Global Sensitivity Analysis — Sobol, Morris, and FAST
- Introduced by
- I.M. Sobol (indices, 2001); Morris (screening, 1991); Cukier et al. (FAST, 1973)
- Year
- 1973–2001
- Type
- Variance-based sensitivity decomposition
What does it do?
Global sensitivity analysis (GSA) is a family of techniques that decompose the variance of a model's output across its input parameters, quantifying how much each input — and each combination of inputs — contributes to the total uncertainty in the result.
Sobol's variance-based indices (2001), Morris's one-at-a-time (OAT) screening (1991), and the Fourier Amplitude Sensitivity Test (FAST, first proposed by Cukier et al. in 1973) are the three most widely used approaches. Together they serve as the standard toolkit for identifying which parameters drive model behaviour and which can be safely fixed.
The intuition behind the method
Imagine a model with ten input knobs. You want to know which knobs matter and which barely affect the output. A local sensitivity check only tells you what happens near one particular setting; global sensitivity analysis explores the whole plausible space of settings at once. Sobol's approach is the most thorough: it allocates a fraction of the total output variance to each input and to every combination of inputs. Morris is a cheaper screening method that ranks parameters by their overall influence without fully decomposing the variance. FAST uses a Fourier transform trick to achieve variance decomposition with fewer model runs than Sobol. All three methods answer the same fundamental question — which inputs are the important ones — at different levels of cost and precision.
History
The decomposition of model output variance into contributions from individual inputs has roots in the ANOVA tradition, but its systematic application to computational models began with the Fourier Amplitude Sensitivity Test introduced by Cukier, Levine, and Shuler in the 1970s. Max Morris formalised the one-at-a-time elementary-effects screening approach in a 1991 paper that remains widely cited for its practical economy. The most influential modern formulation came from Ilya M. Sobol, whose 2001 paper established the Monte Carlo estimators for first-order and total-effect variance-based indices that are now the standard. The textbook by Saltelli and colleagues (2008) consolidated the field and established the Saltelli sampling scheme as the reference implementation for Sobol indices. Today, global sensitivity analysis is a routine component of model validation in engineering, environmental science, epidemiology, and economics.
When to use it
Global sensitivity analysis is appropriate whenever you need to understand which inputs drive the uncertainty in a computational model, or when you want to simplify a model by identifying non-influential parameters that can be fixed. It applies to any deterministic or stochastic model where inputs can be described by distributions or plausible ranges. Sobol indices are the most informative choice when the number of model runs is affordable (typically hundreds to thousands per parameter). Morris screening is preferred for initial exploration when the model is expensive and there are many candidate parameters to rank quickly. FAST is a good middle ground when Sobol's sample cost is prohibitive but full variance decomposition is desired. The method does not require normally distributed inputs; continuous input types are expected. It is not suitable when inputs cannot be assigned probability distributions or plausible ranges, or when the model is purely data-driven without a forward simulation structure.
Limitations and considerations
- Sobol indices require N×(2k+2) model runs, which can be prohibitive for expensive simulators with many parameters.
- Morris screening reveals the ranking of parameter importance but does not decompose variance into quantified fractions; it cannot replace Sobol indices when precise apportionment is needed.
- All methods assume that input uncertainty can be characterised by probability distributions or bounded ranges; if this prior knowledge is unavailable or unreliable, the resulting indices reflect assumed rather than true uncertainty.
- For models with discrete or categorical inputs, the standard variance-based framework does not apply directly and extensions or alternative approaches are required.
Source conditions do not replace independent scientific review or validation by an analysis engine.
Strengths
- Explores the entire input space simultaneously rather than making local linearisation assumptions, capturing nonlinear behaviour and parameter interactions.
- Sobol total-effect indices (ST) directly identify non-influential parameters that can be fixed, enabling principled model simplification.
- The Morris method is extremely cheap relative to full variance decomposition, making it practical for high-dimensional or computationally expensive models.
- FAST achieves full variance decomposition with fewer runs than Saltelli-Sobol sampling, offering an efficient compromise.
- Results are interpretable: sensitivity indices are normalised fractions of total variance, enabling straightforward ranking and communication.
Applications
- Identifying which model parameters most influence predictions in climate and earth-system models, directing calibration efforts toward the parameters that matter.
- Simplifying complex pharmacokinetic and pharmacodynamic models by fixing non-influential parameters before fitting to clinical data.
- Prioritising uncertainty reduction in risk and safety assessments for nuclear, chemical, or structural engineering systems.
- Guiding experimental design in computational biology and systems modelling by revealing which input factors most drive variability in model outputs.
- Supporting policy analysis by quantifying which uncertain input assumptions most affect the conclusions of economic or epidemiological models.
