Sensitivity Analysis in Scientific Models

Updated 16 May 2026

Sensitivity analysis is the systematic investigation of how varying model inputs impact outputs, clearly attributing uncertainty to individual factors.
It combines local techniques, such as derivative-based measures, with global methods like Sobol’ indices to capture both linear and nonlinear effects.
It supports robust model calibration, validation, and decision-making in fields ranging from environmental modeling to causal inference.

Sensitivity analysis is the systematic investigation of how variations in model inputs affect outputs, central in quantifying, attributing, and managing uncertainty in mathematical, statistical, and computational models. It provides critical contextual information about the dependence of model conclusions on underlying assumptions and parameter choices, supporting robust model design, calibration, validation, and interpretation across scientific, engineering, and statistical domains.

1. Conceptual Framework and Definitions

Sensitivity analysis (SA) quantifies how uncertainty in model outputs can be apportioned or attributed to different sources of uncertainty in model inputs. Two distinct yet complementary questions structure the uncertainty modeling workflow:

Uncertainty Analysis (UA): "How uncertain is the prediction?" This involves propagating input uncertainty through the model to obtain summary statistics (mean, variance, quantiles) of the output.
Sensitivity Analysis (SA): "Which uncertain input factors are responsible for the uncertainty in the prediction, and by how much?" Here, the goal is to decompose output variance or performance metrics into contributions from each input and their interactions (Saltelli et al., 2017).

A rigorous sensitivity analysis explores the full hypercube of plausible input values, not just one-factor-at-a-time (OAT) perturbations, ensuring identification of interaction effects and nonlinearity. Local sensitivity examines infinitesimal perturbations at a fixed input point (typically via derivatives or Jacobians), whereas global sensitivity assesses input importance integrated over the full domain (typically variance-based or moment-independent measures) (Fröhler et al., 2022). Distinction from uncertainty analysis is essential: UA maps output distributions; SA apportions their uses or variances back to inputs.

2. Methodological Approaches

2.1 Local Sensitivity

Local (OAT) sensitivity measures the effect of infinitesimal or finite perturbations of individual inputs near a reference point. For deterministic differentiable models $y = f(x)$ , the basic quantity is the derivative $\partial f / \partial x_i$ at $x^*$ or its normalized variant $\phi_i = (x_i^* / f(x^*)) (\partial f / \partial x_i)$ , interpreted as the output’s percent change per percent change in input (Bolker et al., 2021).

2.2 Global Sensitivity

Global sensitivity analysis aggregates effects over the entire admissible input space, capturing interaction and nonlinear dependencies. The dominant paradigm is variance decomposition:

Sobol’ Indices: The ANOVA (variational) decomposition of $f(x)$ produces partial variances $D_i = \text{Var}_{x_i}[\mathbb{E}_{\sim x_i} f(x)]$ . The first-order Sobol’ index $S_i = D_i / \text{Var}(f(x))$ measures the main effect of $x_i$ ; the total index $S_i^{\text{tot}}$ aggregates all terms (main and interaction) involving $x_i$ (Dimov et al., 2017, Saltelli et al., 2017).
Screening Methods (Morris): For high-dimensional models, the Morris method computes “elementary effects” along randomized OAT trajectories, ranking factors by the mean ( $\partial f / \partial x_i$ 0) and standard deviation ( $\partial f / \partial x_i$ 1) of local differences (Qian et al., 2020).

Global indices are estimated via Monte Carlo, Latin hypercube, or quasi-Monte Carlo (QMC) methods, with recent advances in mixed deterministic-stochastic QMC (“symmetrized shaking”) providing faster convergence for smooth integrands (Dimov et al., 2017).

2.3 Analytic and Surrogate-Based Sensitivity

For differentiable models with known input distributions, analytic Taylor expansions yield explicit formulas for output variance as polynomials in input central moments and derivatives of $\partial f / \partial x_i$ 2 (the “analytic variance-propagation” approach) (Zhu et al., 2016). Polynomial chaos expansions (PCE) serve as surrogate models facilitating rapid sensitivity computation, especially when inputs are correlated or models are computationally expensive (Kardos et al., 2023).

