RESULT QC reviews tables, figures, estimates, model outputs, execution logs, and result narratives produced after analysis in R. It assesses alignment with the plan, numerical internal consistency, correspondence with code and logs, presentation, interpretation, and traceability.
Purpose and scope
RESULT QC determines whether results are consistent with the plan, code, execution record, and narrative interpretation within the available evidence. Internal-consistency review, code correspondence, rerunning, and independent recalculation are distinct verification depths.
RESULT QC does not automatically include independent review of all code or reprogramming of the statistical analysis. Do not describe work as reproduced or independently verified unless those operations were performed.
Entry criteria
At least one table, figure, model output, log, or result narrative is required. Each stronger judgment requires its corresponding input.
| Judgment | Required input |
|---|---|
| Presentation and internal consistency | Table, figure, or model output |
| Conformance to plan | Result plus analysis plan or SAP |
| Correspondence with code | Result plus R script |
| Execution status | Result plus execution log |
| Rerun | Data, code, and execution environment |
| Independent recalculation | Data, independent specification, and independent code |
Verification depth
| Level | Review performed | Main limitation |
|---|---|---|
| 1 | Internal consistency of tables, figures, and text | Generation process cannot be verified |
| 2 | Correspondence with the plan or SAP | Actual code implementation cannot be verified |
| 3 | Static correspondence with code and logs | Identical regeneration is not demonstrated |
| 4 | Rerun using the supplied code | The code itself is not independent |
| 5 | Recalculation from an independent specification and code | Depends on source-data and independent-specification accuracy |
Report the actual materials reviewed and operations performed, not merely the highest possible level suggested by available files.
When to use it
- After creating tables, graphs, regression models, survival analyses, or competing-risk analyses
- After producing an R Markdown or Quarto report
- After an AI agent drafts result text
- Before drafting a paper or report result section
- Before transferring results to collaborators
- When assessing warnings, errors, or the need for reanalysis
Analysis plan or SAP
↓
R script
↓
Analysis execution in R
↓
RESULT QC
↓
Correction and reanalysis when required
↓
Report or manuscript
↓
Human statistical reviewReview targets
Review descriptive, cross-tabulation, and baseline tables; group comparisons and models; survival and competing-risk outputs; estimates, confidence intervals, and p-values; subgroup, sensitivity, and safety analyses; R scripts, console output, and logs; R Markdown or Quarto reports; and drafted interpretations.
Review the generating code, execution record, plan, and dataset information when available—not only the displayed result.
Materials and assessable scope
| Material supplied | Main assessable scope |
|---|---|
| Tables or figures only | Internal consistency, presentation, labels, interpretation |
| Results plus plan | Conformance to planned outputs |
| Results plus R code | Apparent correspondence between code and result |
| Results plus code and logs | Warnings, errors, and execution state |
| Data plus code and environment | Stronger reproducibility check |
| Independent verification code | Numerical comparison by independent recalculation |
A table alone may support checks of displayed internal consistency, but not proof that it was generated from the correct data and code.
Six review domains
1. Conformance to planned outputs
Review analysis population, endpoint and time point, comparison groups, method and effect measure, confidence intervals and p-values, adjustment and stratification variables, subgroup and sensitivity analyses, output type, and classification as primary, sensitivity, or supplementary.
For example, presenting only PPS results when FAS is specified for the primary analysis is a plan inconsistency.
2. Numerical internal consistency
Review overall and group counts, exclusions, numerators, denominators, percentages, row and column totals, missing counts, event and censoring counts, model-specific analysis counts, and rounding differences.
Identify whether a percentage denominator is all participants, nonmissing participants, or evaluable participants. Record recalculation formulas, display precision, and rounding rules. Use a specification-defined tolerance when available; do not invent a tolerance during QC.
3. Estimates, confidence intervals, and p-values
Review estimate direction and reference group, effect measure and scale, confidence level, whether the estimate lies within its interval, broad consistency with the p-value, transformed versus original scales, and adjusted versus unadjusted results.
An estimate outside its displayed interval requires review of presentation, scale transformation, comparison direction, or transcription.
4. Correspondence with code and logs
Trace the input dataset, filters and analysis flags, grouping and reference categories, variable derivations, missing-data handling, model formula and covariates, confidence-interval method, packages, labels, and output files.
