Probabilistic Sensitivity Simulation¶
Simulation — instantiates Sensitivity Analysis Protocol
Draws thousands of joint samples from input distributions and reports the share of draws in which the recommendation holds versus flips.
Probabilistic sensitivity simulation stops moving inputs by hand. Instead it assigns each uncertain input a probability distribution, draws thousands of joint samples, re-runs the model on each, and reports not a swing but a frequency: in what share of plausible draws does the recommendation hold, and in what share does it flip. Its defining output is a probability attached to the conclusion — "the option remains preferred in 82% of simulated draws" — which turns "is this robust?" into a countable answer rather than a set of illustrative cases. Note the boundary carefully: it counts decision reversals; it does not partition the outcome's variance to name which single input dominates the spread. That attribution question is a different diagnostic; this mechanism's job is how often the answer changes.
Example¶
A national health-technology body must decide whether to reimburse a new drug, and the rule is simple: fund it if the cost per quality-adjusted life-year comes in under a stated willingness-to-pay threshold. The base-case estimate clears the threshold, but almost every input is uncertain — the drug's effect size, the comparator's real-world adherence, unit costs, and how long the benefit persists. Rather than argue each assumption, the analysts assign every input a distribution and run a Monte Carlo simulation of tens of thousands of joint draws, recomputing cost-effectiveness on each. The result is a distribution of outcomes and a single decision-relevant number: the drug falls under the threshold in 78% of draws. Sweeping the threshold produces an acceptability curve showing how that share rises and falls. The committee no longer debates a point estimate; it weighs a 78% probability of cost-effectiveness — and can see that at a slightly lower threshold the case collapses.
How it works¶
- Turn ranges into distributions. Give each uncertain input a probability density — normal, triangular, beta — rather than a two-point low/high.
- Sample jointly and re-run. Draw thousands of combined samples (respecting known correlations) and evaluate the model on every draw.
- Collect the outcome distribution. Assemble the full spread of results across draws, not just a central estimate.
- Count the flips. Report the fraction of draws that cross the decision boundary — the probability the recommendation holds versus reverses.
Tuning parameters¶
- Distribution shape — how each range becomes a density; this is the highest-leverage and most-abused dial, since an invented shape drives the headline probability.
- Sample count — how many draws; more tightens the estimate and costs compute, too few leaves the reported probability itself noisy.
- Correlation structure — whether inputs are sampled independently or with a dependency structure; independent sampling of correlated inputs manufactures impossible combinations.
- Decision boundary — the threshold used to classify each draw as hold-or-flip; moving it reshapes the acceptability result.
When it helps, and when it misleads¶
Its strength is that it integrates many uncertainties at once and delivers a single defensible probability that a decision-maker can weigh directly, catching combination effects a hand-run sweep would miss. Its failure mode is laundering weak assumptions into false rigor: invented distributions produce a confident-looking probability that merely re-expresses the analyst's priors, and a smooth acceptability curve can hide that every density behind it was a guess.[n1] The classic misuse is presenting a simulation probability as though it were empirical when the input distributions were assumed. The guarding discipline is to publish the distributions and their sources, treat the probability as explicitly conditional on them, and run an informal check on whether those input shapes are themselves defensible.
How it implements the components¶
variation_range— extends each input's range into a full probability distribution, the probabilistic form of a variation range.outcome_response— produces the distribution of outcomes across all draws, not a single swing.robustness_conclusion— states robustness as a probability the conclusion holds under plausible combined uncertainty.decision_flip_point— quantifies the share of draws that cross the decision boundary, the flip expressed as a frequency.
It does not rank single drivers or run one input at a time (parameter_prioritization_rule, One-way Sensitivity Analysis), and it does not reduce uncertainty to a few coherent named worlds (scenario_set, Scenario Variation) — it uses many random draws instead of a handful of stories. Unlike Threshold Analysis, which solves for the single value where the answer flips, this counts how often random draws land past that boundary.
Related¶
- Instantiates: Sensitivity Analysis Protocol — this is the protocol's distribution-sampling, flip-frequency simulation.
- Consumes: Assumption Stress-test Workshop — the input distributions are often elicited and defended there.
- Sibling mechanisms: One-way Sensitivity Analysis · Two-way or Multi-way Sensitivity Analysis · Scenario Variation · Threshold Analysis · Tornado Chart · Sensitivity Table · Assumption Stress-test Workshop
Editorial Notes¶
Form Classification¶
Form family: Analysis, Modeling & Optimization
Rationale: Probabilistic Sensitivity Simulation operates as an analytical, modeling, inference, comparison, or optimization procedure that derives insight or a solution because it draws thousands of joint samples from input distributions and reports the share of draws in which the recommendation holds versus flips.
Independent corroboration: The frozen evidence defines Probabilistic Sensitivity Simulation as 'Draws thousands of joint samples from input distributions and reports the share of draws in which the recommendation holds versus flips', so its operative form is Analysis, Modeling & Optimization.
Review outcome: Independent reviewer agreement; high confidence.
Origin Attribution¶
Primary origin: Statistics & Experimental Design
Origin pattern: Cross-disciplinary synthesis
Present-day reach: Multi-domain
Rationale: Probabilistic Sensitivity Simulation is most plausibly rooted in the statistics_experimental_design tradition because its characteristic form depends on probability, calibrated inference, experimental design, and uncertainty analysis. The assignment tracks that formative lineage, not the many settings in which the mechanism can now be applied.
Related originating lineages:
- Operations Research — The operations_research tradition materially shaped Probabilistic Sensitivity Simulation through its own practice of queueing, optimization, scheduling, prioritization, and constrained allocation.
Review resolution: Both blind reviewers agree that statistics experimental design is the primary origin. Explicit reconciliation resolves origin mode disagreement. Formative alternate lineages are retained as operations_research; later breadth of use is recorded separately as domain_reach=multi_domain, while origin_mode=cross_disciplinary_synthesis describes the relationship among origin lineages.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
Distinct from the identically-named-sounding stochastic sensitivity analysis under Monte Carlo Uncertainty Exploration: that mechanism partitions output variance to name which input dominates the spread; this one counts how often the decision reverses across sampled draws. Same simulation raw material, opposite question — attribution there, flip-frequency here.
[n1] A cost-effectiveness acceptability curve (CEAC), standard in health-economic evaluation, plots the probability that an intervention is cost-effective as a function of the willingness-to-pay threshold, computed from a probabilistic (Monte Carlo) sensitivity analysis over uncertain model inputs. It is a clean example of reporting a decision as a probability rather than a point estimate — and of a result only as trustworthy as the input distributions behind it. ↩