Real-world data: define the cohort before the model

Write an executable cohort definition

Specify the source population, eligibility criteria, observation window and index event in language that a programmer can implement without guessing. Record the coding system and version for diagnoses or treatments. A reproducible cohort is a chain of explicit decisions, not just a final list of identifiers.

Align time zero

Eligibility, treatment assignment and follow-up need careful alignment. If patients must survive long enough to qualify for an exposure group, the definition may introduce immortal-time bias. Work with the study’s epidemiology and statistical leads to resolve the design before turning it into a production derivation.

Show where patients leave the analysis

Build an attrition table with mutually understood exclusion steps and retained counts. Investigate whether exclusions concentrate in particular sites, dates or patient groups. Retain record-level reasons within the authorised analytical environment so totals can be reconstructed.

Distinguish missingness from absence

No recorded diagnosis is not automatically evidence of no disease. A missing laboratory result may reflect care patterns, data linkage or a genuine lack of measurement. Document source limitations and program sensitivity analyses where the protocol calls for them.

Deliver more than a dataset

Include lineage, a cohort specification, quality checks and a reproducible run record. FDA distinguishes routinely collected real-world data from the clinical evidence generated through its analysis. Programming supports that evidence chain; it cannot repair an unsuitable study design on its own.

Sources & further reading

• FDA — Real-world evidence

Sources checked 20 September 2026. Implementation commentary reflects a practical review perspective; source material may change.