How it works
How it works
The analyst specifies the input parameter space: each parameter is assigned a probability distribution (or a plausible range) that captures prior knowledge about its uncertainty. The model whose output is to be analysed is also defined at this stage. Careful choice of input ranges is critical because sensitivity indices are relative to the space that is explored.
A structured set of input combinations is generated according to the chosen method. For Sobol indices, the Saltelli sampling scheme produces N×(2k+2) model evaluations for k parameters; it uses quasi-random (low-discrepancy) sequences to cover the input space efficiently. The Morris method uses a one-at-a-time trajectory design that varies one parameter at a time while holding others fixed. FAST uses a periodic sampling scheme built on distinct frequencies assigned to each input.
The computational model is evaluated at every sampled input point. This stage can be the most expensive part of the analysis if the model is computationally costly. For expensive simulators, surrogate models (metamodels) such as Gaussian process emulators are sometimes substituted to make the required number of evaluations feasible.
The model outputs are analysed to compute sensitivity measures. For Sobol, the total output variance is partitioned: the first-order index S1 captures the fraction of variance due to a parameter acting alone, and the total-effect index ST captures its contribution including all interactions with other parameters. Morris computes an elementary effect for each parameter at each trajectory point; the mean and standard deviation of these effects indicate overall influence and nonlinearity or interaction. FAST uses Fourier analysis of the output variation to attribute variance to each input frequency.
Parameters are ranked by their sensitivity indices. High S1 suggests a parameter has a strong direct effect; a large gap between S1 and ST indicates significant interaction with other parameters. Parameters with near-zero ST can be fixed at nominal values without materially affecting the output distribution — this is the key practical outcome. Results are typically displayed as bar charts or scatter plots of indices with bootstrap confidence intervals.
Common pitfalls
- Specifying input ranges that are too narrow artificially reduces the apparent sensitivity of parameters that would matter across a broader realistic range — always ground input ranges in domain knowledge or data.
- Reporting only first-order indices (S1) and ignoring total-effect indices (ST) misses interaction effects; a parameter with low S1 but high ST is still important because it interacts strongly with others.
- Using too few model runs for Sobol indices produces unstable, unreliable index estimates; bootstrap confidence intervals should always be computed to assess convergence.
- Applying a Morris screening result as a final answer when the purpose requires precise variance attribution — Morris ranks parameters but does not quantify their fractional contributions; use Sobol or FAST for that.
- Confusing local sensitivity analysis (derivatives at a single point) with global sensitivity analysis — local analysis does not capture parameter interactions or nonlinearity across the full input space.
Frequently asked questions
Should I use Sobol, Morris, or FAST?
The choice depends on your budget of model runs and your goal. Start with Morris if the model is expensive and you have many parameters to screen — it ranks them cheaply. If you can afford N×(2k+2) runs (where k is the number of parameters), Sobol indices give the most complete picture by quantifying first-order and interaction effects as fractions of total variance. FAST is a good option when you want full variance decomposition but Sobol's sample size is too large; it achieves similar information with fewer runs.
What do first-order (S1) and total-effect (ST) indices actually mean?
S1 for a parameter is the fraction of the total output variance explained by that parameter alone, ignoring interactions. ST is the fraction explained by that parameter and all its interactions with other parameters. If S1 ≈ ST for all parameters, interactions are negligible. A large gap between S1 and ST signals important interaction effects. Parameters with ST near zero can be fixed without meaningfully changing the output uncertainty.
How many model runs do I need?
The standard Saltelli scheme for Sobol indices requires N×(2k+2) runs, where N is a base sample size (often 512–2048) and k is the number of input parameters. For example, with k=10 parameters and N=1024, that is 22 528 runs. Morris typically needs r×(k+1) runs, where r is the number of trajectories (often 10–50), giving a much smaller total. FAST requirements vary by implementation but are generally between those of Morris and Sobol.
Can I use global sensitivity analysis on a model that is not a simulation?
GSA applies to any model that maps a set of inputs to outputs and can be evaluated repeatedly at different input values. This includes statistical models, machine-learning models, and analytical equations, not only physics-based simulations. The requirement is that inputs can be assigned plausible distributions or ranges and the model can be run enough times to populate the sample matrix.
Related methods
Sources
Sobol, I.M. (2001). Global Sensitivity Indices for Nonlinear Mathematical Models and Their Monte Carlo Estimates. Mathematics and Computers in Simulation, 55(1–3), 271–280.
Open DOI referenceCopied.Saltelli, A. et al. (2008). Global Sensitivity Analysis: The Primer. Wiley.
Open DOI referenceCopied.