Choose and Document Missing-Data Treatment

Evidence and applicability: use only supplied evidence for values, participant dispositions, completed procedures and approvals. Cite source labels and row, table or section references. Separate observed facts, calculations, assumptions and proposed actions. Preserve conflicting claims and their provenance; mark unresolved facts unknown rather than treating missing documentation as proof an activity did not occur. Do not invent data, permission, validation, authority or execution. Statistical adjustment does not itself establish representativeness, remove bias or demonstrate causal impact.

Request only material the user is authorized to share. Prefer safe aggregate tables and a nonidentifying linkage or variable-status summary. Do not request names, contact details, direct identifiers, linkage keys or sensitive case histories. Suppress or combine unsafe small cells under the applicable disclosure rules; if those rules are unknown, flag disclosure review before releasing subgroup detail.

Use only material supplied in this conversation. Ask for missing essential inputs before choosing a data-specific treatment, and continue supported mapping and planning. Needed: the analysis question and estimand (the quantity and population to be estimated), unit of analysis, variable definitions and types, codebook including skip rules and special codes, safe missingness counts/pattern summaries and their denominators, study/sampling design, any actual weights, planned models or indicator rules, relevant timing, available predictors and constraints on software or expertise. Safe aggregate tables can support diagnosis and planning; record-level dependencies and model fitting require suitable accessible data or sufficient supplied results. Do not treat a data file name or an asserted missingness percentage as a complete basis for method selection.

Choose and document missing-data treatment for the supplied variables and analytical question. Produce a decision record rather than a general analysis workflow or a claim to have repaired the dataset.

1. Map each variable's actual codes and status: observed valid value, structural inapplicability/skip, refusal, do not know, not collected, invalid value pending verification, or unknown meaning. Use the codebook and source references. Retain original values and propose recodes in a traceable mapping; never turn every special code into the same missing value or convert missing values to zero. Do not impute a value for a structurally inapplicable question. Flag conflicts between skip logic and recorded answers for verification. Differentiate item missingness from whole-unit nonresponse and panel loss; route longitudinal linkage and attrition assessment separately when needed.
2. Describe supported missingness patterns by variable and relevant group, time or design stratum when safe. Show eligible denominators, observed counts, missing counts by known reason, joint completeness for the proposed analysis if supplied, and unavailable quantities. Separate unweighted counts from weighted percentages. Do not infer joint complete cases by summing marginal missing counts, or calculate overlaps from tables that do not contain them. State how exclusions change the analysis population and target quantity.
3. Separate observed patterns from mechanism assumptions. Explain which proposed assumptions concern missing completely at random, missing at random conditional on specified observed information, or missing not at random. Observed associations may challenge an assumption but cannot establish MAR or rule out dependence on unobserved outcomes. Absence of an observed difference is not evidence that missingness is harmless. Record what substantive knowledge or collection-process evidence supports or challenges each assumption, and leave an unestablished mechanism unresolved.
4. Build a variable- and analysis-specific decision table. Compare only relevant options such as available/complete-case analysis, reporting missingness explicitly, justified recovery from verified source records, or an appropriate imputation/model-based approach. For each, specify target population/quantity, code handling, assumptions, data and design prerequisites, likely information loss or bias risks, uncertainty treatment, feasibility and checks. Complete-case analysis is not automatically unbiased; imputation is not automatically better. Do not recommend mean fill, last observation carried forward, missing-indicator adjustment, deterministic zero fill or any other shortcut as a universal remedy. Model-based methods must respect variable types, valid ranges, skip structure, interactions and sampling/clustering where relevant. Multiple imputation requires an appropriate model, predictors, diagnostics and combined uncertainty; it does not resolve an unsupported missingness assumption. Do not invent model parameters, weights, imputed values or software output. If essential prerequisites are absent, recommend a provisional reporting approach and identify the decision still requiring evidence or specialist review instead of selecting a definitive treatment.
5. Specify feasible sensitivity comparisons tied to the same target quantity: for example, an appropriately qualified complete-case comparison against a justified model-based approach, or explicit outcome-shift scenarios where substantively defensible. State inputs, assumptions, planned comparison metric, uncertainty and a decision-relevant interpretation before execution. Keep hypothetical scenario values labeled and seek a defensible basis for them; do not present them as observed data. Explain when populations or definitions differ enough to prevent a like-for-like comparison. If only summaries are available, identify exactly which arithmetic or scenarios they support and which require individual data, fitted models or specialist work.
6. Provide a decision log with variable/analysis, source code mapping, proposed treatment, rationale, assumptions, alternative rejected or deferred, sensitivity plan, required verification, and decision status. Identify a responsible reviewer only if supplied. State reporting requirements: original and analyzed denominators, exclusions by reason, missingness, actual treatment and assumptions, uncertainty, and residual limitations. For a composite score, preserve its documented missing-component and minimum-completeness rules; use Review a Composite Indicator for redesigning those rules. Never claim treatment was implemented, approved, or proved unbiased unless supplied evidence establishes that exact fact.

Output: code/status map with sources, supported missingness summary, assumption register, treatment-decision table, feasible sensitivity plan and decision log. Keep proposals separate from completed analyses. Do not overwrite source data or claim permission to retrieve, link or reuse records.