FDA RWE: Considerations Regarding Non-Interventional Studies

What it is

— Real-World Evidence: Considerations Regarding Non-Interventional Studies for Drug and Biological Products is a draft guidance for industry issued by the U.S. FDA (CDER and CBER) on 20 March 2024 (Federal Register notice 2024-05969, published 21 March 2024; comment period closed 18 June 2024). It is one deliverable of FDA's Real-World Evidence Program, mandated by the 21st Century Cures Act and the 2016 FDA Reauthorization Act. Unlike an EQUATOR reporting checklist (STROBE, RECORD-PE) or a Cochrane/ISPOR appraisal instrument, this is an agency expectations document: it tells sponsors who intend to rely on a non-interventional study to support a regulatory decision what FDA will look for when judging whether the study can contribute to substantial evidence of effectiveness or to evidence of safety. It is maintained by FDA, not by a methods-guideline consortium, and it sits alongside two sibling FDA RWE guidances it must not be conflated with: the data-focused guidance on Assessing Electronic Health Records and Medical Claims Data To Support Regulatory Decision-Making (data reliability/relevance) and the Considerations for the Use of Real-World Data and Real-World Evidence To Support Regulatory Decision-Making for Drug and Biological Products framework. This guidance is the design-and-conduct layer for the specifically non-interventional case.

When to use

— Apply it whenever a sponsor (or a partner generating evidence for a sponsor) is planning, conducting, or documenting a non-interventional study — a cohort, case-control, or comparative-effectiveness study in claims, EHR, registry, or linked data — that is intended to contribute to a U.S. regulatory submission for a drug or biologic, whether for effectiveness or for safety. The decision rule for which guidance applies: (1) if the question is whether a given data source is fit for regulatory use, the EHR/claims data guidance governs; (2) if the study is interventional or a hybrid/pragmatic trial with a protocol-driven intervention, this guidance does not apply and the RCT-with-RWD or hybrid-design pathway does; (3) if the study is a non-interventional comparison of treatments as used in routine care, this guidance is the operative one. It is not a journal reporting checklist and not an HTA reference case — for EMA/ENCePP submissions or PASS, pair it with the ENCePP Guide and Methodological Standards; for an HTA dossier, pair it with the relevant HTA reference case; for journal publication, pair it with STROBE/RECORD-PE. Early engagement with FDA (e.g., a meeting before the protocol is finalized) is an explicit expectation, not an option, for studies meant to support effectiveness.

What it requires

— The guidance organizes its expectations around making a non-interventional study as transparent and pre-specified as the trial it stands in for. The substantive domains it enforces for real-world data are: (1) Data fitness-for-use — documented reliability (accrual, completeness, provenance, data-quality checks, conformance, plausibility) and relevance (availability of the exposure, outcome, covariates, and population needed to answer the question), with the supporting data guidance invoked by reference; (2) Design transparency and pre-specification — a finalized protocol and statistical analysis plan with date stamps and version control submitted before analysis, ideally before any analytic data access, so that design choices are not data-driven; (3) Study design framed against a target trial — explicit specification of eligibility, treatment strategies, assignment, time zero/index, follow-up, outcome, and the causal contrast, so the observational analysis emulates a hypothetical pragmatic trial; (4) Time-zero alignment — index-date definition that avoids immortal-time bias and prohibits using post-baseline information to define baseline; (5) Estimands and intercurrent events — a clearly stated estimand (target population, treatment condition, outcome, summary measure) and a pre-specified strategy for intercurrent events such as treatment switching, discontinuation, and death; (6) Exposure, outcome, and covariate definitions — validated operational/phenotype algorithms with reported performance (e.g., PPV) where outcomes are derived from coded data; (7) Confounding control — pre-specified covariates and methods (propensity-score or high-dimensional approaches, active-comparator/new-user framing) with diagnostics for balance and positivity; (8) Missing data and attrition — characterization of loss to follow-up and missingness with pre-specified handling; (9) Sensitivity and quantitative bias analysis — robustness checks and bias quantification (negative controls, E-values) to probe residual confounding and design assumptions; and (10) Transparency and reproducibility — versioned code lists, algorithms, and analytic programs that would allow FDA to reproduce the analysis.

When NOT to use — limitations and common misapplications

— This is a regulatory expectations document, not a scoring tool, and several failure modes recur. (1) Treating it as a fit-for-purpose data assessment. The guidance repeatedly defers to the EHR/claims data guidance for reliability and relevance; checking design boxes here does not establish that the underlying data can answer the question. (2) Conflating it with its sibling guidances. Using this guidance to justify a data source (that is the data guidance's job), or invoking it for an interventional hybrid or single-arm externally controlled trial (a different pathway), is a category error — name and route to the correct guidance. (3) Checklist-as-theater. A complete, well-formatted protocol does not make an observational comparison causal; unmeasured confounding, channeling, and an indefensible comparator are not cured by documentation. (4) Assuming method endorsement. The guidance does not bless any particular estimator (PS matching, IPTW, g-methods) — it requires justification and diagnostics, not adoption of a favorite method. (5) Retrofitting pre-specification. Submitting a protocol after the analysis has been run, or after iterating on results, defeats the central safeguard the guidance is built around; the date-stamped, pre-analysis protocol is the point. (6) Using it where a different guideline is required. For an EMA submission rely on the ENCePP standards; for journal reporting use STROBE/RECORD-PE; for an HTA dossier use the payer reference case — this FDA guidance is necessary but not sufficient for those audiences.

How it maps to this catalog

— Each requirement above is implemented by a concept in this repository. Data fitness-for-use → `fit-for-purpose-data-assessment-rwe` and `database-feasibility-attrition-funnel-rwe`, with payer-specific nuances in `medicare-ffs-ma-commercial-claims-differences-rwe` and source mechanics in `claims-analysis`. Target-trial framing → `target-trial-emulation`, with `clone-censor-weight-per-protocol` when the estimand is a sustained per-protocol strategy. Time-zero alignment → `time-zero-index-date-alignment-rwe` and `immortal-time-bias-handling`. Estimands and intercurrent events → `estimands-ate-att-intercurrent-events-rwe` and `estimand-analysis-traceability-rwe`. Phenotype/algorithm validation → `diagnosis-phenotype-algorithm-1ip-2op-time-window-rwe`, `claims-outcome-algorithm-ppv-sensitivity-rwe`, and `ehr-phenotyping-algorithms-rwe`. Confounding control → `active-comparator-new-user`, `high-dimensional-propensity-score-hdps-rwe`, `propensity-score-methods-psm-iptw`, and `dags-backdoor-criterion-drug-studies`. Attrition and missing data → `attrition-and-loss-to-follow-up-rwe` and `missing-data-pattern-table-rwe`. Sensitivity and quantitative bias analysis → `e-value-sensitivity-analysis`, `negative-control-outcomes-rwe`, and `quantitative-bias-analysis-toolkit-rwe`. Question framing → `picots-framework-rwe`; overall regulatory packaging → `regulatory-readiness-rwe` and `study-protocol-or-sap-elements`. Applied note (claims/EHR/registry): in claims, the highest-yield FDA-facing artifacts are a versioned attrition funnel from source population to analytic cohort, a validated outcome algorithm with reported PPV, an active-comparator new-user design with continuous-enrollment requirements over the full lookback (excluding Medicare Advantage person-time where fee-for-service claims are absent), and a pre-specified estimand with an intercurrent-event strategy — all locked in a date-stamped protocol before analytic data access.