Missing-data assumptions & treatment handoff

Statistical handoff clinic · Document what is known about missingness, what remains an assumption, and which treatment decisions must be made in specialist statistical software.

Use the Denominator, exposure & censoring clinic first. That clinic preserves the mechanics of observed zero, partial observation, missing trial, absent-by-design targets, undefined denominators, and right-censored latency. Use this page next when the remaining question is what assumptions are needed for unavailable measurements and how will the statistical analysis address them?

MCAR, MAR, and MNAR are not software labels

An observed table cannot by itself prove that missingness is MCAR, MAR, or MNAR. GazeForge records questions, evidence, grouping, and planned sensitivity work; it does not infer a missingness mechanism.

Core contract: source/reason ≠ mechanism ≠ statistical treatment.

Treatment selection belongs to specialist statistics

GazeForge does not automatically choose complete-case analysis, single or multiple imputation, inverse-probability weighting, a joint model, pattern-mixture or selection-model sensitivity analysis, a survival model, GLM/GLMM, SEM, Bayesian modelling, or another inferential strategy.

Missing-data source is not missing-data mechanism

First record why the value is unavailable or not applicable using the strongest study evidence available.

Source class	Meaning	Missingness-mechanism assessment?
observed	target outcome observed	no
acquisition_unavailable	usable measurement unavailable	yes, study-specific
partial_observation	reduced usable exposure	yes, study-specific
design_absence	target did not exist by design	no; not stochastic missingness
undefined_derived_metric	required denominator/support unavailable or zero	no; derived metric is undefined
right_censored_event	target observable but event not seen before follow-up ended	no; retain censoring state

Do not collapse the last three rows into one generic missing category. Doing so can change the estimand and can destroy information needed by later models.

Questions to ask before using MCAR/MAR/MNAR language

A defensible missing-data record asks questions rather than assigning an automatic classification.

1. Is the unavailable value generated by the experimental design?
2. Is availability associated with prespecified observed participant, trial, stimulus, condition, session, device-state, or QC variables?
3. Could the unobserved outcome itself plausibly affect the probability of being unavailable?
4. Is missingness clustered within participant, stimulus, session, site, or condition?
5. Are acquisition failures plausibly related to task difficulty, motion, display state, participant behaviour, or another observed/unobserved process?
6. Would the proposed treatment change the target population, denominator, or estimand?
7. Which assumptions are required by the intended specialist statistical method?

The answer may motivate a missing-data assumption or sensitivity analysis, but it does not mechanically prove one.

Keep missingness separate from QC and exclusion

These are different records:

Layer	Question
missingness/source state	What was unavailable, partial, or observed?
QC evidence	What technical/data-quality concern was detected?
review decision	What did the reviewer decide?
exclusion ledger	Which analysis unit was excluded and why?
statistical treatment	How will unavailable observations be handled in the model?

A QC flag can coexist with an observed outcome. A missing outcome can be retained in the analysis handoff. An explicit reviewed exclusion can occur even when an outcome was observed. Never use “missing” as a synonym for “invalid”.

Continue to the QC review & exclusion ledger when an actual exclusion decision is required.

Complete-case analysis is not neutral preprocessing

A complete-case subset can change:

• participant/trial/stimulus denominators;
• the target population;
• the estimand;
• condition balance;
• repeated-measures structure;
• the distribution of quality/exposure variables.

Therefore GazeForge does not silently drop rows with missing values. If a complete-case analysis is scientifically justified, record it as an explicit analysis strategy with its denominator and estimand implications.

Imputation is not an automatic cleaning step

Possible strategies include single imputation, multiple imputation, weighting, joint models, and other model-based approaches. Their validity depends on assumptions, variables included in the model, grouping structure, diagnostics, compatibility with the estimand, and the substantive design.

This clinic records candidate strategy families as not_selected_by_gazeforge. It does not implement or rank them.

