Stochastic Sensitivity Analysis¶
Test or assessment — instantiates Monte Carlo Uncertainty Exploration
Analyzes simulated runs to identify which uncertain inputs or assumptions dominate outcome variation.
Stochastic Sensitivity Analysis is a diagnostic that runs after the simulation to answer a different question than any of the generators: not what the outcome distribution is, but which uncertainties are responsible for its spread. Its identity is attribution — partitioning the variance of the output across the inputs that fed it, so a wide, worrying distribution can be traced to the two or three inputs actually driving it. Crucially, it does this stochastically, across the full sampled run set, so it captures interactions — inputs that matter only in combination — that a one-at-a-time check would miss entirely. Its product is a ranking, sometimes a budget: input A explains 55% of the outcome variance, input B another 30%, and the rest barely register. That ranking is what tells a team where reducing uncertainty (better data, a tighter distribution) would actually shrink the risk, and where it would be wasted effort.
Example¶
A developer models the annual revenue of a proposed wind farm. The output distribution is alarmingly wide, and the instinct is to gather better data everywhere. Stochastic Sensitivity Analysis mines the existing runs and partitions the revenue variance across the inputs: the sampled long-run wind-speed distribution accounts for about 60% of the spread, electricity price about 25%, turbine availability roughly 10%, and maintenance cost almost nothing. The interaction terms reveal that price and wind matter most jointly in low-wind years. The finding redirects the whole effort: a year of on-site wind measurement to tighten that one input distribution would collapse most of the revenue uncertainty, while the meticulous maintenance-cost estimate everyone was polishing is nearly irrelevant. The assessment does not change the model or the forecast; it tells the team which uncertainty is worth buying down.
How it works¶
- Reuse the run set. Operate on the runs a generator already produced, reading each run's inputs and its output together — no new model, just analysis.
- Partition the variance globally. Decompose total output variance into contributions attributable to each input and to their interactions, across the whole sampled range rather than around a single point.
- Trace back to input distributions. Express each contribution against the input's own distribution, so a driver's importance is tied to how uncertain that input actually is, not just how influential it would be if known.
- Check that the ranking is stable. Verify the sensitivity indices themselves have converged, since a driver ranking estimated from too few runs can reorder as more runs arrive.
Tuning parameters¶
- Sensitivity method — variance-based global indices vs. cheaper regression or one-at-a-time screening. Global methods capture interactions but need more runs; screening is fast but blind to coupling.
- Interaction depth — how many joint effects are estimated. Deeper interaction analysis explains more of the variance but multiplies the runs required.
- Grouping — analyzing inputs individually vs. in themed bundles. Grouping simplifies the story but can hide which specific input inside a bundle matters.
- Index-convergence threshold — how stable a ranking must be before it is trusted; loose thresholds risk publishing a driver order that is really sampling noise.
When it helps, and when it misleads¶
Its strength is focus: it turns a diffuse "everything is uncertain" into a short, prioritized list of what to measure, model better, or hedge — and because it works across the full sampled range, it catches drivers that only bite in combination.
Its failure mode is the temptation to economize with one-at-a-time sensitivity, wiggling each input alone around a baseline; that misses interactions and can rank the wrong driver first when effects are coupled, which is exactly why variance-based global indices exist.[n1] The classic misuse is running the analysis on too few simulation runs, so the driver ranking is unstable and reorders on rerun. The guarding discipline is to use a global, interaction-aware method and to confirm the sensitivity indices have converged before acting on their order.
How it implements the components¶
sensitivity_partition— the core output: the decomposition of outcome variance into per-input and interaction contributions, ranked.uncertain_input_distribution— ties each input's importance back to its own distribution, so influence is weighted by genuine uncertainty rather than raw leverage.convergence_diagnostic— checks that the sensitivity indices themselves are stable, not just that the mean has settled.
It does not build the model or its couplings — simulation_model and input_dependency_model are Uncertainty Propagation Model's — and it does not render results for stakeholders (scenario_inspection_view, that's Simulation Result Dashboard).
Related¶
- Instantiates: Monte Carlo Uncertainty Exploration — this is the archetype's sensitivity-attribution diagnostic.
- Consumes: Scenario Sampling Workflow — the retained run set it partitions (or the runs directly from Monte Carlo Simulation Method).
- Sibling mechanisms: Monte Carlo Simulation Method · Uncertainty Propagation Model · Probabilistic Risk Simulation · Scenario Sampling Workflow · Portfolio Risk Simulation · Operational Capacity Simulation · Simulation Result Dashboard
Editorial Notes¶
Form Classification¶
Form family: Analysis, Modeling & Optimization
Rationale: Stochastic Sensitivity Analysis operates as an analytical, modeling, inference, comparison, or optimization procedure that derives insight or a solution because it analyzes simulated runs to identify which uncertain inputs or assumptions dominate outcome variation.
Independent corroboration: The frozen evidence defines Stochastic Sensitivity Analysis as 'Analyzes simulated runs to identify which uncertain inputs or assumptions dominate outcome variation', so its operative form is Analysis, Modeling & Optimization.
Nearest alternative: Assessment, Review & Assurance — Stochastic Sensitivity Analysis includes features of a bounded evaluation of existing evidence or work that produces a finding or disposition, but its defining operation is an analytical, modeling, inference, comparison, or optimization procedure that derives insight or a solution.
Review outcome: Independent reviewer agreement; medium confidence.
Origin Attribution¶
Primary origin: Statistics & Experimental Design
Origin pattern: Single lineage
Present-day reach: Multi-domain
Rationale: Attributing outcome variance across simulated uncertainty is global sensitivity analysis.
Related originating lineages:
- Data Science & Analytics — Simulation analytics estimates drivers.
- Mathematics — Mathematical modeling, proof, and abstract-structure practice supplies a parallel or contributing lineage for the mechanism's defining operation: analyzes simulated runs to identify which uncertain inputs or assumptions dominate outcome variation.
- Operations Research — It guides robust decisions.
Review resolution: The blind reviewers agree that statistics_experimental_design is the primary origin and differ only on alternate origin disagreement, domain reach disagreement. I preserve every independently explained alternate from both records rather than imposing a numeric cap. I retain single_lineage because the combined evidence shows one traceable formative lineage. The broader reach of multi_domain records portability separately from historical provenance; encyclopedia_synthesis=false preserves the affirmative synthesis judgment where either reviewer identified one.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
[n1] Sobol' sensitivity indices decompose the variance of a model output into fractions attributable to each input and to their interactions, estimated globally over the whole input space rather than around one baseline point. They are the reference method for catching interaction-driven importance that one-at-a-time sensitivity checks systematically miss. ↩