What is real-world evidence, and what can it actually answer?

Real-world data is health information collected outside a research protocol: electronic records, claims, registries, pharmacy dispensing, device outputs. Real-world evidence is the conclusion derived from analysing it.

The distinction matters because the second is a claim and the first is only material.

What it is good at

• Populations trials exclude: older patients, multiple comorbidities, pregnancy, advanced kidney disease.
• Long follow-up, at a scale trials cannot fund.
• Rare outcomes, where a trial would be impossibly large.
• Treatment patterns as they actually occur, rather than as protocols specify.
• Generating hypotheses precisely enough to be worth testing.

What it is bad at

Anything that depends on the comparison groups being alike. Nobody randomised these patients. Whoever received the treatment received it for reasons, and those reasons are usually related to their prognosis.

That is confounding by indication, and it is not a limitation to be listed at the end — it is the central methodological problem of the design, and the analysis has to be built around it.

The bias that catches people

Data collected for care records what clinicians chose to record. A variable missing from the record is not a variable that was normal; it is a variable nobody measured, and why nobody measured it is often related to the outcome.

Codes are entered for billing and administration, not for research. Their accuracy varies by outcome, by site and over time, and misclassification that differs between exposure groups biases the estimate in a direction that is hard to predict.

What makes a real-world study credible

• A protocol written before the data are examined.
• An explicitly stated target comparison — ideally the trial you would run if you could.
• A confounding strategy chosen from causal reasoning, not from a variable screen.
• Handling of time: when follow-up starts, when exposure is defined, and how the gap between them is treated.
• Sensitivity analyses that could plausibly overturn the result, reported whether they do or not.
• Conclusions that stop where the design stops supporting them.

The last is the one that most often fails. A well-conducted observational study with a carefully worded conclusion is more useful than the same study with a causal claim attached.