Three terms, one direction of travel, and they are not interchangeable. Data is the record that routine care left behind. Insight is the reading of that record: what it appears to show, and whether it is worth pursuing. Evidence is the conclusion that survives a specified method, a stated limitation and an independent reviewer. This page sets out the sequence, then works through every data type, what each one actually contains, and the questions each one can and cannot answer.
Most confusion in this field comes from collapsing three stages into one word. They are sequential. Each stage costs more than the one before it, and each is defensible against a different audience.
Routinely collected information about health status or the delivery of care: claims, dispensing, hospital activity, clinical records, laboratory values, registries, devices. Nobody designed it to answer your question. That is simultaneously its greatest strength, because it is unselected and at scale, and its central difficulty, because the thing you want measured may never have been recorded.
Once the data is structured, de-duplicated and correctly denominated, it will tell you how large a population is, what is being used and in what order, where the gaps sit and which direction the trend runs. This is descriptive work, it is fast, and for most commercial and medical planning decisions it is sufficient. It is not proof of anything, and should not be presented as though it were.
A causal or comparative claim about use, benefit or risk, generated from specified data by a method fixed before the data was seen, with confounding addressed, sensitivity analyses run and limitations declared. Change the method and you change the result, which is why the method note, not the dataset, is the part a regulator, a payer or a journal actually assesses.
The distinction is not pedantry. A stakeholder who asks for "real-world data" almost always wants insight or evidence. A supplier who delivers a dataset has technically complied and practically failed. Equally, a descriptive read presented as evidence is the most common way a well-run analysis loses its credibility in review.
Each step is where a study is usually lost. None of them is the analysis, which is the fastest part of almost every programme.
Population, exposure, comparator, outcome and time window, written down before anyone looks at the data. A question that cannot be written in that form cannot be answered from routine data.
Does the cohort exist at usable size, is the outcome actually captured in this data type, and is the follow-up long enough. This is where a question gets refused, and refusing it early is the cheapest outcome available.
Codes, algorithms and washout periods turned into a reproducible definition. Two teams using different definitions on the same data will report different numbers, and both can be right.
Nobody was randomised. A difference between groups may reflect the treatment, or it may reflect why each patient received it. Matching, adjustment and negative controls are how that is tested rather than assumed.
What the data could not see, which results moved under sensitivity analysis, and where the conclusion stops. An analysis with no declared limitation has not been examined closely enough.
Where it usually goes wrong. Not in the modelling. In an outcome that was never recorded, a denominator that was a projection rather than a count, a cohort defined after the first result was seen, or a comparison drawn across two markets whose records cannot be joined.
"Real-world data" is not one thing. It is at least twelve, and they differ more from each other than most buyers expect. The single most common error in this field is asking a data type for something it never recorded. Read the last column first.
| Data type | What the record actually contains | Questions it answers well | Where it stops |
|---|---|---|---|
| Claims, government | A payment record created whenever a subsidised medicine is dispensed or a subsidised service delivered. Item, date, quantity, benefit paid, prescriber type, patient category. | Market size and share on a counted denominator, treatment sequencing, persistence from dispensing gaps, cost and utilisation, the signature of a policy change. | It records what a scheme paid for, not why. Usually no diagnosis on a prescription, no clinical result, and nothing about privately funded care. |
| Claims, insurance | The same shape of record with a private or employer-based insurer as payer. Billing codes, dates, amounts, and frequently attached laboratory values. | Utilisation and cost in the insured population, treatment patterns outside public funding, budget impact, comparative effectiveness where labs are attached. | The covered population is not the whole population. Employer schemes skew to working age, so elderly and unemployed patients are under-represented. |
