1 Definition and scope

Sensitivity analysis refers to a family of methods for examining how changes in a model’s inputs, assumptions, or parameters influence its outputs. It is used with both simple formulas and large computational models. The central idea is to identify which factors matter most, how strongly they matter, and whether the overall conclusions remain stable when conditions vary.

1.1 Basic concept

In its simplest form, sensitivity analysis asks what happens to a result when one input is adjusted while others are held fixed, or when several inputs change together. If a small change in a parameter produces a large change in output, the model is said to be sensitive to that parameter. If the output changes little, the parameter has a weaker effect.

1.2 Objectives

Sensitivity analysis serves several purposes. It can reveal dominant variables, test the stability of predictions, and highlight where more accurate data would be most valuable. It also helps compare competing assumptions, simplify models by removing unimportant factors, and communicate the reliability of results to decision-makers.

1.3 Relationship to uncertainty analysis

Sensitivity analysis is closely related to uncertainty analysis, but the two are not identical. Uncertainty analysis focuses on how unknown or variable inputs produce uncertainty in outputs. Sensitivity analysis investigates how strongly each input contributes to that uncertainty. In practice, the methods are often used together, since a model may be uncertain for many reasons, but only some inputs drive that uncertainty strongly.

1.4 Relationship to model validation

Model validation assesses whether a model adequately represents the system it is intended to describe. Sensitivity analysis supports validation by showing whether outputs respond in plausible ways to changes in parameters and assumptions. It can help detect unrealistic behavior, unstable parameterizations, or hidden dependencies that would otherwise remain unnoticed.

2 Types of sensitivity analysis

Sensitivity analysis can be classified in several ways, including by the scale of perturbation, the mathematical tools used, and whether the analysis focuses on exact formulas or on broader patterns. Different types are suited to different model structures and research goals.

2.1 Local sensitivity analysis

Local sensitivity analysis examines behavior near a specific baseline input set. It is useful when a model is well understood around a reference point and when small changes are of interest.

2.1.1 Derivative-based methods

Derivative-based methods measure how rapidly an output changes with respect to an input, usually by calculating a slope or gradient. These methods are efficient and precise for smooth models, but they capture only local behavior and may miss effects that appear under larger changes.

2.1.2 One-at-a-time methods

One-at-a-time methods vary a single input while holding all others constant. This approach is easy to interpret and widely used in introductory analyses. However, it may overlook interactions among inputs, since many real systems respond to combinations of changes rather than isolated adjustments.

2.2 Global sensitivity analysis

Global sensitivity analysis examines input variation across a broad range of possible values. It is designed to capture nonlinear effects, interactions, and behavior over the full input space rather than only near one reference point.

2.2.1 Variance-based methods

Variance-based methods attribute portions of output variance to individual inputs and their interactions. They are among the most widely used global techniques because they provide a detailed picture of how uncertainty in inputs is distributed across the model.

2.2.2 Screening methods

Screening methods aim to identify the few inputs that have the largest effects from among many candidates. They are especially useful in early-stage model studies or in high-dimensional settings where a full global analysis would be too costly.

2.2.3 Moment-independent methods

Moment-independent methods assess sensitivity without relying solely on variance or standard deviation. Instead, they examine broader changes in output distributions, which can be valuable when the response is skewed, multimodal, or otherwise not well summarized by moments alone.

2.3 Qualitative sensitivity analysis

Qualitative sensitivity analysis considers whether outputs change directionally or categorically when assumptions are altered. It is often used when numerical data are limited or when the model is too informal for precise quantitative measures. The focus is on patterns of dependence rather than exact magnitudes.

2.4 Deterministic and probabilistic approaches

Deterministic approaches use fixed input values or structured perturbations to trace model response. Probabilistic approaches treat inputs as random variables with specified distributions and evaluate sensitivity over many possible combinations. Probabilistic methods are generally better suited to uncertainty-rich problems, while deterministic methods are often simpler and more transparent.

3 Methodology

A sensitivity study usually follows a sequence of steps, from selecting relevant variables to interpreting the resulting measures. The exact procedure depends on the model, the data available, and the desired level of detail.

3.1 Selecting inputs and outputs

The first step is to determine which inputs should be examined and which outputs matter most. Inputs may include parameters, initial conditions, boundary values, or empirical assumptions. Outputs should be defined clearly so the analysis addresses meaningful quantities such as forecasts, risks, or performance measures.

3.2 Defining baseline assumptions

A baseline provides the reference point against which changes are measured. It may come from prior knowledge, estimated parameters, or standard operating conditions. A well-chosen baseline makes results easier to interpret and helps ensure that the analysis reflects realistic conditions.

