Scenario Analysis

Scenario Analysis: Concept Architecture

Orientation and learning roadmap

Scenario analysis asks whether a conclusion remains credible when several linked features of an analysis are changed together. This page explains how to construct coherent scenarios, distinguish them from other uncertainty methods, calculate and interpret health-economic results, and avoid false probability claims. It then sets out design, implementation, validation, and reporting practices for decision models, budget-impact analyses, and policy planning.

A scenario is a complete analytical case

A scenario specifies a deliberate combination of assumptions, parameter values, model choices, or external conditions and then reruns the analysis under that combination. Each scenario should tell one intelligible analytical story rather than collect unrelated favorable or unfavorable values. The outputs belong to the full case and cannot be attributed automatically to any single changed input.

The reference case provides the comparison point

Health-economic analyses usually begin with a reference case that follows the applicable decision problem and methodological requirements. Alternative scenarios are compared with that case to expose the consequences of contested assumptions, implementation choices, structural alternatives, or plausible future conditions. The reference case is not necessarily the most likely forecast, and calling it “base case” should not imply a probability that has not been estimated.

Scenario analysis changes linked assumptions together

Some assumptions are logically, clinically, or operationally connected and should not be varied independently. A scenario might combine faster uptake with greater delivery capacity, or shorter treatment duration with lower acquisition cost and different adverse-event exposure. The combination must preserve causal and accounting relationships so that the scenario remains possible within the model.

Scenarios can address different kinds of uncertainty

Scenario analysis is useful for uncertainty that cannot be represented adequately by a probability distribution around one parameter set. It can examine structural choices, alternative evidence sources, implementation pathways, policy designs, behavioral responses, and future environments. The reason for each scenario should be stated because parameter uncertainty, structural uncertainty, and policy uncertainty require different interpretations.

Parameter scenarios use alternative input sets

A parameter scenario replaces selected numerical inputs with a coordinated set of alternatives. The values may come from another credible study, subgroup, time period, elicitation, or internally consistent low and high case. Each value still needs a source or transparent derivation rather than being selected solely to produce a desired result.

Structural scenarios change how the model works

A structural scenario changes the representation of the decision problem rather than merely substituting new numbers. Examples include alternative survival extrapolations, treatment-sequencing rules, state definitions, waning assumptions, dependency structures, or methods for handling treatment switching. Structural scenarios are especially important when several defensible model forms fit the available evidence but imply different long-term outcomes.

Implementation scenarios describe alternative delivery pathways

Implementation conditions can affect population reach, uptake, adherence, capacity, waiting times, unit costs, and realized effectiveness. Scenario analysis can combine these conditions into credible rollout pathways rather than assuming immediate steady-state use. These scenarios should distinguish constraints that affect affordability from those that change health outcomes or cost-effectiveness.

Future-state scenarios support policy planning

Planning scenarios may combine demographic change, epidemiology, prices, workforce capacity, technology diffusion, and policy responses over time. They are conditional projections of what the model would produce if the stated conditions occurred. They should not be labeled forecasts unless the analysis also supplies a forecasting method, calibration, and evidence about predictive performance.

Scenario analysis differs from one-way sensitivity analysis

One-way sensitivity analysis changes one input at a time while holding other inputs fixed, which helps reveal local parameter influence. Scenario analysis changes a defined bundle of inputs or assumptions because the elements form a meaningful joint case. A scenario result therefore shows the consequence of the bundle, while a one-way result is easier to attribute to the selected parameter.

Scenario analysis differs from probabilistic sensitivity analysis

Probabilistic sensitivity analysis repeatedly samples uncertain parameters from specified joint distributions and estimates the distribution of model outcomes. Scenario analysis evaluates a limited set of named cases and does not create a probability distribution unless defensible probabilities are assigned to the scenarios. Running three scenarios does not mean that each has a one-third probability or that the middle result is an expected value.

Scenario analysis differs from threshold analysis

Threshold analysis solves for the value at which a result or decision changes, such as the price that makes incremental net monetary benefit equal zero. Scenario analysis instead evaluates predefined combinations of assumptions and observes their outcomes. A threshold can strengthen scenario analysis by showing how far each case lies from a decision boundary.

Scenario analysis differs from stress testing

Stress testing deliberately examines severe conditions to identify vulnerability, resilience, or failure points. A stress scenario may be intentionally extreme and need not be the analyst's view of a plausible central future. Ordinary scenario analysis can include stress cases, but their severity, purpose, and lack of probability should be labeled clearly.

