From real-world data to decision-relevant evidence
Real-world evidence is not produced by data alone. It emerges from the combination of a well-defined research question, data that are fit for purpose, an appropriate study design, transparent analytical choices and an interpretation that reflects the strengths and limitations of the evidence.
This page provides a practical overview of the main components of real-world evidence generation. The emphasis is on durable principles and institutional resources rather than on a continually changing list of publications.
1. Foundations of Real-World Evidence
Real-world data (RWD) describe information relating to health status, healthcare delivery or patient experience collected outside the context of highly controlled clinical trials, including both routinely collected data and data gathered prospectively in real-world settings. Real-world evidence (RWE) is the evidence generated when such data are analysed to address a research or decision-making question.
The distinction between data and evidence is important. A data source does not become evidence simply because it contains information on patients or treatments. The credibility of the resulting evidence depends on the question being asked, the suitability of the data, the study design, the analytical approach and the transparency of the process.
RWE can complement evidence from randomised clinical trials by addressing questions about treatment patterns, natural history, safety, effectiveness, healthcare utilisation, longer-term outcomes and populations that may be underrepresented in trials. RWD can also contribute to randomised studies, so RWE should not be treated as synonymous with non-randomised research.
Institutional starting points
FDA — Real-World Evidence
Definitions, guidance and regulatory resources on the use of RWD and RWE across medical-product development and oversight.
EMA — Real-world evidence
Information on the European medicines regulatory network’s use of RWD and RWE, including regulatory-led studies and DARWIN EU.
2. Fit-for-Purpose Data
No real-world data source is universally good or poor. Its value depends on whether it is suitable for a particular research question and intended use. A source that is well suited to studying treatment utilisation may be inadequate for estimating comparative effectiveness, and a rich clinical database may still be unsuitable if key confounders or outcomes are not captured reliably.
Assessment should consider whether the data adequately capture the population, exposure, comparator, outcomes and important covariates; the duration and completeness of follow-up; the provenance and meaning of variables; missingness and measurement error; coding and validation; changes in data capture over time; and the feasibility and validity of linkage to complementary sources.
Data quality and relevance are therefore inseparable from the intended analysis. Fitness for purpose should be assessed before the study is conducted and documented clearly enough for others to understand the strengths and limitations of the evidence. A formal feasibility assessment can then determine whether the available data can support the proposed design before the protocol is finalised, including whether there are sufficient eligible patients, exposure and comparator information, outcomes, key covariates, follow-up and expected precision.
Selected frameworks
EMA Data Quality Framework for medicines regulation
An EU framework for assessing data quality, including specific guidance for the fitness-for-use of real-world data.
FDA — Assessing EHR and medical claims data
Considerations for the use of electronic health records and claims data in studies intended to support regulatory decisions.
3. Study Design & Causal Inference
Credible evidence starts with a clearly specified question. For comparative-effectiveness research, this means defining the population, treatment strategies, comparator, time zero, follow-up, outcomes and causal contrast before choosing an analytical method. Many important biases originate in the way these elements are aligned rather than in the statistical model itself.
Design strategies such as active-comparator and new-user designs, explicit alignment of eligibility and follow-up, and target-trial emulation can reduce avoidable sources of bias. Statistical approaches such as regression adjustment, matching, weighting and propensity-score methods can then address measured differences between treatment groups, provided their assumptions are considered carefully.
Residual confounding, selection bias, immortal-time bias, measurement error and informative missingness remain important threats. Sensitivity and quantitative bias analyses can help assess how dependent conclusions are on assumptions that cannot be verified directly from the data.
External controls can be valuable when randomisation is infeasible or inappropriate, but they require particular attention to comparability of populations, calendar time, outcome ascertainment, treatment pathways and the definition of the start of follow-up. More broadly, internal validity does not guarantee that an effect estimate applies to the population or setting of interest. Generalisability and transportability therefore require attention to differences in patient characteristics, care pathways, treatment-effect heterogeneity and other contextual factors that may modify the observed effect.
Methodological guidance
NICE — Methods for real-world studies of comparative effects
Recommendations on target-trial approaches, time-related bias, external controls, confounding and assessment of robustness.
EMA — Reflection paper on use of RWD in non-interventional studies
European regulatory guidance on planning non-interventional studies using real-world data, including bias, confounding, data quality and study feasibility.
