L-09 · Method

Missing data in AI-answer observations

Primary intent
Record missing AI-answer data without silently changing denominators or confusing absence with zero.
Evidence state
Source-grounded reference
Review owner
Matthias Ramahi · independent review not claimed
Last reviewed
2026-08-22
Direct answer

Missing data

Missing AI-answer data should be represented with an explicit reason and eligibility state. No answer, no visible citation, blocked access, capture failure and rights-limited retention are different conditions; none should be silently converted to a zero or dropped row.

Use this taxonomy when designing a schema, cleaning a batch or calculating any answer or source measure.

Missing is not zero

An answer with zero visible citations is an observed value. An answer that failed to load has unknown citation count. A route that does not expose citations may be ineligible for a citation measure. Treating all three as zero changes the result and hides why.

Store the raw state and derive measure eligibility through code. This keeps the record reusable when a later analysis applies a different defensible rule.

Use a bounded reason taxonomy

Free-text notes are useful but difficult to aggregate. Define a small versioned set of machine-readable reasons, then allow an optional note for case detail. The taxonomy should match the actual collection process rather than copying a generic data-science list.

A useful starting set separates response unavailable, response interrupted, no visible source, source UI unavailable, access blocked, capture failure, field not exposed, retention not permitted and protocol exclusion.

Publish the eligibility flow

For every measure, show how the initial scheduled observations became the analyzed denominator. Report scheduled, attempted, completed, eligible, excluded and missing counts. Reasons should reconcile to the totals.

When missingness differs by surface or window, report that before comparing the main metric. A change in missingness may explain an apparent change in diversity or stability.

  • Never delete a scheduled observation because it failed.
  • Never use one denominator label for different eligibility rules.
  • Never infer “no citations” from a missing capture.
  • Version changes to the reason taxonomy.

Rights-limited is a data state, not a loophole

Provider terms or applicable rights may permit metadata while restricting raw output retention or publication. Record the permitted representation and the decision source. Do not imply that a hash or summary reproduces content that reviewers cannot inspect.

If the limitation prevents the research question from being audited, the correct outcome may be to withhold the claim or redesign the study.

S

Source notes

These sources support the definitions, standards or project boundaries named in this reference. They do not prove that a public observation dataset exists.

  1. portfolio-dossier
    Canonical ai-fanout.com domain dossier

    Confirmed ownership, accepted public Evidence Lab purpose, named Research Owner, indexable website launch and separately gated provider research.

    Owner record
  2. nist-ai-rmf-genai
    NIST AI RMF Generative AI Profile

    Supports explicit measurement, documentation, monitoring and limitations for generative-AI evaluations.

    Open
  3. w3c-prov-o
    PROV-O: The PROV Ontology

    Provides provenance concepts for entities, activities, agents, derivations, sources and versions.

    Open
  4. fair-principles
    The FAIR Data Principles

    Supports reusable research data with metadata, provenance and clear usage licenses.

    Open