STaRT-RWE

What it is

— STaRT-RWE (Structured Template for Planning and Reporting on the Implementation of Real World Evidence Studies) is a reporting-and-planning template, published in BMJ in 2021 by Wang, Schneeweiss, and an international group of pharmacoepidemiologists, methodologists, regulators (FDA), and HTA/industry scientists. Its defining feature is tabular pre-specification: rather than prose, it forces every implementation decision into structured tables and figures — a design diagram with explicit time anchors (cohort entry, exposure assessment, covariate look-back, washout, follow-up, outcome ascertainment), code-list and algorithm tables, and an analysis specification — so that an independent team could reproduce the study from the document alone. It is a companion to, not a competitor of, the prose-protocol templates: it pairs with the ISPE/ISPOR HARPER harmonized protocol template (which carries the same implementation logic in narrative form) and feeds the downstream reporting checklists (RECORD-PE, STROBE, ENCePP). STaRT-RWE is maintained as a community good-practice instrument rather than by a single standards body; it is referenced in FDA real-world-evidence and EMA/ENCePP transparency expectations.

When to use

— Use STaRT-RWE for hypothesis-evaluating, treatment-effect non-interventional studies in routinely collected data: active-comparator new-user cohorts, comparative effectiveness/safety studies, target-trial emulations, and post-authorization safety/effectiveness studies (PASS) built on claims, EHR, registry, or linked data. It is appropriate at three moments — the protocol/planning stage (lock the design before data access), an amendment/audit stage (document deviations against the locked specification), and the reporting stage (publish the completed tables as a manuscript appendix, an HTA dossier annex, or a regulatory submission artifact). Decision rules for choosing it over siblings: use STaRT-RWE (or HARPER) — not PRISMA-P — when the object is a primary RWE study rather than a systematic review of studies; use STaRT-RWE's structured tables alongside HARPER's narrative protocol (they are designed to interlock, not substitute); and use STaRT-RWE for the implementation specification, then report the finished study against RECORD-PE/STROBE (item-level reporting checklists) and, for EU PASS, the ENCePP checklist. For descriptive/disease-natural-history or hypothesis-generating studies, STaRT-RWE's treatment-effect scaffolding is heavier than needed.

What it requires

— STaRT-RWE enforces the implementation domains where unforced errors actually occur in RWE: (1) a design figure with explicit time-zero and all assessment windows, which surfaces immortal-time and look-back/look-forward errors before they happen; (2) data-source fitness-for-use — provenance, capture, lags, linkage, and the rationale that the source can measure exposure, outcome, and confounders well enough for the question; (3) exposure, outcome, and covariate operational definitions as code-list/algorithm tables with windows, settings (inpatient/outpatient counts, e.g. 1-IP/2-OP rules), grace periods, and — for outcomes — phenotype/algorithm validation (PPV/sensitivity and the validation source); (4) eligibility and cohort construction with an attrition table from source population to analytic cohort; (5) the estimand — target population, treatment strategies, and handling of intercurrent events (switching, discontinuation, death) under an ITT-like or per-protocol contrast; (6) confounding control — the covariate set, the adjustment method (propensity-score or high-dimensional PS, matching/weighting), and balance diagnostics; (7) missing data and loss-to-follow-up/attrition handling; and (8) a pre-specified sensitivity and quantitative-bias analysis plan (alternative windows, negative controls, E-value). It also requires version-controlled code lists and parameters so the specification is auditable.

When NOT to use — limitations and common misapplications

— STaRT-RWE is a transparency template, not a validity guarantee, a risk-of-bias instrument, or a quality score. Concrete failure modes: (1) Template-as-theater — filling every cell while the design is biased; a perfectly tabulated immortal-time error is still an immortal-time error. Completing STaRT-RWE does not make an observational estimate causal or unconfounded; it makes the design legible so reviewers can judge it. (2) Confusing it with a critical-appraisal tool — STaRT-RWE describes what was done; bias is graded with ROBINS-I, and reporting completeness with RECORD-PE/STROBE. Do not cite a completed STaRT-RWE table as evidence of low risk of bias. (3) Wrong instrument for the object — using it for a systematic review (that is PRISMA-P/PRISMA 2020) or for an RCT protocol (SPIRIT). (4) Wrong scope — applying its treatment-effect machinery to a purely descriptive or hypothesis-generating study, where it adds friction without protecting against the relevant errors. (5) Specification drift — locking a template and then silently deviating; the value is in pre-specification plus documented amendments, not the blank form. (6) Stopping at STaRT-RWE for reporting — it specifies implementation; journals and regulators still expect the item-level RECORD-PE/STROBE reporting checklist and, for EU PASS, ENCePP.

How it maps to this catalog

— Each STaRT-RWE requirement is implemented by a concept in this repo: - Time-zero and design figure → target-trial-emulation (specify the hypothetical trial and align follow-up at a defensible time zero) and active-comparator-new-user (the new-user washout + active-comparator structure that fixes time zero and curbs confounding by indication). - Exposure/eligibility construction in routine data → active-comparator-new-user and claims-analysis (NDC/fill-date/days-supply exposure, enrollment requirements, MA-vs-FFS capture caveats). - Outcome/covariate definitions and phenotype validation → diagnosis-phenotype-algorithm-1ip-2op-time-window-rwe (1-IP/2-OP rules, time windows, position/setting, PPV). - Estimand and intercurrent events → estimands-ate-att-intercurrent-events-rwe (ATE/ATT, ITT vs per-protocol, switching/discontinuation/death handling). - Confounding control → high-dimensional-propensity-score-hdps-rwe (proxy selection and PS adjustment when key confounders are unmeasured). - Attrition and loss to follow-up → attrition-and-loss-to-follow-up-rwe (CONSORT-style flow, informative censoring). Read STaRT-RWE as the specification layer that ties these concepts together; each catalog concept supplies the operational depth a single STaRT-RWE cell only summarizes.

Applied note (claims/EHR/registry RWE)

For a claims-based active-comparator new-user safety study, the STaRT-RWE design table makes the high-leverage choices explicit and checkable in one view: the continuous enrollment + drug-free washout that establishes incident-user status, time zero set at the first qualifying fill (not at diagnosis — the classic immortal-time trap), covariates measured only in the pre-index window feeding a high-dimensional PS, an attrition table from source population to matched cohort, the estimand with its switching/discontinuation rule, and a sensitivity row (washout length, grace period, negative-control outcome, E-value). In EHR or registry data the same template forces declaration of order-vs-administration exposure capture, linkage to fills, and explicit observation windows so that visit-driven, potentially informative loss to follow-up is handled rather than ignored.