4. Conduct, Transparency & Reproducibility
Good design needs to be accompanied by transparent conduct. Protocols and statistical analysis plans should define the principal design and analytical decisions before results are known, particularly when the study is intended to support consequential clinical, regulatory or reimbursement decisions.
Transparent studies describe cohort definitions, code lists, variable construction, data transformations, validation activities, analytical choices and deviations from the original plan. Sensitivity analyses should be distinguished from post hoc exploration, and changes made during the study should be documented with their rationale.
Where governance permits, reusable analytical code, computable phenotype definitions and common analytical specifications can improve reproducibility across data sources. Registration of studies and access to protocols also help make the evidence-generation process more visible and reduce unnecessary analytical flexibility.
Transparency and study conduct
ICH M14 — Non-interventional studies using RWD
International principles for planning, designing, analysing and reporting non-interventional studies that use real-world data for medicine safety assessment.
HMA–EMA Catalogues of Real-World Data Sources and Studies
A structured European environment for information on RWD studies, protocols, data sources and research networks.
STaRT-RWE — Structured template for planning and reporting RWE studies
A practical template for specifying study design and implementation decisions in a transparent and reproducible format.
5. Regulatory & HTA Use of RWE
Regulators and health technology assessment bodies do not evaluate RWE simply according to whether a study is labelled “real world”. The evidentiary value depends on the decision being informed, the suitability of the data, the study design, analytical validity, transparency and the degree of uncertainty that remains.
Regulatory uses include disease epidemiology, medicine utilisation, safety and effectiveness, feasibility of clinical studies, external controls and post-authorisation evidence generation. HTA questions may place additional emphasis on comparative effects in routine practice, generalisability, long-term outcomes, resource use, patient-relevant outcomes and the characterisation of uncertainty.
Requirements differ according to jurisdiction and intended use. When RWE is being generated for a specific regulatory or HTA decision, early engagement with the relevant body can be as important as the technical design of the study.
6. Analytics, Standards & Tools
Analytical tools are most useful when they support good design rather than substitute for it. Reproducible workflows, standard vocabularies and common data structures can make analyses easier to audit, reuse and implement consistently across multiple databases.
Common data models can harmonise the structure and representation of heterogeneous observational data. When combined with shared analytical specifications, they can support distributed research in which patient-level data remain under the control of local data partners while the same analytical question is implemented consistently across sites.
Open statistical environments such as R support transparent and reproducible analysis, while the OHDSI ecosystem provides a mature example of combining the OMOP Common Data Model, standardised vocabularies and reusable analytical tools for observational research.
Core resources
OHDSI — OMOP Common Data Model & Analytical Tools
An open-source ecosystem combining the OMOP Common Data Model and standardised vocabularies with analytical tools for reproducible observational research across heterogeneous data sources.
The R Project for Statistical Computing
An open environment for statistical computing, graphics and reproducible analytical workflows.
7. Integrated Evidence Generation
Evidence needs rarely sit within a single methodological discipline. Questions arising during the development and use of a health technology may require information from clinical trials, epidemiology, real-world studies, health economics and outcomes research, patient-reported outcomes, digital measures, biomarkers and other sources.
Integrated evidence generation starts from the decisions that need to be made and the evidence gaps that need to be addressed. Those gaps should reflect the needs of the relevant stakeholders, including regulators, HTA bodies, clinicians, payers and patients, and where appropriate those stakeholders should help shape the questions, outcomes and evidence priorities. The strategy then coordinates different evidence streams across the product lifecycle rather than treating individual studies as isolated activities.
This approach can align populations, endpoints and research questions; identify where evidence can be reused across stakeholders; clarify the sequencing of studies; and make explicit which method is best suited to each question. Real-world evidence is one important component of this broader strategy, contributing where routine healthcare data and real-world settings add information that other evidence sources cannot provide as efficiently or directly. Integration also requires synthesis and triangulation across evidence streams: examining where findings converge or diverge, whether differences can be explained by populations, estimands, study designs or biases, and what uncertainty remains after the evidence is considered as a whole.
A practical sequence: define the decision and research question → identify stakeholder evidence needs → identify potentially relevant data → assess fitness for purpose and feasibility → choose the design and analysis → conduct and report transparently → assess applicability → triangulate and interpret the evidence in its regulatory, HTA or clinical context → integrate it with the wider evidence plan.