Skip to main content

Reliability, Confidence, and What a Dataset Doesn't Say

Overview​

Every dataset carries two things that never appear on the map: how confident it is in what it does show, and what its silence means where it shows nothing. Reading both is the core skill that separates a defensible desk assessment from a confident-looking guess. This page pulls together a theme running through all of WildKnowledge: a polygon is a claim, not a fact — interrogate the claim.

Why it matters for nature strategy​

BNG turns habitat into numbers, and numbers hide their own uncertainty. A unit total looks equally solid whether it rests on ground-truthed survey or a low-confidence model. Two questions recover the missing context:

  1. How confident is this data in what it shows?
  2. What does it mean when this data shows nothing?

Get these wrong and you make the two classic errors: over-trusting a confident- looking guess, and reading silence as reassurance.

The two questions in practice​

1. Confidence — does the dataset tell you how sure it is?​

  • Some datasets publish confidence — Living England's reliability field is the standout example. Use it: a "very low" reliability class is a hypothesis, not a finding.
  • Some imply confidence through method — PHI's survey/compiled provenance signals generally higher confidence, but unevenly and without a per-feature score, so you must reason about the source yourself.
  • Some carry no confidence signal at all — treat those with the most caution, not the least.

2. Silence — what does "nothing here" mean?​

Absence in a dataset almost never means "confirmed absent". It usually means one of:

  • Not surveyed / not mapped — e.g. no PHI polygon means "not a mapped priority habitat here", not "no habitat".
  • Not recorded — e.g. an NBN Atlas blank reflects recorder effort, not species absence.
  • Not complete — e.g. the Conservation Areas national layer is knowably missing LPAs, so absence ≠ not designated.

Absence of evidence is not evidence of absence. It is the single most-repeated lesson across these profiles because it is the single most-repeated error.

How to apply it​

  • Carry the method with the data. A habitat is never just "lowland meadow" — it's "lowland meadow (PHI, surveyed)" or "lowland meadow (Living England, reliability: low)". Different claims.
  • Let low confidence trigger fieldwork, not conclusions. A high-stakes result resting on a low-confidence input is a survey trigger.
  • State what silence means every time you rely on it — "no records" and "confirmed absent" must never be written as the same sentence.

WildStack's take​

WildStack's take

The most valuable question you can ask any nature dataset is the one it doesn't answer on its face: how sure are you, and what does your blank space mean? BNG's whole apparatus is built to produce a confident number, and confidence is exactly what a raw unit total fakes. Our entire opinionated posture — scoring data confidence explicitly, letting low reliability trigger survey, refusing to read silence as reassurance — comes down to this one habit. A dataset that tells you how much to doubt it (like Living England's reliability field) is doing you a favour; a dataset that stays silent about its own uncertainty is the one to handle most carefully. Treat every polygon as a claim with a confidence and a scope, and most desk-assessment errors disappear.

Official sources​

Last reviewed​

5 July 2026. A stable methodological topic; revisit if the major datasets change how they express (or hide) confidence.

Last updated on