Data Standards and Why They Are a Regulatory Question

Part of Clinical Data Management and Integrity: Coding and Data Standards.

About 5 minutes. Free, in your language, no account needed.

Why this matters

Standards look like a technical convenience: agree on a structure, and tools work across studies.

They are more than that. Several authorities require submitted data in a defined standard format, which makes conformance a condition of a submission being accepted rather than a preference. A study whose data cannot be produced in the required structure has a regulatory problem, not a formatting one.

The three layers

CDASH describes how to collect. It specifies field names, formats and conventions for the case report form, so that what sites enter aligns with what the downstream structures expect.

SDTM describes how to structure collected data for submission. It organizes observations into domains by type, each with defined variables, and it is the tabulation layer: a faithful representation of what was collected, not an interpretation of it.

ADaM describes how to structure data for analysis. It holds derived variables, flags for population membership, and the values the statistical analysis actually reads. It carries a specific requirement of traceability: it must be possible to get from an analysis result back through ADaM to SDTM to the collected value.

The relationship is the point. Collect in a structure that maps cleanly to tabulation, tabulate in a structure that supports analysis, and preserve the path back at every step.

Why collecting in the standard matters

A study can collect however it likes and map to the standard later. Mapping is always required and its difficulty varies enormously.

Collecting in a structure close to the standard makes mapping mechanical. Collecting in an arbitrary structure makes it interpretive, and interpretation performed months later by someone who was not present when the form was designed is where meaning gets lost.

The specific loss is a distinction that existed in collection and does not survive mapping. A form with three fields for a concept the standard expresses in one requires a decision about which to carry, and if nobody documents the decision the other two silently disappear.

Traceability

The requirement is that any number in a submission can be traced back to a collected value.

This sounds procedural and is the whole basis on which a reviewer decides whether to believe an analysis. A reviewer who wants to know where a particular number came from expects to follow it: from the table to the analysis dataset, to the tabulation, to the collected value, to the source.

A break anywhere in that chain means the number is an assertion rather than a derivation. The commonest break is a derived variable whose derivation is described in prose rather than specified, so two people implementing it produce different values and neither can be reconstructed.

Define files and the data guide

Submitted data comes with metadata describing it: what each dataset contains, what each variable means, its type, its controlled terminology, and how derived variables were derived.

This is not documentation in the optional sense. It is how a reviewer reads the data at all, and it is assessed. A submission with clean data and poor metadata is harder to review than one with adequate data and excellent metadata, and reviewers respond to that.

In practice

A data manager on a study intended for submission works backwards from the standard at design time rather than forwards from the protocol.

For each planned analysis, she identifies the analysis dataset variables, then the tabulation variables they derive from, then the collected fields those require.

Three gaps appear. An analysis requires a baseline flag with no rule defined for which assessment is baseline when a visit is repeated. A tabulation variable requires a unit that the form collects inconsistently. And a planned subgroup requires a stratification factor the form does not collect at all.

All three are cheap now. The third, discovered at analysis, would mean the planned subgroup analysis simply cannot be done.

What this looks like when it goes wrong

A study collects laboratory data in a custom structure because the electronic data capture system was configured before the standards mapping was considered.

At submission preparation, mapping to the tabulation structure takes four months rather than the planned four weeks.

The hard part is not the structure. It is that the collected data holds a field called "result comment," used by sites for several different purposes: sometimes explaining an abnormal value, sometimes recording that a sample was hemolyzed, sometimes noting a repeat draw.

The standard structure has defined places for each of those, and separating them requires reading several thousand free-text comments and assigning each an interpretation.

The resulting data is good. It was produced by a person, two years after the fact, deciding what each site meant. That is not traceable in the sense the requirement means, and the chain from the submitted value back to what the site recorded now passes through an undocumented judgment.

Key takeaways