Global Sensitivity Analysis

Sensitivity analysis techniques assessing how a model's output responds to simultaneous variation across the full range of all uncertain inputs, unlike one-way analysis.

Source: Saltelli et al. 2008

One-way analysis moves a single input at a time, so it cannot detect effects that arise only when several inputs change together, nor does it weight inputs by how uncertain they actually are. Global sensitivity analysis varies all inputs simultaneously across their full plausible ranges, so it captures interactions between them and reflects their real uncertainty in apportioning influence. This gives a truer picture of what drives the output when everything is uncertain at once. It sees the combined behaviour a one-way sweep misses. Saltelli and colleagues (2008) describe these methods.

Source: Saltelli et al. 2008

Global sensitivity analysis varies all inputs simultaneously across their full ranges, capturing interactions and non-linear effects, whereas local sensitivity analysis examines the effect of small changes in one input around a fixed base-case point, missing interactions and behaviour away from that point. Global methods explore the whole input space and can attribute output variance to inputs individually and jointly, while local methods give a limited, point-based view. The difference is between whole-range, simultaneous variation and small, one-at-a-time variation around a single point.

Source: Saltelli et al. 2008

Global sensitivity analysis uses methods such as variance-based indices, including first-order and total-effect Sobol indices, which decompose output variance among inputs and their interactions; screening methods like the Morris method using elementary effects for efficient identification of influential inputs; and Monte Carlo filtering for factor mapping. These methods differ in cost and detail, from cheap screening to full variance decomposition. They share the feature of varying all inputs across their ranges, allowing influence, interactions, and the conditions producing particular outputs to be assessed.

Source: Sobol 2001

Global sensitivity analysis is used because one-way, local methods can miss interactions and misjudge influence when inputs vary over wide ranges or act jointly, whereas exploring the full input space with all inputs varying together gives a more reliable picture of what drives output uncertainty. It supports prioritising influential inputs, fixing negligible ones, and identifying conditions producing particular outcomes. Where models are non-linear or inputs interact, global sensitivity analysis provides a sounder basis than local analysis for understanding and communicating the drivers of uncertainty.

Source: Saltelli et al. 2008

Global sensitivity analysis, especially variance-based methods, can be computationally intensive, requiring many model runs, which is demanding for complex or slow models, though screening methods reduce this cost. Its results depend on the assumed ranges and distributions of the inputs, so these must be chosen carefully. Interpreting joint and interaction effects can be complex. These limitations mean the method and effort are matched to the model, with input assumptions considered, and cheaper screening sometimes used before fuller variance-based analysis.

Source: Saltelli et al. 2008