\) with respect to \(x\).
Practice 4: Profit Sensitivity
Profit is
\[ P=(p-c)q-F. \]Find
\[ \frac{\partial P}{\partial c}. \]Practice 5: Break-Even Threshold
A project has
\[ P=(20-12)q-4000. \]Find the break-even quantity.
Practice 6: Input Uncertainty
Demand is estimated to lie between \(900\) and \(1100\) units.
What simple uncertainty representation is already available?
Practice 7: Monte Carlo
State the main idea of Monte Carlo uncertainty propagation in one sentence.
Practice 8: Robust Decision
Strategy A outperforms Strategy B under every plausible parameter combination considered.
What does this suggest about the decision?
Practice 9: Sensitivity and Uncertainty Together
Parameter \(a\) is highly uncertain but has almost no effect on the output. Parameter \(b\) is only mildly uncertain but has a very large effect.
Which parameter may deserve more immediate attention for the output?
Practice 10: Identify the Error
A modeler reports only one baseline forecast even though several major inputs are uncertain over wide ranges.
What important modeling information is missing?
Practice Answers
Practice 1
This is primarily sensitivity analysis because the purpose is to measure how responsive the output is to a controlled parameter change.
Practice 2
\[ \frac{dy}{dx}=2x. \]At \(x=4\),
\[ \boxed{\frac{dy}{dx}=8}. \]Practice 3
For
\[ y=5x^3, \]the power-model elasticity is the exponent:
\[ \boxed{E=3}. \]Practice 4
\[ \boxed{\frac{\partial P}{\partial c}=-q}. \]Increasing unit cost reduces profit at a rate equal to the quantity sold.
Practice 5
\[ P=8q-4000. \]At break-even,
\[ 0=8q-4000. \]Therefore,
\[ \boxed{q=500}. \]Practice 6
The uncertainty can be represented by the interval
\[ \boxed{D\in[900,1100]}. \]Practice 7
Monte Carlo simulation repeatedly samples plausible uncertain inputs, runs the model for each sample, and uses the resulting outputs to approximate the distribution of possible outcomes.
Practice 8
The preference for Strategy A appears robust to the modeled parameter uncertainty.
Practice 9
Parameter \(b\) may deserve more immediate attention because the model is highly sensitive to it. Even modest uncertainty in \(b\) can strongly affect the output.
Practice 10
The forecast lacks an explicit representation of uncertainty. Ranges, scenarios, intervals, distributions, or other uncertainty summaries should be considered so the baseline is not mistaken for a certain outcome.
What to Remember
- Sensitivity measures how strongly outputs respond to changes in inputs or parameters.
- Uncertainty measures how imperfectly inputs, parameters, outcomes, or model structures are known.
- Local sensitivity can be measured with derivatives or partial derivatives.
- Elasticity compares percentage changes and is useful across different units.
- One-at-a-time analysis is a useful first screen but can miss interactions.
- Global sensitivity examines a broader region of the model's input space.
- Scenarios, ranges, and probability distributions are different ways to represent uncertainty.
- Monte Carlo simulation propagates uncertainty by repeated sampling and model evaluation.
- Expected values do not describe the full range of possible outcomes.
- Robust conclusions remain stable across reasonable changes in assumptions and parameters.
- Fragile conclusions should be communicated as conditional rather than certain.
- Sensitivity analysis can guide data collection, calibration, and model simplification.
- Robustness does not replace validation.
The central modeling habit is: do not report a model result without asking how much it depends on what you assumed, measured, estimated, or left uncertain.
Where We Go Next
Once a model has been built, calibrated, validated, and tested for sensitivity and uncertainty, its results must be explained correctly.
The next lesson, Interpretation and Communication, develops how to translate mathematical outputs back into the language of the original problem, communicate assumptions and uncertainty, and avoid presenting model results with more precision or certainty than the model supports.