EART66001 Chap.8 Data Quality and Model Uncertainty
Data Quality and Model Uncertainty
Data Quality and Model Uncertainty explains how evidence limitations move through a model into a decision. Data quality includes spatial and temporal coverage, resolution, units, datum, completeness and representativeness, not accuracy alone. Model assumptions about geometry, resistance, boundaries, demand or future climate define the simulated world.
Calibration evidence has a domain: agreement at one event, location or metric does not establish performance everywhere. Sensitivity analysis is useful when a defensible change to an uncertain input tests a decision claim or option ranking.
The limitations section should identify the affected output, explain the pathway and propose an evidence, monitoring or modelling response rather than listing generic uncertainty.
Review every dataset against the project use. Spatially sparse geometry may miss a critical control, a short time series may exclude the event range, and a datum mismatch can corrupt a water-level comparison despite precise measurements.
Record provenance, transformations and exclusions. A data table is not sufficient if the reader cannot see how quality affects the selected indicator or receptor.
Model assumptions require the same discipline. Geometry, resistance, boundaries, demand rules and future climate trajectories determine the simulated system. Identify which values are observed, estimated, inherited or defined for a scenario.
Test the assumption with the clearest pathway to the decision, while watching for compensation between inputs that can create a good fit for the wrong physical reason.
Communicate uncertainty according to consequence. State the element, propagation pathway, affected result and possible decision change.
Then propose a response: another observation, calibration event, sensitivity range, alternative representation, staged action or monitoring trigger. This structure distinguishes uncertainty from correctable error and prevents a limitations section from becoming a generic disclaimer detached from the recommendation.
What this chapter covers
- 01
Input quality
- 02
Model assumptions
- 03
Calibration domain
- 04
Sensitivity analysis
- 05
Error versus uncertainty
- 06
Decision confidence
Worked example · free
Handle an unstable option ranking
- 2Check whether calibration evidence constrains the resistance setting and whether the option difference is small.
- 1Report that the ranking is sensitive rather than averaging it into a confident answer.
- 2Recommend targeted evidence or a robust option that remains acceptable across the tested range.
Key terms
- Calibration domain
- The events, locations, variables and conditions covered by model-performance evidence.
- Sensitivity analysis
- A controlled test of how an input or assumption changes a result or decision.
- Structural uncertainty
- Uncertainty arising from how a model represents or omits real system processes.
Data Quality and Model Uncertainty FAQ
What makes a limitation specific?
A specific limitation names the uncertain element, the model process or output it affects, the decision consequence and a practical response. Generic statements that all data or models are uncertain do not establish importance.
How are error and uncertainty different?
A wrong unit, broken file path or mislabelled scenario is an error that should be corrected. An unobserved future demand or imperfect parameter is uncertainty that should be tested, bounded or managed.
Does matching one observation validate the model?
No. Agreement with one peak or gauge is evidence for that metric and condition. Timing, hydrograph shape, other locations and spatial extent may still be weak, so later claims should remain inside the tested domain.
How should limitations be prioritised in a short report?
Rank them by the possibility of changing the main indicator, crossing a decision threshold, reversing an option ranking or affecting an important receptor. A small uncertainty at a critical location can matter more than a broad range in a peripheral output. Pair each priority limitation with a practical evidence or management response.
When is sensitivity analysis decision-ready?
It is useful when the tested range is defensible, other settings remain controlled and the result is compared through an indicator that affects the recommendation. Report whether the conclusion stays stable, crosses a threshold or reverses an option ranking, then name the evidence or management response that follows.
Assessment move
Create an uncertainty chain for each case: input quality, model assumption, calibration evidence, scenario design, output and decision claim. Use the chain to explain how a weakness could change the result rather than writing a generic limitations list. Review datasets by spatial and temporal coverage, resolution, units, datum, completeness and relevance to the decision. Separate errors from uncertainties.
Correct errors immediately and document the change; use sensitivity, ranges, alternative scenarios or monitoring for uncertainties. When reading calibration evidence, list the event, location, variable and metric. Identify what remains untested, such as timing, low flows or floodplain extent. Choose sensitivity tests by decision importance.
Vary one justified factor, keep other settings stable and observe whether the main indicator, threshold or option ranking changes. Report stable and unstable conclusions differently. If a recommendation reverses, seek more evidence or choose a robust action rather than averaging conflicting runs. Practise writing one limitation sentence beside each major figure so readers encounter the boundary with the claim.
Finish with an actionable response, such as improved survey data, another observed event, a different model representation, staged action or a monitoring trigger.
Working through Data Quality and Model Uncertainty in EART66001? Sia is AskSia’s AI Environmental Science tutor — ask any EART66001 Data Quality and Model Uncertainty question and get a clear, step-by-step explanation grounded in how EART66001 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.