Observational research · Missing data
How to report missing data: variable counts, analysis samples and assumptions
Have a data dictionary and missingness summaries? Build a reviewable report by defining who is eligible for each measurement, reconciling each analysis sample, and documenting the method, assumptions and sensitivity analyses. This guide supplies a synthetic table and an adaptable AI request—not a decision to delete or impute.
This guide concerns reporting in medical cohort, case-control and cross-sectional studies. STROBE items 12(c) and 14(b), with their explanation, ask authors to describe how missing data were addressed and report missing counts for each variable of interest. Participant flow should also explain exclusions due to missingness. Complete reporting does not establish an appropriate method or remove bias.
All counts below are synthetic teaching material, not patient data, real findings or a software run. The intended outputs are a variable-level missingness table, an analysis-denominator crosswalk and a draft with sources and unresolved questions. Formal numbers must come from verified, controlled analysis outputs.
Prepare materials that can be reconciled
Study sample and data dictionary
Supply authorized sources, versions, the unit of study, eligibility rules and observation times. Define each field’s applicable population, units, skip logic and special codes. Keep structural inapplicability, refusal, measurement failure and unknown reasons separate; do not infer meaning from a numeric code.
Deterministic summaries and overlap counts
Provide applicable, observed and missing counts and reasons by variable, with groups or time points where relevant. Obtain joint missingness or complete-record counts for the variables required by each analysis. Do not give a model patient-level values or infer joint counts from marginal totals.
Analysis plan and actual outputs
List outcomes, exposures, adjustment variables, target populations and planned missing-data methods for each model. Include actual analysis counts, software/version and deviations from the plan. Label undecided or unperformed methods as pending rather than completed.
Output 1: a missingness table with explicit denominators
Assume 500 unique eligible participants. Y, X and C apply to all 500. Z is collected only in a prespecified subgroup of 80; the other 420 are inapplicable and outside Z’s missingness denominator. Reasons have not been verified and must not be invented.
| Variable | Applicable N | Observed N | Missing N / % | Note |
|---|---|---|---|---|
| Outcome Y | 500 | 480 | 20 / 4% | Reason pending |
| Exposure X | 500 | 465 | 35 / 7% | Reason pending |
| Adjustment variable C | 500 | 485 | 15 / 3% | Reason pending |
| Subgroup measure Z | 80 | 76 | 4 / 5% | 420 inapplicable; reason pending |
Each row satisfies applicable N = observed N + missing N. Counts and percentages here are unweighted; weighted estimates for complex surveys require separate labels and cannot be inferred from this table. Add groups or time points as the design requires. Loss to follow-up, unmatched records and missing fields are not interchangeable.
Output 2: reconcile analysis samples using joint counts
Assume an additional verified synthetic count: 20 lack Y, 35 lack X, and 8 lack both. Within the same 500 participants, at least one is missing for 20+35−8=47 people; 500−47=453 have both Y and X observed. This count applies only to the Y/X variable set.
If the plan justifies an analysis requiring complete Y and X only, 453 is a candidate complete-case sample, subject to other eligibility rules and actual output checks. After adding C, do not simply subtract 15: missing C may overlap with those 47 people. Obtain the joint complete count for Y/X/C. Z is not automatically a model variable merely because it appears in the table.
Draft excerpt (synthetic description only): “Of 500 included participants, 20 lacked Y, 35 lacked X and 8 lacked both; 453 had observations for both Y and X. The jointly complete count after adding C remains to be calculated. Reasons and the analytical missing-data approach require researcher confirmation.” This does not claim that complete-case analysis was performed.
Output 3: distinguish methods, assumptions and sensitivity analyses
Report the method actually decided and used
For complete-case analysis, state required variables, actual samples and exclusion rules, and discuss sample changes and possible bias. For multiple imputation, report included variables, methods/models, number of imputations, diagnostics, pooling rules and actual analysis samples. Use the real plan and outputs; do not copy settings from another paper.
A missingness mechanism is an assumption, not a percentage label
MCAR means missingness does not depend on observed or unobserved data. Under MAR, conditional on observed data, missingness no longer depends on the missing value itself; under MNAR it can still depend on it. Percentages, group comparisons or a single test cannot prove MAR rather than MNAR. Discuss plausibility using the collection process and retain unknown reasons as unknown.
Specify which assumption a sensitivity analysis challenges
Address a defined missingness assumption or analysis-population difference, compare the same target estimate and uncertainty with the primary analysis, and explain differences. “Sensitivity analyses were conducted” is insufficient. Mark unperformed work as proposed/pending; invent no estimates, and do not treat similar results as proof of no bias.
Use AI to organize a reviewable draft
Adaptable suggested input—not an execution record
Use only my authorized dictionary, missingness summaries, joint counts, agreed plan and actual outputs. Prepare: (1) applicable/observed/missing counts and denominators by variable; (2) required variable sets and actual sample counts by analysis; (3) a methods and assumptions draft; (4) performed versus planned sensitivity analyses; (5) source mappings and unresolved questions. Check arithmetic and overlap first. If joint counts cannot be calculated, state exactly what is missing. Separate structural inapplicability from missingness. Do not infer MCAR/MAR/MNAR, choose deletion or imputation, generate findings, or describe planned work as performed. Materials: [versions/authorization, dictionary, summaries, plan, outputs].
The handoff should map each statement to its evidence, status and unresolved questions, not merely return polished prose. The 500-person example is teaching material only. Recheck all counts when replacing inputs; do not carry over illustrative numbers or invent run records.
Six human checks before handing over the draft
- Are applicable denominators and special codes source-bound?
- Can counts and percentages be recalculated, with overlap independently verified?
- Are Table 1, flow-diagram and model denominators individually defined and differences explained?
- Do methods, imputation settings and samples come from actual execution rather than suggestions?
- Are reasons, mechanism assumptions, sensitivity status and limitations honestly retained?
- Have you provided only authorized/public aggregate material to AI, not patient-level records?
Common questions
Can missingness below 5% go unreported?
No universal threshold waives reporting. STROBE asks for missing counts for variables of interest and an account of their handling. A low percentage neither proves absence of bias nor automatically permits deletion.
Can marginal missing counts determine the final model sample?
Usually not. You need the model variable set, applicability rules and joint missing/complete counts. Marginal totals do not reveal overlap and cannot simply be subtracted column by column.
Does this guide perform imputation or statistics for me?
No. It provides reporting methods and suggested input; no imputation or model was selected, executed or validated. Researchers must decide methods in light of the design and appropriate statistical expertise.
Verified sources
The official sources below were checked on 2026-10-17. Access conditions and update dates are rechecked before publication.
- STROBE: choose the checklist for your observational study design
Checked requirements for missing-data handling and variable-level counts; not a quality score.
- STROBE Explanation and Elaboration: items 12(c), 14(b) and Box 6
Checked reasons, analysis samples, assumptions and the limits of proving MAR rather than MNAR from observed data.
Materials to reviewable outputs
Organize a reporting draft in RigorMed with your summaries
Describe this research-writing task in the browser workspace and provide an authorized dictionary, aggregate summaries and plan. Request a missingness table, denominator crosswalk and unresolved-question list. Check actual materials, task support and credit notices before starting; automatic imputation, statistical execution and professional approval are not promised.