What Is Sensitivity Analysis?

Sensitivity analysis is a mathematical technique used to determine how changes in the input parameters of a model affect its outputs. It examines the relationships between model inputs and outputs to identify which factors contribute most to output uncertainty, which are negligible, and how parameters interact with one another. The method appears across engineering simulation, economic modeling, environmental assessment, and reliability engineering, wherever a computational model encodes uncertain or variable quantities. Its roots lie in classical perturbation theory and first-order derivative analysis, but modern practice extends to global variance-decomposition methods capable of handling highly nonlinear systems.

Local and Global Approaches

Local sensitivity analysis varies one parameter at a time around a nominal reference point, computing the partial derivative of the output with respect to each input. This one-at-a-time (OAT) approach is computationally inexpensive and yields results that are easy to interpret, but its validity depends on the assumption that model behavior near the reference point is representative of the entire feasible space. For linear or nearly linear models, this assumption holds; for nonlinear models it can produce severely misleading rankings of parameter importance. Global sensitivity analysis addresses this limitation by varying all uncertain inputs simultaneously across their full feasible ranges, capturing interaction effects that local methods miss. The Addressing Uncertainty in MultiSector Dynamics research framework provides a thorough comparison of these two classes of methods and their appropriate use conditions.

Variance-Based Methods

Variance-based sensitivity analysis, associated primarily with the Sobol indices developed by Ilya Sobol in 1993, decomposes the total variance of a model output into contributions from individual inputs and their interactions. The first-order index S_i for parameter x_i measures the fraction of output variance attributable to x_i alone. The total-effect index S_Ti additionally captures the variance arising from all interactions in which x_i participates, including higher-order combinations with other parameters. When S_Ti is close to zero, the corresponding parameter can be fixed without meaningfully changing the output distribution, a result of direct practical value for model simplification and experimental design. The Sobol method captures all-order interactions but is computationally intensive, typically requiring many thousands of model evaluations. The Morris screening method offers a less expensive alternative: it applies an OAT design across multiple random trajectories to estimate mean and variance of elementary effects, providing a qualitative ranking that is often used as a first pass before running a full variance-based analysis. Research comparing global sensitivity analysis methods has shown that Morris rankings can serve as reliable proxies for Sobol total-effect indices in many engineering applications.

Sensitivity Analysis in Engineering Design

In engineering contexts, sensitivity analysis informs which design parameters, material properties, or loading conditions most strongly influence a key performance metric such as stress, efficiency, or failure probability. Finite element models of structural components, for instance, may depend on dozens of material constants and boundary conditions; sensitivity analysis identifies which of those constants merit precise experimental measurement and which can be treated as fixed without loss of accuracy. In circuit design, sensitivity methods identify which component tolerances drive the tightest constraints on manufacturing yield. In systems reliability analysis, the technique pinpoints which subsystem failure rates dominate the overall system unavailability. The evaluation of global sensitivity methods in computational structural mechanics benchmarks several leading methods against structural models with varying degrees of nonlinearity and parameter interaction.

Applications

Sensitivity analysis has applications in a wide range of modeling and design fields, including:

• Engineering reliability and structural risk assessment
• Circuit tolerance budgeting and manufacturing yield optimization
• Environmental and climate model calibration
• Pharmacokinetic and biological systems modeling
• Financial model stress testing and scenario analysis