Sensitivity testing a result that sits near a threshold
A result two per cent from a policy threshold is the weakest part of any evidence base. Here is how to test it, report a range and close the argument before a consultee reopens it.
A result that sits two or three per cent from a policy threshold is the most fragile part of an evidence base, and sensitivity testing is how you stop a case officer reading that margin as doubt. You vary the inputs you trust least, show what the output does, and state plainly whether the conclusion holds. Done properly it closes the argument down. Left out, it invites a consultee to run their own numbers and land somewhere else.
Officers and statutory consultees rarely have time to rebuild a model. They do have time to ask whether one input could be wrong, and that is the question a sensitivity test answers before anyone else asks it.
Why a marginal result needs testing
Thresholds in the English system are mostly hard edges. A retail proposal above the 2,500 m2 default figure in the National Planning Policy Framework (edition of 17 August 2026) attracts an impact assessment, and below a locally set threshold it may not. A Schedule 2 urban development project under the Town and Country Planning (Environmental Impact Assessment) Regulations 2017 crosses into screening territory at defined site areas and floorspace figures. A statutory biodiversity metric output of 10.2 per cent meets the gain requirement under the Environment Act 2021. An output of 9.8 per cent does not.
Where the answer is comfortable, nobody asks how you reached it. Where it is marginal, every choice in the method becomes material: the survey date, the baseline condition, the rounding convention, the version of the tool.
There is a second reason, less often discussed. A marginal result that is presented as a single number reads as overconfident, and overconfidence is what invites a holding objection. A range with reasoning attached reads as competence.
Choosing the inputs to vary
Vary the inputs that are both uncertain and influential. Everything else is decoration, and a test that shuffles twelve variables tells the reader nothing about which one matters.
In practice that means a short list, chosen after you have found where the arithmetic is most sensitive:
- Inputs taken from a single measurement or a single day of survey, such as one noise measurement position or one traffic count.
- Inputs carried across from another discipline, where a late design change may not have reached you.
- Professional judgement inputs, such as habitat condition scoring in the biodiversity metric or reflectance assumptions in a daylight model.
- Inputs that depend on a published dataset with a known margin, such as modelled background pollutant concentrations or a mapped flood extent.
- Inputs where the guidance itself offers a choice of method.
Modelled time of day and time of year belong on that list too. Anyone who has argued about shadow plots and the dates they are run for knows how much the chosen date does to the answer.
Habitat condition is the other classic. Moving one parcel of modified grassland from moderate to good condition can swing a metric result by more than the margin you are arguing about, so the sensible test scores that parcel both ways and reports both answers.
Reporting a range rather than a point
State the central case first, with the inputs that produced it, then the tested cases beneath it. The central case is your professional opinion and should be labelled as such. The tested cases show the reader where the conclusion would change.
Three habits make the difference. Give each case a name that describes the assumption rather than a letter. Quote the output to the same number of decimal places throughout, because a jump from one decimal place to three looks like a result being massaged. And say which case you would defend at appeal.
If one tested case breaches the threshold, say so in the same paragraph. Burying it in an appendix is how a reason for refusal gets written for you.
Reasonable worst case, and how it differs from worst case
The absolute worst case is usually a physical impossibility: every uncertain input simultaneously at its least favourable value. Stacking independent pessimistic assumptions produces a number nobody can design to, and it is the fastest way to lose a viability or noise argument you should have won.
A reasonable worst case picks the pessimistic end of the inputs that plausibly co-vary and holds the rest at their central values. For a noise assessment under BS 4142:2014+A1:2019 that might mean plant running at full duty during the quietest measured background period, which can happen, rather than full duty combined with the lowest background ever recorded anywhere on the site and the closest theoretical receptor window.
Say which you have used. Consultees read "worst case" as a term of art and will assume the strict version unless you define it.
| Case | What is varied | Why it is credible | Effect on the conclusion |
|---|---|---|---|
| Central | Measured and agreed inputs | Professional judgement on the evidence | Conclusion as reported |
| Reasonable worst case | Co-varying pessimistic inputs | Operationally possible together | Conclusion tested |
| Single input test | One variable at a time | Isolates the sensitive input | Shows the margin |
| Threshold reversal | The value needed to change the answer | Defines how much error is tolerable | States the safety margin |
That last row does more work than the rest put together. If the reader can see that the input would need to be 40 per cent different before the answer changes, the margin stops being an issue.
