Sometimes the hardest uncertainty is not “what value should this parameter have?” It is “are we even using the right kind of model?”
Structural uncertainty arises when several different model forms, process descriptions or causal structures remain plausible. The uncertainty lives in the architecture of the explanation itself.
A flood model may disagree over how runoff should be represented. An economic forecast may differ depending on whether behaviour changes smoothly or through regime shifts. A disease model may use different assumptions about contact structure. A railway model may treat passenger dwell as a simple function of crowd size or as an interaction between crowd distribution, doors, interchange flow and platform geometry.
This is a specialist branch beneath How Models Work and How Uncertainty Works. Parameter uncertainty asks which values belong inside one fixed model. Structural uncertainty asks which model deserves to be fixed in the first place.
One Dataset Can Support Several Plausible Stories
Different models can fit the same observed data reasonably well and diverge sharply when asked to predict new conditions.
This happens because historical data constrain only part of the possible mechanism. Several structures can pass through the observed points while making different assumptions about what happens outside them.
The danger appears when one model is selected early and all later uncertainty is calculated as though that structural choice were certain.
Model Selection Does Not Make Structural Uncertainty Disappear
Statistical model selection can compare candidate models using information criteria, cross-validation or other objectives. That is useful. It does not prove the winning model is the true structure of the world.
The winner may simply be the best among the candidates considered.
The formal owner remains How Statistical Model Selection Works. Structural uncertainty asks what remains uncertain even after selection.
Different Structures Can Produce Different Decisions
If all plausible models recommend the same action, structural uncertainty may matter scientifically without changing the operational decision.
If one plausible model says “safe” and another says “unsafe,” the model-family uncertainty becomes decision-critical.
This is why Wintour House treats uncertainty as a receiver question. The size of the uncertainty matters less than whether it crosses the boundary that changes action.
Ensembles Can Carry Several Models Forward
One response to structural uncertainty is to retain several plausible models rather than force premature certainty.
An ensemble can compare or combine outputs across model families. The spread between models becomes evidence about structural disagreement.
But averaging is not magic. If all models share the same wrong assumption, the ensemble can be confidently wrong together. Diversity of structure matters only when the models genuinely represent different plausible mechanisms.
Scenario Models Are Useful When Probability Is Weak
Sometimes we cannot assign trustworthy probabilities to competing structures.
In that case, scenarios can be more honest: under structure A, this follows; under structure B, that follows. The decision can then be tested for robustness across the plausible worlds without pretending one model has earned a precise probability it has not.
Existing How Sensitivity Analysis Works owns the broader route for alternative assumptions and tipping points.
Structural Uncertainty Often Hides in Boundaries
Models disagree not only about equations, but about what belongs inside the system.
Should a transport model include only trains, or also buses, walking, passenger behaviour and land use? Should a financial model include liquidity feedback, collateral calls and behavioural response, or treat them as external shocks?
Move the boundary and the model structure changes. This connects directly to How System Boundaries Work.
Worked Example: Flood Risk
Two flood models use the same rainfall forecast. One represents drainage capacity with a simplified static rule. Another represents dynamic interaction between rainfall intensity, soil saturation, drains, pumps and tide level.
Both may fit past moderate storms. During an extreme event, they can diverge sharply.
The uncertainty is not merely in rainfall or pump capacity. It is in which model structure captures the dominant feedback under stress.
Worked Example: MRT Crowd Modelling
One model treats station demand as average arrivals per minute. Another represents passenger waves, interchange timing and platform distribution.
Under ordinary demand, both may predict similar throughput. Near saturation, the second structure may show local crowd traps the first cannot represent.
The railway owner remains How MRT Works | It’s Mathematics. Structural uncertainty asks which representation continues to work when the system enters the regime that matters.
A Careful Analogy: Education
A learner is struggling. One explanation says the main constraint is weak retrieval. Another says the concept itself is misunderstood. A third says the language of the question is blocking access.
Those are competing model structures, not merely different estimates of the same parameter. A good teacher uses discriminating tasks to discover which structure better explains the pattern of errors.
How to Carry Structural Uncertainty Responsibly
- List the plausible model families, not only one preferred specification.
- State the assumptions that distinguish them.
- Compare fit and out-of-sample performance.
- Use residuals and domain knowledge to identify missing structure.
- Check whether different models change the decision.
- Carry several models forward where the evidence does not justify collapse to one.
- Use scenarios when precise model probabilities are not defensible.
- Update structural weights or preferences when new evidence arrives.
Read the Mechanism in Three Directions
Forward: several plausible structures → fit and diagnostics → divergent outputs → decision robustness. Backward: start from a decision that changes across models and identify the structural assumption responsible. Across: compare statistician, engineer, policymaker and affected receiver; each may tolerate a different level of unresolved model disagreement.
Structural uncertainty is the admission that the world may fit more than one defensible explanatory architecture — and that choosing one too early can make every later decimal look more certain than the model deserves.
Continue through How Model Misspecification Works, How Statistical Model Selection Works and the master How X Works hub. Next: parameter uncertainty — when we accept the model structure but remain unsure about the values inside it.