Preserve grouping for specialist analysis

The missing-data handoff should retain the identifiers needed to model or diagnose dependence:

participant_id
trial_id / session_id
stimulus_id
condition
AOI identity where relevant
source / device / site identity where relevant
outcome identity
missing-data source class
QC/review state
denominator / observable exposure

Missingness clustered within a participant or stimulus cannot be studied correctly if those identities were removed during aggregation. Use the Grouping, repeated measures & pseudoreplication clinic to audit those identities before a missing-data treatment or specialist model is chosen.

Censoring remains censoring

No-event first-fixation latency belongs to the censoring contract already documented in the Denominator, exposure & censoring clinic.

Do not convert a right-censored event into:

• ordinary missingness;
• zero latency;
• an observed latency equal to the trial maximum;
• an arbitrary large latency;
• an automatically excluded trial.

Whether and how the censoring is modelled belongs to specialist statistical software.

Sensitivity and deviations

Missing-data sensitivity analyses should enter the same governance path as other analysis variants:

• register the intended strategy before result inspection where possible;
• preserve the primary estimand;
• state when a variant changes the estimand;
• retain non-evaluable or failed variants;
• do not report only the most favourable specification.

Use the Sensitivity & robustness clinic for the executed comparison and deviation ledger.

Worked assumptions audit

Run:

python examples/18_worked_missing_data_assumptions_audit.py \
  --output-dir worked-missing-data-assumptions-audit

The deterministic bundle writes:

01_missing_data_source_registry.csv
02_mechanism_assumption_questions.csv
03_analysis_treatment_registry.csv
04_exclusion_missingness_separation.csv
05_sensitivity_handoff.csv
06_reporting_language.csv
README.md
missing_data_assumptions_manifest.json

It verifies that:

• source classes remain explicit;
• MCAR/MAR/MNAR are never inferred;
• candidate treatment families remain not_selected_by_gazeforge;
• no treatment is selected automatically;
• missingness does not imply automatic exclusion;
• right-censored events remain censored;
• design absence and undefined metrics are not reclassified as missing;
• no device, measurement, construct, causal, or psychological-state validity claim is created.

The example is synthetic_demo_not_empirical_evidence.

Reporting examples

Mechanism

The observed data did not by themselves establish an MCAR, MAR, or MNAR mechanism. Missing-data assumptions were specified in the specialist statistical analysis plan and assessed using the prespecified diagnostic and sensitivity procedures.

Complete-case analysis

Any complete-case analysis was treated as an explicit analysis strategy rather than a cleaning step; the resulting denominator and estimand implications were reported.

Imputation or weighting

The missing-data treatment was prespecified and implemented in specialist statistical software. GazeForge preserved outcome/source/QC/grouping metadata but did not select or execute the treatment method.

Exclusions

Missing-data source state, QC flags, and reviewed exclusions were retained as separate records. Missingness alone was not treated as an automatic exclusion.

Censoring

No-event first-fixation observations remained right-censored and were not reclassified as ordinary missing outcomes.

Limitations

This clinic improves auditability but does not establish:

• the true missingness mechanism;
• that MCAR, MAR, or MNAR holds;
• that a complete-case analysis is unbiased;
• that an imputation/weighting/model strategy is appropriate;
• that censoring is non-informative;
• device or tracker validity;
• measurement or construct validity;
• a causal effect;
• a latent psychological state.

API and workflow links

This handoff references existing public surfaces rather than introducing a new statistical API:

• Schema API — preserve participant/trial identity.
• Quality-control API — keep QC evidence separate from missingness and exclusion.
• Eye-events API — preserve event definitions used by latency/count outcomes.
• Semantic AOI API and Dynamic AOI API — preserve AOI support and design/observability context.
• Scanpath API — preserve unavailable sequence states.
• Sampling-sensitivity API — route relevant acquisition/derivation sensitivity work.

Continue through Outcome & estimand preregistration, Denominator, exposure & censoring, Analysis handoff, Sensitivity & robustness, Reporting & interpretation, and Publication readiness.