Before collecting anything
- The protocol and approved/authorised data source are known.
- Every variable has a reason to exist.
- A data dictionary exists before extraction/recruitment begins.
- Identifiers and research variables are separated where feasible.
- Access roles and storage location are authorised.
- Missing-value codes and date/unit conventions are defined.
Step-by-step workflow
Map each objective to required variables
For every objective and planned analysis, identify the outcome, exposure/intervention, covariates/confounders and operational definitions. Remove variables with no analytic or governance justification.
Create the data dictionary
For each field document variable name, human-readable label, type, units, allowed values, coding rules, source, time point, missing-value code and derivation rule.
Design the collection form
Use consistent field order, controlled options where appropriate, validation ranges, explicit unknown/not-done choices, and instructions for ambiguous source records.
Pilot the form
Test on a small number of permitted records or mock cases before full collection. Check whether different reviewers interpret variables consistently and whether crucial information is routinely unavailable.
Control identifiers
Use the minimum identifiers necessary. If a linkage key is required, keep it separately with restricted access according to the approved data-management plan.
Record provenance
Know where each field came from and, where relevant, which date/time point or source document generated it. Derived variables should have a documented formula.
Perform quality checks
Check impossible values, duplicates, inconsistent dates, outliers, missingness patterns and cross-field logic. Never overwrite raw values without an auditable cleaning rule.
Freeze an analysis dataset
Create a dated/versioned analysis dataset after cleaning. Subsequent changes should generate a new version with a documented reason.
Archive according to policy
Retain or destroy data, keys, consent documentation and study files according to the approved protocol, institutional requirements and applicable regulation.
Minimum data-dictionary fields
| Field | Purpose |
|---|---|
| variable_name | Short analysis-safe name. |
| label | Full human-readable definition. |
| type | Numeric, date, categorical, text, Boolean, etc. |
| allowed_values | Permitted categories/range and units. |
| source | Where the value is obtained. |
| time_point | When it is measured. |
| missing_rule | How unavailable/not applicable/not measured are distinguished. |
| derivation | Formula or rule for calculated variables. |
Governance and privacy
Use only authorised data
A dataset being technically accessible does not establish permission for research use.
Minimise
Collect only what is necessary for the approved objectives and analyses.
Do not place patient data here
The Research Hub is an operational/educational portal, not a clinical research data-capture system.
Final checklist
- Variable definitions are unambiguous.
- Units and time points are specified.
- Missing data are coded consistently.
- Raw data are preserved or otherwise handled according to authorised procedures.
- Cleaning rules are documented.
- Analysis data are versioned.
- Access is limited to authorised team members.
- Retention/destruction requirements are known.
Collection form design rules
- Use one variable per field. Do not combine several concepts into free text when they need separate analysis.
- Define units explicitly: kg versus lb, mg/dL versus mmol/L, hours versus days.
- Separate “not measured,” “unknown,” “not applicable” and true zero where those meanings differ.
- For dates, define which date matters: diagnosis, admission, procedure, sample, discharge, follow-up or outcome.
- For repeated measurements, use a clear time-point structure rather than adding ad-hoc columns later.
- Do not infer clinically important values from ambiguous notes without a prespecified abstraction rule.
Quality-control plan
Range and logic checks
Define impossible or implausible values before analysis. Examples include impossible dates, values outside physiologic/technical ranges, or an outcome date preceding eligibility.
Duplicate checks
Define how duplicate records or repeated encounters are detected and which record represents the analysis unit.
Inter-rater consistency where abstraction is subjective
Train reviewers on a small common sample, resolve disagreements and update the abstraction manual before scaling up.
Change log
Record corrections, recoding decisions, exclusions and derivations. A final number should be traceable back to the defined source and transformation rule.
Analysis handoff package
When data collection is complete, the analyst should receive more than a spreadsheet. Provide the frozen dataset version, data dictionary, protocol, statistical analysis plan, derivation rules, inclusion/exclusion log, missing-data definitions and a short readme explaining the analysis unit and key dates.
Common data-management failures
- Changing variable definitions halfway through collection without versioning the change.
- Using blank cells for several different meanings.
- Keeping the only linkage key in the same broadly shared file as the research dataset.
- Collecting dozens of “maybe useful” variables that were never justified in the protocol.
- Overwriting raw values after cleaning, leaving no auditable source.
- Letting each reviewer interpret ambiguous clinical documentation differently.