Registry and routinely collected data studies
Why this level?
The design family and its defining features have not been specified for this hierarchy. NHMRC 2009 sourceA registry is a data source, not a design. The level follows whatever design is built on it: a cohort assembled from a registry is a cohort, and the questions to ask are about how completely and how consistently the registry was populated.
What question does this help answer?
Questions needing large numbers, long follow-up, rare outcomes or unselected populations: natural history, treatment patterns in practice, safety signals, and variation between providers.
What makes it attractive
Size, cost, and the absence of the selection that trial recruitment imposes. A national registry can include everyone with a condition, including the elderly and comorbid patients trials exclude, and can follow them for decades at a fraction of the cost of prospective collection.
What the data actually are
- A diagnosis code is a billing artefact as much as a clinical one, and its accuracy varies by condition, by incentive and by country. Validate the key codes against charts where possible, and report the validation.
- A missing laboratory value usually means the test was not ordered, which correlates with how well the patient was. That is missing-not-at-random and cannot be imputed away without an explicit assumption.
- Loss to follow-up in an administrative database means leaving the system, which is not a random event.
- The denominator is whoever is covered by the system, and coverage is rarely the population the question is about.
- Timing is recorded when the record was made, not when the event occurred, which matters greatly for exposure windows.
Design defences
Define the cohort with an explicit, reproducible algorithm, including the code lists, and publish them. Set the exposure window and the outcome window relative to a clearly defined time zero, since an ill-defined time zero is the usual source of immortal time bias here. Prespecify the analysis and, where feasible, register it, because the temptation to iterate against a dataset that is already sitting on disk is much stronger than in a prospective study. Use negative-control outcomes to detect residual confounding.
Why Mallard may recommend it
Mallard may point here when the description mentions a registry, claims data, an electronic health record extract, or a question about long-term natural history that no feasible prospective study could answer.
What to watch out for
The data exist because someone was billed or treated, not because someone was studying. So a variable is missing when nobody ordered the test, which is informative rather than random; a diagnosis code records what was CODED, which is shaped by reimbursement; and a patient disappears from the record when they change provider, not when they die. Coverage is the first question to answer and the one most often assumed: who is in this database, and who is systematically absent.
Clinical example
Question: What is the long-term renal outcome after a particular surgery? Design: Define the cohort from procedure codes validated against a chart sample, set time zero at the operation, follow through linked laboratory and dialysis records, and state explicitly who is missing from the database and how their absence could bias the estimate.
- The REporting of studies Conducted using Observational Routinely-collected health Data (RECORD) statement (opens in a new tab) · PLoS Med · doi 10.1371/journal.pmed.1001885
- Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement: guidelines for reporting observational studies (opens in a new tab) · BMJ · doi 10.1136/bmj.39335.541782.AD