An “adjusted for age and sex” label is inconsistent if the model formula contains neither variable. Logs should be checked for errors, warnings, record counts, exclusions, R and package versions, random seeds, convergence, singularity, separation, collinearity, and relevant model assumptions.
The existence of code is not by itself evidence that the displayed output was generated by that code.
5. Clarity of tables and figures
For tables, review title, population, groups, row and column labels, denominator, units, time points, missingness, estimates, reference categories, abbreviations, footnotes, and digits.
For figures, review title, axes, units, groups and legends,
confidence intervals, time points, risk sets, censoring marks, and
estimation method. Distinguish a cumulative incidence function from
1 - Kaplan–Meier when competing risks exist.
For model outputs, identify variables, categories and references, adjustment, effect measures, analysis counts, events, covariates, and units.
6. Interpretation and traceability
Check that text matches numbers, effect direction and units are correct, uncertainty is acknowledged, nonsignificance is not equated with no difference, association is not described as causation, exploratory findings are not presented as confirmatory, and subgroup heterogeneity is not inferred from separate significance tests.
Confirm traceability to dataset and version, data-cut date, plan version, scripts, logs, R and package versions, outputs, limitations, and unresolved items.
RESULT QC versus code QC
RESULT QC evaluates correspondence between results and code. It is not an independent review of every branch and exception in the code. Without independent code and recalculation from source data, do not treat it as independent numerical verification.
AI assumption risks
| Missing information | Possible AI assumption | Main impact |
|---|---|---|
| Denominator | Use the displayed total | Percentage meaning may change |
| R code | Assume values were computed correctly | Calculation errors cannot be assessed |
| Execution log | Assume successful completion | Warnings and errors cannot be assessed |
| Group direction | Use the first group as reference | Effect direction may be wrong |
| Study design | Describe an association causally | Results may be overinterpreted |
Record unavailable information rather than infer it.
Decision rules and evidence requirements
Use OK, Needs clarification,
Problem, Cannot assess, or
Not applicable for each domain.
- A mismatch in the primary population, group labels, endpoint, or model is normally Critical or Major.
- Do not mark reporting or interpretation
Readywhen a primary model has not converged or an interpretation-relevant warning is unresolved. - If code or logs are unavailable, use
Cannot assessfor those domains and do not infer computational accuracy. - Presentation-only issues may be Minor or Note depending on impact.
- Distinguish issues requiring reanalysis from those resolved by presentation or text changes.
Evidence identifies table or figure numbers, cells or
row labels, model names, output filenames, script names and processing
labels, warnings, plan sections, and recalculation formulas. Reference
long logs or sessionInfo() in separate files rather than
pasting them into the report.
Output file
ai_project/qc/result-qc-001.mdUse sequential numbers for later reviews and retain earlier reports. The standard report contains Skill information, Overall judgment, Readiness by next step, Materials reviewed, Domain assessment, Issues, Cannot-assess items, AI assumption risks, Required clarification questions, Recommended next actions, Quick assessment, and Handoff to next workflow step.
The Issues table contains ID, Severity,
Issue, Evidence, and
Recommended action. When reanalysis is required, specify
the changed process, the rerun scope, and outputs invalidated by the
change.
Actions after QC
| Finding | Main action |
|---|---|
| Variable, row unit, or population unclear | Return to CONTEXT QC |
| General analysis specification unclear | Return to PLAN QC |
| Clinical-trial statistical specification unclear | Return to SAP QC |
| Code and result disagree | Review the R script |
| Code, data, or conditions changed | Rerun the analysis in R |
| Values are correct but presentation is incomplete | Revise the table or figure |
| Result narrative is inappropriate | Revise the narrative |
Decision examples
Internal-consistency issue
ID: M-001
Severity: Major
Issue: The percentage for Treatment A in Table 2 does not match the displayed numerator and denominator.
Evidence: Table 2, "Any adverse event": 18/60, displayed as 25.0%; 18/60 = 30.0%.
Recommended action: Review the aggregation code, transcription, and denominator definition; regenerate Table 2.Review limitations
RESULT QC alone cannot establish source-data accuracy, complete correctness of all R code, independent reproducibility of every calculation, optimal statistical methods, complete SAP conformance, or final scientific validity. Record the assessed scope and unresolved items, then hand them off for correction or human review.