| EMR and EHR | What clinicians wrote down: diagnoses, orders, results, notes, often free text. The richest clinical detail available in routine care. | Standard of care as practised rather than as guidelines describe it, clinical outcomes including response and progression, cohorts defined clinically, reason for discontinuation. | Coverage is institutional, not national. A patient who moves between hospitals usually restarts. Free text is often excluded or needs separate permission. |
| Hospital activity | Counts of what hospitals did: admissions, procedures, separations, length of stay, day cases, emergency presentations. Usually facility or jurisdiction level. | Where a procedure is performed and at what volume, facility-level targeting, peer benchmarking, capacity as a proxy for site feasibility, procedure trends. | Tells you what was done, not what it cost the hospital, who decided it, or what happened to the patient afterwards. Rarely patient level. |
| Dispensing | The record created when a medicine is physically handed over at a pharmacy. Closer to real use than a prescription, which may never be filled. | Actual uptake rather than intent to treat, refill and persistence patterns, within-class switching, seasonality and stockpiling, week-by-week launch tracking. | Shows the medicine leaving the pharmacy, not the patient taking it. Under-co-payment dispensing is incompletely captured in some schemes. |
| Laboratory results | Numeric values from diagnostic tests, with dates, so trajectories rather than single points. | Response and progression rather than treatment alone, biomarker-defined cohorts and eligible volume, testing rates and testing gaps, severity stratification, early safety signals. | Reference ranges and units differ between laboratories, so harmonisation is real work. Where only a selected subgroup is tested, the tested population is not the treated population. |
| Registry | A collection built on purpose for one disease, procedure or product. Captures fields nobody records elsewhere: stage, grade, histology, biomarker status. | Disease epidemiology at a defined completeness, staging and histology, outcome and survival research with a validated denominator, quality benchmarking, cohort enrichment. | Contains only what it was built to capture, for only the population it enrols. Reporting lags are usually longer than claims because cases are curated. |
| Biobank | Biological samples linked to health information about the people who gave them: questionnaire, physical measurement, genotype, and often follow-up. | Genetic risk and polygenic scores, biomarker discovery and validation against real outcomes, target identification, exposure analysis alongside clinical data. | Participants volunteer, so the cohort is not a random sample. Sample sizes limit rare-event work, and follow-up intervals are set by the study, not by care. |
| Genomics and precision medicine | Sequencing and molecular profiling results, ideally joined to the clinical record of what was then done and what followed. | Population-specific variant frequency, biomarker-eligible population sizing, testing pathway analysis, matching of molecular subtype to treatment and outcome. | A sequence with no linked outcome answers a biology question, not a clinical one. Linkage, not sequencing, is almost always the binding constraint. |
| Linked data | Two or more sources joined at the person level under an approved linkage, most often claims with hospital, or a registry with dispensing. | Full pathway questions that no single source can answer: diagnosis through treatment to outcome, cost across settings, and true incident-case identification. | Requires a custodian-approved linkage and usually a formal approval pathway, which sets the timeline. Records cannot be joined across national systems. |
| Population data | Census, demographic, socioeconomic and geographic reference data published by statistical agencies. | Denominators, per-capita normalisation, catchment and territory design, equity and access analysis, standardisation for age and sex. | No clinical content at all. It is the denominator layer that makes other data types comparable, and useless on its own for a clinical question. |
| Epidemiological studies | Published prevalence, incidence, survival and burden estimates, derived from surveys, cohorts and surveillance systems. | Baseline burden and unmet need, sanity checks against a claims-derived estimate, market sizing where record-level access does not exist. | Published aggregates cannot be re-cut. If your question needs a subgroup the authors did not report, the estimate cannot answer it. |
Two data types beat one. Claims describe what was funded, laboratory values describe what happened, and a registry describes who the patient was. Any two of those answers materially more than either alone, which is why the linkage question, rather than the volume question, decides what a programme can realistically deliver.
The same dataset serves different functions differently. What follows is the mapping we use when scoping: the decision on the left, the data types that can carry it, and whether the honest output is an insight or evidence.
