Real-world evidence: when routine care data becomes research

In April 2019, the US Food and Drug Administration extended the approval of palbociclib, a breast-cancer drug, to cover men. There was no new randomized trial behind the decision. Male breast cancer is rare enough that running one would have taken many years — so the agency weighed electronic health records, insurance claims, and postmarketing safety reports instead. Records generated not by a study, but by the ordinary business of getting care.

That is real-world evidence doing exactly the job it was given: answering a question a trial would never practically reach.

Data first, evidence second

The field's most useful distinction is the one in its name. Real-world data (RWD) is the raw material — information about patient health and the delivery of care, collected routinely and outside conventional trials. Real-world evidence (RWE) is what a designed analysis of that material can support about the benefits or risks of a treatment.

The distinction does real work. 'We have real-world data' claims nothing; every hospital basement has real-world data. A pile of records becomes evidence only when someone commits to a question, a defined population, a comparison, and a way of being proven wrong. Most of the controversy around RWE is really controversy about analyses that skipped one of those steps.

Where real-world data comes from

Each source of real-world data has a character — a shape of what it sees clearly and what it cannot see at all.

Registries deserve their own treatment — they are the quiet workhorses of chronic-disease research, and we give them one in patient registries, explained. Patient-reported outcomes matter most exactly where lab tests fail: pain, fatigue, flares — the substance of invisible illness.

Increasingly, these sources are queried where they live rather than pooled into one warehouse: research networks send the analysis to the data and collect aggregate results back, a pattern we explain in federated analytics.

The law that opened the door

Regulators did not discover real-world data recently — safety surveillance has leaned on claims and records for decades. What changed is effectiveness. The 21st Century Cures Act of 2016 directed the FDA to evaluate the use of real-world evidence to support new indications for approved drugs and to satisfy post-approval study requirements. The agency published its framework for the real-world evidence program in December 2018 and has followed with a series of guidances on data quality, registries, and study design.

Safety came first for a reason. Detecting harm signals across millions of claims is exactly the kind of needle-counting routine data does well, and the FDA's Sentinel system has run that surveillance at national scale since the 2010s. Effectiveness is the harder promotion — it asks observational data to support the kind of claim a trial would have earned — which is why the framework leans so heavily on study design and data quality.

The direction of travel is unambiguous all the same: routine care data is now a recognized input to regulatory decisions — provided the analysis can survive the scrutiny a trial would have faced.

What it does well — and where it breaks

Real-world evidence is strongest where trials are weakest, and weakest exactly where trials are strong.

The strengths are structural. Trial populations are narrow — eligibility criteria trim away the elderly, the comorbid, the pregnant, the rural — while real-world data covers everyone who actually showed up for care, which is the population the treatment will actually meet. Real-world studies run faster and cheaper, reach rare diseases and rare outcomes that no trial can power, and follow patients across horizons no sponsor will fund. They can also ask questions no one will ever randomize: comparisons between two off-patent drugs, harms it would be unethical to assign.

The weakness is one word: confounding. In routine care, treatment is not assigned at random — sicker patients get treated differently, and a naive comparison will conclude that the treatment caused what the sickness did. Methods exist to narrow the gap — propensity matching, new-user designs, target-trial emulation — and the OHDSI community has shown how far disciplined, reproducible analysis at network scale can go. But randomization is the one trick observational data cannot replicate, and honest RWE never claims otherwise.

Two quieter failure modes deserve naming. Missingness in routine data is not random — sicker patients generate more records, and an absent lab value can mean 'normal,' 'untested,' or 'tested somewhere this database cannot see.' And outcomes are only as real as their documentation: a study counting emergency visits is actually counting coded claims for emergency visits, which is not quite the same thing. The craft lives in the details we cover in what is a research cohort? and data quality in health research — define the population precisely, measure what the data can actually see, and say so plainly.

Evidence is made upstream

Here is the part the acronym hides: the quality ceiling of real-world evidence is set long before any analysis runs. It is set at collection — by whether records are structured or scanned, by whether their provenance survives the journey, by whether the people they describe ever agreed to the use. No statistical method recovers what sloppy plumbing destroyed, and no finding is durable if the consent under it is shaky.

That upstream layer is what Mithras Health builds. Patient records live as FHIR R4 / US Core as the system of record, projected into OMOP views so researchers can run standard, reproducible analyses. Every research use runs under a per-study, revocable consent rather than a blanket waiver, federated queries execute in-region and return aggregates only, and patients are paid participation fees that post on IRB-approved research-use events. The aim is real-world data that arrives research-ready: structured, provenanced, and permissioned — so the evidence built on it does not wobble.

Routine care data becomes research the moment someone asks it a question. The work — the entire field, really — is making sure it can answer honestly: with permission attached, provenance intact, and enough structure to be believed.