Skip to main content Link Menu Expand (external link) Document Search Copy Copied
Biology Skills

Measurement, observability and negative evidence

Summary

A biological state can be confidently described as absent only when the measurement or assessment had an adequate opportunity to observe that state.

Absent, not detected, not measured, not callable, below detection limit, unknown, and not applicable are different scientific states. Collapsing them can create false negative evidence and can make downstream statistical or causal inference stronger than the data justify.

Core rules

  • A negative result is evidence against a biological state only relative to a measurement process capable of observing that state.
  • Lack of a positive record is not itself a negative measurement.
  • Preserve the difference between observed absence, non-detection, missingness, lack of assessment, lack of callability, below-detection values, unknown states, and not-applicable states.
  • Observability can be location-specific, event-class-specific, sample-specific, phenotype-specific, state-specific, and time-specific.
  • A measurement can have adequate data volume while still being insensitive to the relevant event or biological state.
  • Preserve the reason for missing or unresolved information when that reason affects interpretation.
  • Do not convert an unresolved measurement into evidence against a hypothesis.
  • Detection limits, analytical sensitivity, assessment criteria, and relevant QC can be part of the meaning of a negative result.
  • Derived datasets must not strengthen the meaning of an upstream negative state merely by reducing its representation.

Observation states

The exact vocabulary varies by domain, but these concepts should remain separable when relevant:

observed present
observed absent
not detected
not measured
not assessed
not callable
below detection limit
missing
unknown
not applicable

A binary representation may be useful downstream, but the upstream state and reduction rule should remain recoverable when the distinction can affect scientific interpretation.

Required context

For a negative or unresolved observation, establish where relevant:

  • what biological state is being assessed
  • the specimen, tissue, cell type, individual, or other measurement target
  • the assay or assessment method
  • whether the relevant location or state was in assay scope
  • analytical sensitivity or detection limit
  • technical quality required for interpretation
  • temporal context
  • whether the result is explicit negative evidence or merely absent from the data
  • why a value is missing or unresolved
  • the denominator or coverage supporting a negative claim

AI behaviour

Before stating that something is absent, negative, reference, normal, or not present:

  1. identify the exact state being claimed absent,
  2. determine whether it was explicitly measured or assessed,
  3. determine whether the method could detect it in the relevant context,
  4. distinguish a resolved negative result from an unresolved or missing result,
  5. preserve the relevant detection, quality, or denominator information.

Do not infer a negative phenotype from lack of documentation.

Do not infer a reference genotype from lack of a variant record.

Do not infer absence of expression from failure to detect a transcript without considering assay sensitivity and context.

When the available data support only non-detection, say not detected rather than strengthening the claim to biological absence.

Common failure modes

Missing record treated as negative evidence

Observed:
  no record for locus X

Unsafe:
  locus X is reference

Required:
  determine whether the locus was measured and callable

No emitted record can reflect a variant-only representation, filtering, missing coverage, parser behaviour, or a genuinely resolved reference state.

Unassessed phenotype treated as absent

A phenotype omitted from a clinical record may never have been assessed. This is not equivalent to an explicit negative examination.

Below detection limit treated as zero

A result below a method’s detection threshold does not establish a true biological quantity of zero.

Unsupported event class excluded

A sequencing assay that reliably detects SNVs does not automatically exclude copy-number changes, repeat expansions, structural rearrangements, mosaic events, or other event classes outside its validated scope.

Authoritative standards

Use domain-specific assay and reporting standards for exact requirements. Biology Skills defines the interpretive distinction: negative evidence requires adequate observability.

Where structured phenotype exchange is required, use maintained standards such as GA4GH Phenopackets rather than inventing local missing-versus-excluded semantics. Where genomic experiment metadata affect observability, use maintained experiment metadata standards and assay documentation.

Examples

Genomic observation

no variant record
!= callable reference genotype
!= unresolved locus

Phenotype observation

not mentioned in record
!= assessed and absent

Quantitative assay

signal < detection limit
!= biological concentration = 0

Sources

  • GA4GH Phenopackets: https://www.ga4gh.org/product/phenopackets/
  • GA4GH Experiments Metadata Checklist: https://www.ga4gh.org/product/experiments-metadata-standard/
  • gnomAD: https://gnomad.broadinstitute.org/
  • Wilkinson MD et al. The FAIR Guiding Principles for scientific data management and stewardship: https://doi.org/10.1038/sdata.2016.18