Why data management decides whether a trial can be trusted
Why data management decides whether a clinical trial can be trusted: audit trails, validation checks and the data dictionary, with REDCap as the example.
By Manouchehr Hessabi, MD, MPH
A clinical trial's analysis can only be as reliable as the data record it runs on. Data management is the set of rules and systems that make that record complete, checkable and traceable, which is why the current international guideline for good clinical practice treats it as part of trial quality rather than as office work.
Why are data errors harder to see than design errors?
Readers usually judge a trial by its design: randomization, blinding, size, confidence intervals. Those features are visible in the published paper. Errors in the underlying data usually are not. A mistyped dose or a value quietly corrected without a record of the original can pass into an analysis without leaving a mark on the final table. Good data management exists so that, if anyone asks how a number came to be, there is an answer.
What do the current standards say?
The reference point is the International Council for Harmonisation's guideline on good clinical practice, ICH E6(R3). Its final version was adopted on 6 January 2025 (ICH, Step 4). The guideline defines good clinical practice (GCP) as an international ethical, scientific and quality standard for the conduct of trials that involve human participants.
The revised guideline includes a section titled "Data Governance," addressed to both investigators and sponsors. It describes its purpose as the appropriate management of data integrity, traceability and security, so that trial information can be accurately reported, verified and interpreted. Three ideas run through it, and each maps onto a practical piece of data management:
- What a data point means has to be defined (metadata and the data dictionary).
- Errors should be caught where data are entered (validation checks).
- Every change has to leave a trace (the audit trail).
The guideline also gives a compact definition of what good data look like. Source records, it says, should be attributable, legible, contemporaneous, original, accurate and complete (section 2.12.2). In plain terms: you can tell who recorded it; it can be read; it was recorded at the time the event happened; it is the first record of the observation; it is correct; and nothing required is missing.
What is a data dictionary, and why is it written before enrollment?
A data dictionary is the document that defines every variable a study will collect: its name, what it measures, its type (a number, a date, a category), its units, the values it is allowed to take, and how categories are coded. "Systolic blood pressure, millimeters of mercury, whole number between a set minimum and maximum" is a dictionary entry. So is "smoking status, coded 0 for never, 1 for former, 2 for current."
Writing this before the first participant enrolls forces decisions that would otherwise be made inconsistently (kilograms or pounds, whether "unknown" is an answer or a blank), and it fixes the meaning of each variable before anyone has seen results. A definition that shifts mid-study is an opening for confusion at best and bias at worst. ICH E6(R3) makes a related point for source records: the investigator should define what counts as a source record, the methods of data capture and their location before the trial starts (section 2.12.2).
Some data systems make the dictionary the center of the whole design. REDCap (Research Electronic Data Capture), an academic data-capture platform, is a widely used example. In the paper that first described it, Harris and colleagues presented REDCap as a metadata-driven methodology, in which "study-specific databases are created using data dictionaries" supplied by the research team and the system accomplishes its key functions through a single study metadata table (Harris et al., Journal of Biomedical Informatics, 2009). Metadata here simply means data about the data: the definitions, rules and context needed to understand each value. The forms, checks and exports all follow from one written definition, so they cannot quietly drift apart.
REDCap has since become an international collaboration. A 2019 paper by the same group reported that the REDCap Consortium had more than 3,200 partner institutions in 128 countries as of December 2018 (Harris et al., Journal of Biomedical Informatics, 2019). It serves here as an illustration; the principles apply to any well-run system, including paper.
How do validation checks catch errors at entry?
A validation check is a rule that tests a value as it is entered and flags it when it cannot be right. Common types include:
- Range checks: a body temperature of 3.7 degrees is almost certainly a missing digit.
- Required fields: a consent date cannot be blank for an enrolled participant.
- Consistency checks: a follow-up visit dated before the enrollment visit is impossible, even if each date is valid on its own.
- Branching logic: questions about pregnancy outcomes appear only where they apply, which removes a whole class of nonsensical entries.
When a check fires, the usual mechanism is a data query: a recorded question to the site asking someone to confirm or correct the value against the source. The query and its answer become part of the record. The early REDCap paper listed real-time data validation and integrity checks among the system's critical features (Harris et al., 2009).
ICH E6(R3) frames this in terms of risk rather than as a blanket rule. At the point of data capture, it says, automated data validation checks to raise data queries should be considered as required based on risk, and their implementation should be controlled and documented (section 4.2.1). The logic is proportion: a check on the primary outcome matters far more than one on a rarely used descriptive field.
What is an audit trail, and what does it protect against?
An audit trail is a secure record of every action taken on the data. The guideline's glossary defines it as metadata showing initial entry and changes to data fields or records, "by whom, when and, where applicable, why." The key property is that a change does not overwrite history. The original value, the new value, the person, the time and the reason all remain visible.
ICH E6(R3) is specific on this point. Changes to source records "should be traceable, should not obscure the original entry and should be explained if necessary" through an audit trail (section 2.12.2). Systems should document the initial data entry and any later changes or deletions, including, where appropriate, the reason for the change, and audit trails, reports and logs should not be disabled (section 4.2.2). The REDCap developers likewise described audit trail logging as a key feature from the system's earliest planning (Harris et al., 2009).
What does this protect against? Ordinary problems, mostly: a value corrected in good faith but wrongly, a file overwritten with an older version, or a later question about whether an outcome was recorded before or after treatment groups were revealed. An audit trail does not prevent mistakes. It makes them findable, so an independent reviewer can reconstruct how the dataset reached its final form.
Why does this belong in good clinical practice?
The revised guideline does not treat this as housekeeping. Its data governance section ties data handling directly to whether results can be verified and interpreted. The same document asks that systems for data capture and management be fit for purpose and implemented in a way that is proportionate to the risks to participants and the importance of the data collected (principle 9.2).
A finding becomes credible when others can check it, both by repeating the study in new data, the subject of a separate explainer on why replication matters in science, and by tracing a single study's numbers back to their source. Data management makes the second kind of checking possible.
What can a reader reasonably ask about a study's data?
Most readers will never see a trial's database. But a few questions, many answerable from a protocol, a registry entry or a methods section, say a good deal about how seriously data were handled:
- Were the outcomes and the analysis plan fixed in advance? A pre-specified plan and registered outcomes limit the room for choices made after seeing the data.
- How were data captured? Directly into an electronic system, transcribed from paper, or pulled from routine health records? Each step of transcription is a chance for error, which is why the guideline asks that unnecessary transcription steps be avoided.
- Were validation checks and an audit trail in place? A trial protocol or statistical analysis plan, where it is published, is the usual place to look.
- How were missing data handled? Missing values are not neutral. How they were treated can change a result, which is part of why the explainer on what intention-to-treat analysis means matters for reading trials.
- Were measurements defined and checked consistently across sites? The same question arises for laboratory measures, discussed in the explainer on biomarker validation versus qualification.
A study that cannot answer all of these is not automatically wrong. The questions simply ask whether the record behind the result could withstand a careful look.
Good data management does not make a finding true. It makes a finding checkable, which is the precondition for everything else science does with it. Readers interested in how these methods appear in applied work can find a list of peer-reviewed publications on this site.