3.3 Perturbation strategies

Perturbation strategies specify how inputs are varied. Common choices include small increments, percentage changes, random sampling, or systematic exploration of ranges. The selected strategy should match the question being asked: narrow perturbations for local behavior, broader sampling for global effects, and structured scenarios for comparative studies.

3.4 Measuring output response

Output response can be measured in many ways, including absolute change, relative change, variance contribution, rank order, or distributional shift. The appropriate metric depends on the model and the interpretation desired. A single study may use more than one measure to capture different aspects of sensitivity.

3.5 Interpreting sensitivity measures

Interpretation requires attention to context. A parameter with a large numerical effect is not always the most important in practice if it is unlikely to vary. Conversely, a moderately influential input may deserve attention if it is highly uncertain or costly to estimate. Sensitivity results are therefore best read alongside knowledge about realism, uncertainty, and decision relevance.

4 Mathematical foundations

The mathematics of sensitivity analysis ranges from elementary calculus to probability theory and numerical approximation. These foundations provide the tools for quantifying how outputs depend on inputs.

4.1 Partial derivatives

Partial derivatives describe the rate at which a function changes when one variable varies and all others remain fixed. In sensitivity analysis, they provide a direct measure of local influence. They are especially useful for smooth models, where gradients can be computed analytically or numerically.

4.2 Elasticity

Elasticity measures relative change rather than absolute change. It is commonly defined as the percentage change in output associated with a one percent change in an input. This makes results easier to compare across variables with different units or scales.

4.3 Finite difference approximation

Finite difference approximation estimates derivatives by evaluating the model at nearby points. It is widely used when analytic derivatives are unavailable. Although simple, it can be affected by step size, numerical noise, and rounding error, so careful choice of perturbation is important.

4.4 Variance decomposition

Variance decomposition partitions the variability in an output into contributions from different inputs and combinations of inputs. It provides the basis for many global sensitivity measures and is particularly valuable for nonlinear models with interacting factors.

4.4.1 Sobol indices

Sobol indices are variance-based measures that quantify the contribution of individual inputs and their interactions to total output variance. They are widely used because they separate main effects from higher-order effects and offer a comprehensive picture of influence.

4.4.2 Main-effect and interaction effects

Main effects describe the influence of each input acting alone, while interaction effects capture cases in which the effect of one input depends on the value of another. Interaction terms are important in systems where variables do not act independently, since they can dominate behavior even when individual effects seem modest.

5 Applications

Sensitivity analysis is used across many fields to clarify model behavior, focus attention on influential variables, and support robust decision-making. Its practical value lies in helping analysts understand complex systems with many moving parts.

5.1 Engineering and physical sciences

In engineering, sensitivity analysis is used to study design tolerances, safety margins, and performance trade-offs. In the physical sciences, it helps evaluate how model parameters affect simulations of processes such as fluid flow, heat transfer, or material behavior. It is often essential in systems where small parameter shifts can change stability or efficiency.

5.2 Economics and finance

Economists and financial analysts use sensitivity analysis to test how forecasts, valuations, or risk measures respond to changing assumptions. Examples include interest rates, discount factors, demand elasticities, and volatility estimates. The method helps identify which assumptions most strongly shape outcomes in models that guide planning or investment.

5.3 Climate and environmental modeling

Environmental models often involve many uncertain inputs, including emissions rates, physical constants, and land-use assumptions. Sensitivity analysis helps determine which drivers most affect projections of temperature, runoff, pollution, or ecosystem response. It also guides where to improve data collection and where model refinement is most urgent.

5.4 Medicine and epidemiology

In medicine and epidemiology, sensitivity analysis supports the study of disease spread, treatment effectiveness, and clinical decision models. It can reveal how assumptions about transmission, recovery, adherence, or measurement error influence predictions. This is useful when evidence is incomplete or when several plausible parameter values exist.

5.5 Operations research and optimization

Operations research uses sensitivity analysis to examine how changes in costs, capacities, constraints, or demand affect optimal solutions. It helps decision-makers understand whether a recommended plan remains effective when conditions shift. In optimization, it is often paired with tools that examine stability around the optimum or the impact of parameter uncertainty.

6 Computational techniques

Because many modern models are too complex for purely analytic treatment, sensitivity analysis often relies on computation. Efficient numerical methods make it possible to explore large input spaces and expensive simulations.

6.1 Monte Carlo simulation

Monte Carlo simulation evaluates a model repeatedly using randomly sampled inputs. The resulting distribution of outputs can be analyzed to estimate sensitivity, uncertainty, and interaction effects. This approach is flexible, though it may require many runs to achieve stable estimates.