2.4 Sensitivity in Structured Statistical Models

In discrete probabilistic graphical models, sensitivity analysis leverages the algebraic/monomial structure of atomic probabilities. For Bayesian networks and certain generalizations, output probabilities are rational functions (or multilinear polynomials) of parameters; sensitivity functions and global distance measures (e.g., Chan–Darwiche distance) admit closed-form or polynomial representations (Leonelli et al., 2015, Leonelli, 2018).

2.5 Sensitivity in Causal Inference

Sensitivity analysis underpins robustness assessments to violations of core identification assumptions in causal inference, such as unmeasured confounding, selection bias, or ignorability. Notable frameworks include:

Bounding Factor Approaches: Sharp analytical bounds for the influence of unobserved confounders based on maximal exposure–confounder and confounder–outcome relative risks (Ding et al., 2015).
Marginal Structural Models: Sensitivity models for deviations from ignorability in MSMs parameterized by sensitivity parameters (e.g., odds ratio, outcome shift, or contamination fraction), with efficient estimation procedures (Bonvini et al., 2022).
Transportability and Partial Identification: Exponential-tilt and change-of-measure models quantify the impact of assumption violations when generalizing from one population to another, yielding bounds on estimands such as risk or average treatment effect (Dorn et al., 2023, Steingrimsson et al., 2023, Huang, 2022).

3. Computational Algorithms and Practical Computation

Efficient computation is central in large-scale models or high-dimensional Bayesian networks. Notably:

Variance-Based Indices: Standard estimation uses the Saltelli sampling scheme: two independent design matrices, construction of “hybrid” samples replacing individual columns, and estimation via sample variances and covariances (Saltelli et al., 2017, Qian et al., 2020).
Junction Tree Sensitivity in Bayesian Networks: For discrete Bayesian networks, coefficients of sensitivity functions for all marginals with respect to all parameters can be computed via at most three belief propagations in a junction tree (one inward and two outward), yielding all relevant quotient-of-linears forms for OAT analysis. This is orders-of-magnitude more efficient than naive re-evaluation for each parameter (Kjærulff et al., 2013).
Surrogate Modeling: Polynomial chaos expansion (PCE) surrogate fitting, especially in the presence of correlated or non-Gaussian inputs, enables efficient computation of both “Full” and “Independent” indices via change-of-variables and basis permutation (Kardos et al., 2023).
Visualization: Advanced computational environments combine matrix, spatial, and constellation plot views to visually summarize local and global sensitivities, supporting complex algorithmic pipelines and interactive diagnostic evaluation (Fröhler et al., 2022).

4. Applications and Empirical Findings

Sensitivity analysis is foundational in model calibration, validation, uncertainty quantification, model reduction, and policy development across disciplines.

High-Dimensional PDE Models: In environmental modeling, such as pollution forecasting, variance-based SA quantifies contributions of input uncertainties, with hybrid QMC/Monte Carlo yielding lower estimator variance even in non-smooth regimes (Dimov et al., 2017).
Bayesian Neural Networks: Decomposition of predictive uncertainty into epistemic and aleatoric components, with sensitivity defined as the derivative of uncertainty with respect to features, allows modelers to identify inputs responsible for reducible (data-limited) versus irreducible (inherent noise) uncertainty. Such decomposition aligns well with physical understanding and guides data acquisition (Depeweg et al., 2017).
Causal Mediation and Transportability: Sensitivity of decomposed causal effects (e.g., mediated health disparities) to unmeasured mediator–outcome confounding is quantitatively assessed using regression coefficient or $\partial f / \partial x_i$ 3 parameterizations, providing contour plots and robustness values to describe the minimum confounder strength required to overturn conclusions (Park et al., 2022, Huang, 2022).
Principal Stratification and LATE: In noncompliance and instrumental variable analyses, multifaceted sensitivity models (odds ratio, mean ratio, standardized mean difference) have been developed, allowing estimator modification and nonparametric inference under deviations from untestable assumptions (Nguyen et al., 2023).