Where a precautionary assumption works in your favour
Precaution is normally framed as a cost. It can also be a tactic. Assume the less favourable value on a point the consultee is likely to challenge, show that the scheme still passes, and you have removed the challenge rather than answered it.
This works best where the precautionary assumption costs almost nothing in design terms. Setting a finished floor level 150 mm higher than the modelled level requires, adopting the shorter of two possible sight lines, or scoring a hedgerow as species-rich when the survey is equivocal will often cost little and buy a clean recommendation. It works badly where the precautionary figure drives a structural or commercial change, which is the point at which you argue the central case instead and test around it.
Zoned criteria reward the same move. Lighting limits tighten as you move between environmental zones and their obtrusive light limits, so where a site sits on a zone boundary the cleaner course is to design to the stricter zone and say you have done it.
Conditions that manage residual doubt
Some uncertainty cannot be tested away because the input does not exist yet. The plant has not been selected, the tenant is unknown, the contractor has not been appointed. That is a condition, not a sensitivity test.
Offer the wording. A condition requiring the installed plant to achieve a stated rating level at a defined assessment position, verified by post-completion measurement, converts an untestable assumption into an enforceable obligation. Planning Practice Guidance on the use of planning conditions sets out the six tests each one must meet, and a condition you have drafted yourself is far more likely to survive than one improvised by an officer on a Friday afternoon.
Where a condition is the answer, say in the assessment which specific residual doubt it closes. An offered condition with no stated purpose looks like an admission that the assessment is incomplete.
Setting the test out in one table
One table, one page, near the front of the report. Keep it out of the appendices.
The table needs the case name, the varied input and its value, the resulting output, and whether the threshold is met. Add a single column for the source of each input, because that is the column a consultee reads first. Then a short paragraph under it giving your conclusion in one sentence.
Illustrative example, and it is illustrative only: a drainage calculation returns a discharge rate of 5.1 litres per second against a 5.0 litres per second limit. The infiltration coefficient came from three tests in one trial pit. Testing at the lowest of the three measured values takes the rate to 5.4, and at the mean of all three to 4.8. The honest report says the limit is met on the mean value, names the tests, and offers a condition for a verified as-built rate. The dishonest version quotes 4.8 and calls it the result. Officers notice.
Reports drafted in PlanScope follow the same discipline: where a site-specific figure is missing, the draft inserts a "TO BE CONFIRMED" flag rather than filling the gap with a plausible number, which leaves the sensitivity question visible to whoever reviews it. Where you want a structure to work from across a multi-consultant submission, keeping every report on one set of figures matters as much as the figures themselves. You can see the full set at all 66 PlanScope report types and start a draft from the one you need.
PlanScope provides AI-assisted draft reports for planning applications. Reports should be reviewed before submission. We do not guarantee planning approval. This is general guidance, not legal advice; check the position with your local planning authority or a qualified professional before relying on it.
Frequently asked questions
What is sensitivity testing in a planning assessment?
It is a short set of re-runs of your calculation with the least certain inputs changed, reported alongside the central case. The point is to show the reader how much an input would need to differ before the conclusion changes. It is standard practice in noise, drainage, air quality, daylight and biodiversity metric work, and it is expected wherever a result sits close to a threshold.
How many inputs should I vary?
Usually two to four. Choose inputs that are both uncertain and influential on the output, and vary them one at a time so the reader can see which one matters. A test covering a dozen variables loses the message and reads as padding. Add one combined case if two of your inputs plausibly move together in practice.
Is a reasonable worst case the same as a worst case?
No. An absolute worst case puts every uncertain input at its least favourable value at once, which is often physically impossible. A reasonable worst case takes the pessimistic end of inputs that credibly occur together and holds the others at central values. Define which you have used, because consultees will assume the strict version otherwise.
Can a planning condition replace a sensitivity test?
Only where the input does not yet exist, such as plant not yet selected or a tenant not yet known. In that situation offer a condition requiring a measured performance standard after completion. Where the input exists but is uncertain, test it. Planning Practice Guidance sets six tests a condition must meet, so draft the wording yourself.
Where should the sensitivity table go in the report?
Near the front, in the assessment chapter, not in an appendix. Give it the case name, the varied input and value, the output, the source of the input and whether the threshold is met. Follow it with one sentence stating which case you would defend. Officers and consultees read that table before the text.
Sources and further reading
This article is general guidance about the planning system in England and, where stated, the rest of the UK. It is not legal or professional advice. Check your local planning authority’s validation requirements and take professional advice for your specific site.