6.2 Surrogate models

Surrogate models are simplified approximations of a more expensive original model. They are used to reduce computation time while preserving the main response structure. Once fitted, a surrogate can be analyzed more quickly to explore sensitivity over a wide range of inputs.

6.3 Emulation and metamodeling

Emulation and metamodeling refer to building secondary models that reproduce the behavior of the original system closely enough for analysis. These techniques are useful when direct simulation is costly. They can support extensive sensitivity studies that would otherwise be impractical.

6.4 Sampling methods

Sampling methods determine which combinations of input values are evaluated. Good sampling design improves the efficiency and reliability of the analysis by covering the input space evenly and reducing the chance of missing important regions.

6.4.1 Latin hypercube sampling

Latin hypercube sampling divides each input range into equal-probability intervals and draws samples so that all intervals are represented. This creates a more balanced exploration than simple random sampling and is often used in global sensitivity studies.

6.4.2 Quasi-random sequences

Quasi-random sequences generate points that fill the input space more uniformly than ordinary random draws. They are commonly used to improve convergence in simulation-based studies and can be especially helpful when the model must be evaluated many times.

7 Interpretation and reporting

The value of a sensitivity study depends not only on the calculations but also on how the results are explained. Clear reporting helps readers understand which inputs matter, how stable the findings are, and what limits apply.

7.1 Ranking influential parameters

A common reporting practice is to rank parameters by their estimated influence. Rankings help identify the most important drivers, but they should not be treated as absolute when differences are small or when interaction effects are substantial. Context and uncertainty around the ranking matter.

7.2 Assessing robustness

Robustness refers to whether conclusions remain similar under reasonable changes in assumptions or inputs. A robust model or decision rule is one that does not depend heavily on a narrow set of precise values. Sensitivity analysis provides one of the main tools for testing this property.

7.3 Visualizing results

Graphs such as tornado diagrams, response surfaces, scatter plots, and heat maps are often used to display sensitivity findings. Visualization can make complex patterns easier to grasp, especially when interactions or nonlinearities are present. Good figures usually clarify both magnitude and direction of effects.

7.4 Common pitfalls

Common mistakes include using too narrow a range of inputs, ignoring correlations, interpreting local results as if they were global, and focusing only on the largest numerical effect without considering uncertainty. Another frequent problem is reporting sensitivity measures without explaining the model context or the assumptions behind them.

8 Limitations and challenges

Although sensitivity analysis is widely useful, it has practical and conceptual limits. These challenges become more pronounced as models grow larger, more nonlinear, or more tightly coupled.

8.1 High-dimensional models

Models with many inputs can be difficult to analyze because the number of possible combinations grows rapidly. This creates a need for screening methods, efficient sampling, and careful prioritization. Otherwise, the analysis may become too broad to interpret well.

8.2 Nonlinear interactions

When outputs depend on complex interactions among variables, simple one-factor-at-a-time approaches may be misleading. Nonlinear behavior can produce thresholds, saturation, or sudden shifts that are hard to detect without global methods. Such systems often require more sophisticated tools and richer sampling.

8.3 Computational cost

Detailed sensitivity studies can be expensive, particularly when each simulation takes a long time. Large numbers of model evaluations may be needed to estimate robust measures. This is one reason surrogate models and emulators are frequently used in practice.

8.4 Dependence among inputs

Many methods assume that inputs vary independently, but real systems often contain correlated variables. Dependence can alter sensitivity estimates and complicate variance decomposition. Analysts must account for such relationships, or the results may overstate or understate the role of particular inputs.

Sensitivity analysis is closely connected to other fields concerned with model behavior, uncertainty, and decision-making. These neighboring topics often overlap in method and purpose.

9.1 Uncertainty quantification

Uncertainty quantification studies how uncertainty in inputs, parameters, and structure propagates through a model to affect outputs. Sensitivity analysis is often a component of this broader framework, helping identify which uncertainties matter most.

9.2 Model calibration

Model calibration estimates parameter values so that a model matches observed data. Sensitivity analysis informs calibration by showing which parameters are identifiable or influential. It can also indicate whether some parameters have little effect and may be difficult to estimate reliably.

9.3 Scenario analysis

Scenario analysis examines model behavior under a set of plausible alternative conditions. It differs from sensitivity analysis in that it usually emphasizes coherent narratives or cases rather than systematic input perturbation. The two approaches are complementary and are often used together.

9.4 Decision analysis

Decision analysis evaluates choices under uncertainty by comparing outcomes, trade-offs, and preferences. Sensitivity analysis supports this process by showing how strongly recommendations depend on assumptions or parameter values. It is especially useful when a decision is robust across many plausible settings.