Naming should describe the case rather than judge it

Labels such as optimistic, pessimistic, best case, and worst case can hide whose outcome is being optimized and whether the combination is genuinely attainable. Descriptive names such as “rapid uptake with constrained capacity” or “shorter duration with no waning” reveal the assumptions directly. A genuinely favorable or adverse case can still be identified, but the perspective and outcome criterion should be stated.

A worked cost-effectiveness example

Suppose a new intervention is compared with current care at a threshold of £30,000 per QALY. Three scenarios jointly vary incremental QALYs and incremental cost to represent different evidence and delivery assumptions. Incremental net monetary benefit for scenario (s) is:

[ \operatorname{INMB}_s=\lambda\Delta E_s-\Delta C_s, ]

where (\lambda) is the cost-effectiveness threshold, (\Delta E_s) is incremental QALYs, and (\Delta C_s) is incremental cost. Positive incremental net monetary benefit favors the new intervention under the stated threshold and scenario.

The reference-case calculation is (£30{,}000\times0.18-£4{,}200=£1{,}200). The change in decision across scenarios shows that the conclusion depends on the joint evidence-and-cost case, but it does not identify which input caused the switch or how likely either alternative is.

Scenario results can be decomposed carefully

When a scenario changes many inputs, an analyst may run intermediate cases that add changes sequentially or by theme. This decomposition can show which groups of assumptions drive the difference, but the result can depend on the order when the model is nonlinear or inputs interact. A full factorial or designed analysis may be needed when interaction itself is the question.

Coherence matters more than symmetry

Scenarios do not need to sit equal distances above and below the reference case. Evidence may support an asymmetric range, and linked quantities may move in different directions. Artificial symmetry can create combinations that look tidy but violate clinical pathways, accounting identities, or observed dependence.

Correlation should be represented rather than implied

Inputs that share a data source, causal mechanism, or accounting relationship may move together. A coherent scenario can encode the direction and magnitude of that joint movement explicitly. If the relationship is uncertain and probabilistic data are available, a joint distribution in probabilistic sensitivity analysis may be more informative than a few hand-built combinations.

Mutually exclusive model choices need separate runs

Some alternatives cannot coexist in one calculation, such as different extrapolation families or incompatible treatment pathways. Each model form should be implemented as a separate scenario with all dependent calculations updated consistently. Averaging their outputs is not justified unless model probabilities or weights have a defensible basis.

Time horizon and perspective can define scenarios

Alternative time horizons can show how much a conclusion depends on distant costs and outcomes, while alternative perspectives can include different resource consequences. These are not simple parameter changes because they alter which events or costs enter the analysis. Results should remain labeled by horizon and perspective so they are not mistaken for interchangeable estimates of the same quantity.

Subgroup analysis should not be relabeled casually

A subgroup analysis estimates results for a population with defined characteristics and may require subgroup-specific baseline risks, effects, utilities, and costs. A scenario that substitutes one subgroup's parameters into an overall-population model may not reproduce a valid subgroup analysis. Population definition, treatment-effect evidence, and interaction claims therefore need separate justification.

Budget-impact scenarios need operational consistency

Budget-impact analysis often uses scenarios for uptake, market share, eligible population, treatment duration, displacement, and service capacity. Those inputs must reconcile across periods and comparators so that patients, treatments, and costs are neither omitted nor double counted. A high-uptake case should also reflect the capacity, timing, and resource consequences required to deliver that uptake.

Expected values require probabilities

An expected result across scenarios can be calculated only when mutually exclusive and collectively relevant scenarios have defensible probabilities (p_s). The probabilities must sum to one, and the expected output is a probability-weighted quantity rather than a simple average. If probabilities are not defensible, the scenarios should remain separate.

[ \operatorname{E}[Y]=\sum_{s=1}^{S}p_sY_s, \qquad \sum_{s=1}^{S}p_s=1. ]

Assigning subjective probabilities introduces another evidence and judgment requirement. The weights, elicitation method, calibration, and sensitivity to alternative weights should be reported.

Decision robustness is the central interpretation

Scenario analysis is most useful when it shows whether the preferred option, affordability conclusion, ranking, or policy action changes. Stable decisions across credible scenarios strengthen robustness, while a switch identifies conditions that deserve scrutiny or risk management. Stability across poorly designed scenarios does not demonstrate that all material uncertainty has been addressed.