5. Limitations and Common Pitfalls

A systematic review of published practice demonstrates that a large fraction of sensitivity analyses fail to explore the full input space, relying on OAT approaches inappropriate for nonlinear or high-dimensional models, resulting in severe underestimation of interaction effects and misleading input importance rankings (Saltelli et al., 2017). Sobol’ indices require input independence unless corrected. In the presence of correlated inputs, both variance-based and derivative indices must be adapted to account for changed marginal and conditional contributions (Kardos et al., 2023).

Gradient-based (local) methods capture only first-order and local behavior, missing higher-order, nonlinear, or global effects, and their interpretability is limited in strongly nonlinear or interaction-rich models. The computational cost of standard global methods scales poorly with dimensionality unless surrogates or dimensionality reduction techniques are employed.

6. Best Practices and Future Directions

Consensus best practices, synthesized across foundational and applied literature, recommend:

Comprehensive Exploration: Use global (variance-based or screening) methods to ensure proper exploration of the input space before interpreting input importance or fixing parameters (Saltelli et al., 2017, Dimov et al., 2017).
Combined Local and Global SA: Employ local derivative-based analysis for calibration and local diagnostics, supplementing with global methods for overall input attribution (Bolker et al., 2021).
Design of Experiments: Use stratified, quasi-random, or Latin hypercube sampling to efficiently cover the multidimensional parameter space.
Reporting and Benchmarking: Visualize both uncertainty and sensitivity outcomes; report point estimates and robustness or bias contours; relate sensitivity parameter magnitudes to benchmarks derived from observed covariates (Park et al., 2022, Dorn et al., 2023).
Correlated Inputs: Explicitly model input dependencies; compute both Full and Independent global indices and interpret accordingly (Kardos et al., 2023).
Model Structure: For structured probabilistic models, exploit algebraic or polynomial properties for analytic or highly efficient sensitivity computation (Leonelli et al., 2015, Leonelli, 2018).

Key future directions include development of higher-order, global, and model-agnostic indices for sensitivity to specific uncertainty types (e.g., epistemic versus aleatoric), algorithms for high-dimensional models with complex input dependencies, and integrated workflows coupling sensitivity analysis with active data acquisition or optimization (Depeweg et al., 2017, Dimov et al., 2017).

References:

Sensitivity analysis for Bayesian neural networks (Depeweg et al., 2017)
Monte Carlo and QMC with “symmetrised shaking” (Dimov et al., 2017)
Analytic multivariate variance propagation (Zhu et al., 2016)
Visualization-driven global/local sensitivity in parameter space (Fröhler et al., 2022)
Sensitivity analysis in integrated assessment climate-economy models (Bolker et al., 2021)
Assumption-free bounding-factor sensitivity in causal inference (Ding et al., 2015)
Global sensitivity in MSMs (Bonvini et al., 2022)
Systematic review and best practice guide for SA (Saltelli et al., 2017)
Algebraic/probabilistic graphical sensitivity analysis (Leonelli et al., 2015, Leonelli, 2018)
Correlated input adaptations (Kardos et al., 2023)
Efficient network sensitivity propagation (Kjærulff et al., 2013)
Causal decomposition/mediation sensitivity to confounding (Park et al., 2022)
Global SA methodology review in biomedicine (Qian et al., 2020)
Sensitivity for principal stratification violations (Nguyen et al., 2023)
Sensitivity to exchangeability in prediction transport (Steingrimsson et al., 2023)
Linear estimator identification-failure bounds (Dorn et al., 2023)
ADE under continuous exposure and marginal odds-ratio bounds (Zhang, 9 Nov 2025)