What is clinical data management?
Clinical data management covers the design, capture, validation, correction, locking and archiving of a study's data. Its product is not a table of numbers: it is a dataset accompanied by everything needed to judge it — the definition of every variable, the origin of every value, the trace of every correction.
The discipline formed in regulated clinical trials, where a sponsor must be able to demonstrate to an authority that submitted data corresponds to the source record. Its requirements then spread to observational and academic research for a simple reason: a study that cannot show where its numbers came from cannot be reproduced, whatever its regulatory setting.
The cycle, in order
- Design — write the question, define the primary endpoint, settle the list of variables and what they mean.
- Structure — build the variable dictionary: type, measurement scale, unit, plausible bounds, permitted values, mappings to standards.
- Capture — enter or import values, each attached to its source.
- Validate — check consistency, resolve discrepancies, document every correction.
- Lock — freeze the database before analysis, so the dataset analysed is exactly the dataset described.
- Export and archive — deliver the data with its dictionary, metadata and decision log.
Why it determines how solid a study is
The most common failure is not fraud, it is forgetting. Three months after capture, nobody remembers whether the delay column counted days or weeks, whether missing values were absent from the record or never sought, or why twelve rows changed between two versions of the file.
These lapses do not show in the results. They show the moment someone asks to verify them — a reviewer, a committee, a colleague who wants to reuse the cohort. At that point a study without data management has nothing to offer, and the question "how did you measure the primary endpoint?" goes unanswered.
What the discipline actually protects
- Interpretability — the meaning of a column is written before the first value, not reconstructed afterwards.
- Comparability — declared, normalised units allow aggregation, comparison and meta-analysis.
- Inferential integrity — an endpoint fixed before capture forbids picking, after seeing the data, the one that yields a result.
- Reproducibility — a third party given the dataset, the dictionary and the log can retrace the path.
Who needs it
Clinical data management is often taken to be the business of industry sponsors and their clinical research associates. That reading is too narrow: most studies conducted worldwide are investigator-initiated — a hospital department mining its own records, a resident preparing a thesis, a team building a disease registry.
| Profile | What is at stake | The usual breaking point |
|---|---|---|
| Resident writing a thesis | Work that will hold up before a jury | A data collection form improvised in a spreadsheet, with no types and no units |
| Investigating clinician | A publication that survives peer review | Undocumented missing values, impossible to handle after the fact |
| Department head | A departmental base reusable study after study | Every study starts again from an empty file |
| Methodologist | A dataset ready to analyse | Receiving a cleanup job rather than a dataset |
| Committee / reviewer | Verifying what is claimed | No trace linking a published figure to a source record |
For non-regulated research the requirement is no lower — it is simply unpoliced. Nobody will audit a thesis, which makes the discipline a team imposes on itself all the more decisive.
How it is carried out
The variable dictionary
This is the keystone, and the piece most often missing. Each variable carries a stable technical identifier, a readable label, a type, a measurement scale, a unit, plausible bounds, the closed list of its categories where categorical, and the permitted reasons for absence. See the clinical data dictionary.
Validation rules
A bound that refuses an age of 210 takes seconds to write and avoids a manual correction months later. The useful checks are: possible and plausible bounds kept distinct, consistency between linked variables, temporal consistency, and completeness measured per variable rather than overall.
Traceability to source
Every value must be able to point back to what it came from — a registry field, a deterministic rule, a sentence in a report, a documented human decision. That is what data provenance covers, and it is what separates a verifiable figure from an asserted one.
Log and lock
The audit trail records who did what and when. The database lock separates the time of capture from the time of analysis. Without that separation, nothing technically prevents editing a value after seeing the result it produces.
Which standards apply
None of these standards is mandatory for academic research. Following them confers no certification: it makes the dataset interoperable, and legible to someone who does not know the team that produced it.
| Standard | What it normalises | Where it helps |
|---|---|---|
| ICH E6(R3) | Good clinical practice | The general frame: integrity, traceability, responsibilities |
| CDISC CDASH | Capture: which variables, under which name | Designing the collection form |
| CDISC SDTM | The structure of submitted data | Tabulation for transmission |
| UCUM | Units of measure | Making mg/dL and mmol/L comparable without ambiguity |
| STROBE | Reporting of observational studies | Writing up — and what must have been recorded to make it possible |
| FAIR principles | Reusability of research data | Metadata, identifiers, access conditions |
How AnzarSeha addresses it
AnzarSeha applies this discipline by construction rather than by procedure: the stages of the cycle are states of the software, not instructions to follow.
- The variable dictionary is generated when the study is created — typed, with UCUM units and mappings to standards — so it exists before the first value.
- The protocol locks before capture, and the lock refuses to engage until the primary endpoint is declared.
- Every value keeps its origin: registry field, deterministic rule, a model proposal citing its source sentence, or a human decision.
- The database freezes before analysis; reopening requires a stated reason, kept in the log.
- Export delivers the data with dictionary, metadata, missing-value table and decisions — readable by SPSS, R and pandas with no manual editing.
What data management does not solve
This needs saying plainly, because the opposite is regularly implied: impeccable data management does not make a study valid.
- It does not correct selection bias. A poorly assembled cohort stays poorly assembled, however clean the capture.
- It does not compensate for insufficient statistical power.
- It does not replace ethics committee approval or participant consent.
- It does not guarantee the clinical accuracy of the source: an erroneous report yields a value that is traceable, and wrong.
Its contribution is more modest and more solid: it guarantees that what is claimed matches what was collected, and that anyone wishing to check can.
Frequent questions
What is the difference between clinical data management and biostatistics?
Data management produces the dataset; biostatistics analyses it. The boundary is the database lock: before it, you define, capture and correct; after it, you analyse a dataset that no longer moves. Conflating the two phases is precisely what the lock prevents.
Does a departmental study need a dedicated data manager?
Not necessarily, but the function must be held by a named person. In investigator-initiated research it usually falls to whoever leads the work; the problem is rarely the absence of a dedicated post, it is the absence of a written method.
Can a spreadsheet be enough?
For entering data, yes. For managing it, no: a spreadsheet has no types, no units, no source and no log, and nothing stops a column being renamed or a value changed without a trace. This is treated in detail in the spreadsheet comparison.
When should the database be locked?
When capture is complete and discrepancies are resolved, and before the first analysis of the primary endpoint runs. Locking too early forces reopening; locking too late strips the lock of meaning, since the results are already known.