Scenario ranges are not confidence intervals

The minimum and maximum results across selected scenarios depend on which cases the analyst chose. They do not have a sampling coverage probability and should not be reported as a 95% interval. A scenario range is a conditional envelope over named cases, not a statistical confidence or credible interval.

Scenarios should be selected before results are known

Prespecifying the purpose, assumptions, and evidence for each scenario reduces the risk of choosing only cases that support a preferred conclusion. Additional post hoc scenarios may still be useful when review identifies a gap, but they should be labeled as such. The report should include unfavorable as well as favorable credible cases when both are decision-relevant.

A scenario register makes the analysis auditable

Each scenario should have a stable identifier and one authoritative record of its changes. The register separates human-readable rationale from machine-readable inputs and allows reviewers to reproduce the run. It should include enough detail to distinguish an input override from a structural or code change.

Spreadsheet implementation should avoid hidden overrides

A spreadsheet can store one column per scenario in a controlled assumptions table, with a single scenario selector feeding one authoritative model build. Every scenario column should contain a complete validated input set or an explicit link to the reference value. Pasting alternative values directly into calculation sheets makes provenance, resetting, and comparison unreliable.

Scenario tables need independent checks

The implementation should verify that scenario names and IDs are unique, required inputs are present, units are consistent, probability weights are valid when used, and dependent totals reconcile. Automated checks should also confirm that selecting each scenario changes the intended inputs and restores the reference case without residue. A stored output snapshot is not evidence that the live model still reproduces the result.

A disciplined scenario-analysis workflow

Strong scenario analysis begins with a decision-relevant uncertainty and ends with a conditional interpretation. It does not begin by creating arbitrary low, central, and high columns. The following sequence preserves meaning and reproducibility.

1. Define the decision. State the population, alternatives, perspective, horizon, outcomes, and decision criterion.
2. Identify material uncertainty. Separate parameter, structural, implementation, policy, and future-state questions.
3. Write the scenario rationale. Explain the mechanism that makes the combined changes coherent.
4. Source every change. Record evidence, derivation, expert judgment, units, and applicable period.
5. Map dependencies. Update linked values, constraints, pathways, formulas, and accounting identities together.
6. Prespecify outputs. Include decision outcomes as well as intermediate diagnostics that explain the case.
7. Run and validate. Reconcile calculations, compare with the reference case, and inspect unexpected behavior.
8. Test the boundary. Use one-way, threshold, probabilistic, or interaction analysis where the scenario result raises a narrower question.
9. Interpret conditionally. State what would follow if the scenario assumptions held without implying an unsupported probability.
10. Report completely. Publish the register, results, model version, and any post hoc changes.

Common errors and safeguards

Scenario analysis becomes misleading when labels replace explicit assumptions or when selected cases are treated as a probability distribution. Strong safeguards preserve coherence, completeness, and conditional interpretation. They also separate genuine uncertainty exploration from advocacy.

• Changing several unrelated inputs together creates a bundle that has no defensible interpretation.
• Moving every input in the same favorable direction can construct an impossible best case.
• Calling a scenario “likely” without probability evidence turns judgment into an unsupported forecast.
• Averaging scenario outputs without defensible weights creates a false expected value.
• Reporting the scenario range as a confidence interval assigns statistical meaning it does not have.
• Changing a structural assumption without updating dependent equations produces an internally inconsistent model.
• Substituting subgroup inputs into an overall model can misrepresent a genuine subgroup analysis.
• Omitting unfavorable credible cases can bias the presentation of decision robustness.
• Using scenario analysis alone can conceal which parameter or interaction drives a changed result.
• Treating identical decisions across a few scenarios as proof of robustness can overlook untested uncertainty.
• Hard-coding overrides in calculation sheets can leave residual values when the reference case is restored.
• Comparing scenarios from different model versions can attribute code changes incorrectly to assumptions.

Reproducible reporting supports accountable decisions

A complete report presents the reference case and every material scenario in a common table with assumptions, sources, outputs, and decision consequences. It distinguishes prespecified from post hoc cases and separates scenario analysis from probabilistic, one-way, threshold, subgroup, and stress analyses. The conclusion should identify which combinations change the decision, which remain plausible, what evidence would resolve the uncertainty, and where expert judgment remains necessary.