A set of techniques used to assess how variation in model inputs, parameters, or assumptions influences model outputs or decision metrics; methods range from local derivatives (one‑at‑a‑time perturbations) to global approaches that vary inputs across their joint distributions to apportion output variance or to map decision boundaries, with the aim of identifying influential parameters and evaluating robustness.

Principle

By systematically varying inputs within plausible ranges or distributions and measuring resulting output changes, sensitivity analysis reveals which inputs drive outcome variability or alter decisions; global methods quantify contributions to output variance and account for interactions, while local methods approximate marginal responsiveness around a baseline.

Demonstration

Illustrative scenario: A health economist evaluates a cost‑effectiveness model. Recognition: key inputs include discount rate, treatment effect size, and unit costs. Action: perform one‑way sensitivity checks on each parameter, then a probabilistic (global) sensitivity analysis sampling joint distributions to compute output distributions and identify parameters explaining most output variance. Consequence: the analyst identifies which assumptions critically affect the incremental cost‑effectiveness ratio and prioritizes data collection accordingly.

Misapplication

Relying solely on one‑at‑a‑time (OAT) analyses when inputs interact, leading to underestimated joint effects; choosing implausible ranges that misrepresent real uncertainty; interpreting low sensitivity as proof of parameter irrelevance without considering model misspecification.

Consequence

Sensitivity analysis guides interpretation of model results, highlights critical assumptions, informs data‑collection priorities and robustness checks, and supports transparent reporting of model uncertainty; however, results depend on chosen ranges, distributions, and the correctness of the model structure.

Reversal

If the model is misspecified (wrong functional form, omitted mechanisms) or inputs are endogenous, sensitivity analysis can mislead by attributing output variability to parameters while structural errors dominate; local measures may be uninformative for nonlinear models with interactions, requiring global techniques instead.

Boundary

Clearly within: quantitative models with parameterized inputs where ranges or distributions can be credibly specified. Boundary case: structural sensitivity (changing model form) goes beyond parametric sensitivity. Clearly outside: purely qualitative judgments without systematic variation are not sensitivity analysis.

Semantic Tension

Comprehensiveness (global, high‑dimensional exploration) versus interpretability and feasibility (simplicity of one‑at‑a‑time checks); diagnosing parameter importance versus diagnosing structural model adequacy.

Synthesis

Sensitivity analysis is a diagnostic and prioritization tool: it does not validate a model but shows where uncertainty matters and where effort to reduce uncertainty or to improve model structure will most improve decision confidence; appropriate method choice (local vs global) depends on nonlinearity, interactions, and available information about input distributions.