Ask Mallard
Core concept

Registry and routinely collected data

4 min read

A registry or routine-data study uses records made for care, billing or surveillance to answer a research question those records were not collected for.

Fit

When it matters, and when it does not

It matters when

  • Your question needs large numbers, long follow-up or rare outcomes that a new study could not reach.
  • You want unselected people, including older and sicker people whom trials often exclude.
  • The records hold the exposure, the outcome and their dates, or can be linked to sources that do.

It matters less when

Before you start

What to check in your study

Decide these first
  1. Who the registry coversWhich people, places and years are included, and how complete capture is within them.
  2. Who enters each record, and whenA clinician at the visit, a coder after discharge or a clerk months later. It shapes accuracy and dates.
  3. The code lists and algorithms for each variableExplicit and reproducible, so another team could rebuild the same cohort from the same data. A rule that finds a condition in records is called a phenotype; validate it against charts.
  4. Any linkage between sourcesWhich identifiers join the records, how many fail to link, and whether failures differ between groups.
  5. Permission for research useRecords made for care need approval for secondary use under the rules that govern that data source.
  6. A missing-data plan that starts from an assumptionDecide why coded values may be missing before choosing a method. The assumption picks the method, not the share missing.
It rests on these
  1. Each key code records what it claims to record.How to check: Check key codes against a sample of charts, and report how accurate they were.
  2. Events are captured the same way across places and years.How to check: Ask whether capture differs by hospital, region, insurer or year. A change in coding rules can look like a change in disease.
  3. Linked records belong to the same person.How to check: Report the linkage method and its error rate. A missed link can look like a missing event.
  4. Missing values follow the assumption you named.How to check: Coding is often less complete at some hospitals and in some years, so missing completely at random (MCAR) is rarely plausible. Missing at random (MAR) means the gaps depend only on recorded factors, such as hospital, year and age. Under MAR, multiple imputation can use those factors, but it is not automatically better than analyzing complete records. If a value is left blank because of what it would have shown, missing not at random (MNAR) is plausible. Then add a tipping-point analysis, which shifts the imputed values until the conclusion changes.
  5. For a causal question, the confounders are recorded.How to check: List them from a causal diagram. Negative-control outcomes can reveal confounding that remains.

Plain language

How it works

A registry is a data source, not a design. Records made for care, billing or surveillance can support several study designs. Choose the design from the question, then check whether the data can supply its population, dates and measurements.

Who the registry holds
Who a registry study can see. Of everyone the question is about, some enter the registry and some never do, because they were not seen, not coded or not submitted. The study cohort is drawn from the registry alone, so people who never entered cannot be in it. Linking the cohort to other sources adds data where identifiers match, and records that fail to link can look like missing events.Registry study: who the records hold, and who they never seeEveryone the question is aboutIn the registry: seen, coded and submittedNever entered: not seen, not coded or notsubmittedStudy cohort, defined by a published codelistLinked to other sources where identifiersmatchFailed to link: events can look missingAn estimate describes the people the registry holds. Who never entered decides how far it reaches thepopulation in the question.
Conceptual diagram. It shows the structure of the idea, not data from any study.
Check the source before choosing the analysis

Review coverage, code definitions, linkage quality and changes in recording. A code records what was documented, not everything that happened. Use the linked cohort or case-control page to set time zero and comparison groups. Follow the missing-data assumption through the analysis.

Worked example

A worked example

Invented for illustration
Setting
A regional hip fracture registry, one year, with every hospital in the region submitting.
Question
What share of hip fracture operations happen within 36 hours of admission?
Approach
Operations are found by procedure codes checked against a sample of charts. The registry's admission and operation times are used, after checking how each is recorded.
Result
672 of 960 operations happened within 36 hours: 70 in 100.
What it means
If one hospital had stopped submitting, its operations would be missing. The share would then describe only the hospitals that report, not the region.

In practice

Recent studies PubMed search · not vetted

Searching PubMed…

Watch out

Common mistakes

The one to remember

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.

  • Using the date a record was entered as the date the event happened.Instead: Check how each date is recorded, and allow for any lag when setting exposure and outcome windows.
  • Starting follow-up at a date defined by later events.Instead: Set one clear time zero, and place the exposure and outcome windows relative to it. This avoids immortal time bias.
  • Excluding people whose exposure the records leave unknown.Instead: Those left may differ from those dropped, which is selection bias. Keep them, and use multiple imputation under a stated assumption.
  • Trying analyses until one works, because the data are already on hand.Instead: Prespecify the analysis, and register it where you can.

Inside Mallard

How Mallard uses this

Plans do not link to this concept yet. A plan built on registry data links to the design it names, such as a retrospective cohort.

No Mallard code example points to this page.
Atlas node concept.registry, the id a plan cue records.

Related Learn pages

Related designs and methods

Evidence

Sources & further reading

  1. 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
  2. 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
RECORD is a reporting guideline extending STROBE to routinely collected data, and the second reference is the STROBE statement it extends. Both govern what is disclosed rather than how the study is designed. Source records rechecked against PubMed on 2026-11-07: record verified, confirm it supports this use. No independent professional or methodological approval is claimed.
Last source review: 2026-11-07
AI-generated educational content
Mallard uses reputable methodological and reporting sources, but this material may contain errors or omit important context. Verify important methodological decisions with the cited sources and, when appropriate, a qualified research methodologist or statistician.