| Decision | The question, stated properly | Data types that can carry it | Honest output |
|---|---|---|---|
| Unmet need | How many patients reach a given line of therapy each year, and what happens to those who fail it. | Claims, dispensing, registry, linked data | Insight, becoming evidence once the outcome is captured rather than inferred |
| Market sizing | How many treated patients exist in this indication, on a counted rather than modelled denominator. | Claims, dispensing, population data | Insight |
| Treatment pathway | In what order are therapies actually used, for how long, and where do patients switch or stop. | Claims, dispensing, EMR, linked data | Insight, robust when built on dispensing gaps rather than prescriptions |
| Comparative effectiveness | Do patients on A experience different outcomes from comparable patients on B. | Linked data, EMR with laboratory results, registry | Evidence, and only with a pre-specified confounding strategy |
| Safety in practice | What is the rate of an event among treated patients, against a defensible denominator. | Claims, dispensing, EMR, laboratory results | Evidence for rates, insight for signals |
| Trial feasibility | How many eligible patients exist under these criteria, and at which sites. | Claims, EMR, registry, hospital activity | Insight, and it is the cheapest study you will ever run |
| External comparator | What would have happened to these patients under standard care. | Registry, linked data, EMR with outcomes | Evidence, subject to the outcome being measured the same way |
| Health economics | What does this pathway cost in this system, and what changes if the mix shifts. | Claims, hospital activity, linked data, population data | Evidence for cost, insight for the counterfactual |
| Reimbursement submission | What does the local population look like, and what do local inputs do to the global model. | Claims, registry, epidemiological studies, population data | Evidence, and the local inputs are usually the part that is challenged |
| Biomarker strategy | Who is tested, who is not, and how many patients are eligible but never identified. | Laboratory results, genomics, EMR, registry | Insight, becoming evidence where testing and treatment are linked |
| Field and territory design | Where is the procedure performed, at what volume, and by which facilities. | Hospital activity, claims, population data | Insight |
| Launch performance | Is uptake tracking to plan, in which segments, and against what baseline. | Dispensing, claims | Insight, week by week rather than quarter by quarter |
Neither is superior. They answer different questions, and the most common error in the field is asking one of them to do the other's job. A trial establishes whether a treatment can work under controlled conditions. Real-world data establishes what happens when it meets an actual population.
| Real-world data | Controlled clinical trial | |
|---|---|---|
| What it measures | Effectiveness. What happened in practice | Efficacy. What happened under protocol |
| Setting | Routine clinical practice | Controlled research conditions |
| Who is included | No strict inclusion criteria. The population as it is | Strict criteria, applied before enrolment |
| Who drives the data | Patient-centred. Care happened, a record followed | Investigator-centred. Data collected on purpose |
| Comorbidities and interactions | Present, because real patients have them | Included only where the protocol allows |
| Role of the physician | Multiple physicians, chosen by the patient | A designated investigator |
| Comparator | Whatever the market and the physician actually chose | Placebo or a defined standard of care |
| Treatment | Variable, as determined by practice | Fixed, according to the protocol |
| Response monitoring | Variable, whenever care happened to occur | Continuous, on a schedule |
| Follow-up | Determined by clinical practice | Defined by the protocol |
| Assignment to treatment | Chosen by clinician and patient, so confounded by indication | Randomised, which is what makes the comparison causal |
| Strongest use | Generalisability, long-term outcomes, cost, and populations a trial excluded | Establishing that an effect exists at all, under control |
The practical consequence. Because nobody randomised anyone, a difference between two groups in real-world data may reflect the treatment, or it may reflect why each patient received that treatment in the first place. Handling that is the entire craft: cohort definition, covariate control, sensitivity analysis, and the honesty to say when the design cannot settle the question.
The same analysis is judged differently depending on who receives it. Knowing which audience the output is for determines how far up the sequence the work has to go.
Better access to information has raised awareness of both trial and real-world findings, and with it the reporting of safety concerns, comorbidity effects and long-term outcomes. It has also raised willingness to take part in real-world studies at all, which is what keeps registries and biobanks viable.
A trial population rarely resembles a clinic list. Real-world evidence describes what happens to older patients, patients on four other medicines, and patients who would have been excluded from the registration study. The standard applied is whether the cohort looks like their patients.
Post-market safety and effectiveness at population scale, and a growing role in label extensions and single-arm settings where a randomised comparator is not feasible. The standard applied is whether the method was fixed before the data was seen.
They ask what a treatment does in their own population, at their own prices, in their own care pathway. That is a real-world question by definition. The standard applied is whether the local inputs are local, and they will check.
Unmet need, comparator selection, feasibility, launch tracking, safety context and reimbursement evidence. Most commercial questions inside a life sciences company are real-world questions wearing a different name. The standard applied is whether a decision can be made on it.
Service planning, quality improvement and resource allocation, using the record the system already generates rather than commissioning new collection. The standard applied is whether the finding holds at facility level, where the budget sits.
Why standardisation has become the live issue. Registered real-world studies have grown steadily, and with them the importance of standardising how such work is conducted and reported. An unregistered, unspecified real-world analysis is difficult to distinguish from an exercise in finding the answer somebody wanted. Pre-specification and a published method note matter more here than in almost any other setting.
Describe the decision you are trying to make. We will tell you whether routine data can answer it, from which data type, in which market, and how long it would take.
