STROBE-MR (STROBE Extension for Mendelian Randomization)

What it is

— STROBE-MR (Strengthening the Reporting of Observational Studies in Epidemiology using Mendelian Randomization) is a reporting checklist that extends the STROBE statement to Mendelian randomization (MR) studies — observational analyses that use germline genetic variants (typically SNPs from GWAS) as instrumental variables to estimate the causal effect of a modifiable exposure on an outcome. The statement (Skrivankova, Richmond, Woolf et al., JAMA 2021) defines the items an MR study must report; a companion explanation-and-elaboration paper (BMJ 2021) gives item-by-item guidance and worked examples. STROBE-MR is hosted and maintained as a STROBE extension within the EQUATOR Network. Its purpose is to make the genetic instruments, the data sources (one-sample vs two-sample, individual-level vs summary-level GWAS), and the assessment of the three core IV assumptions transparent and auditable, so that readers and reviewers can judge whether an MR estimate is credible rather than an artefact of pleiotropy, weak instruments, or population structure.

When to use

— Apply STROBE-MR whenever you are reporting a Mendelian randomization study: a one-sample MR in a single cohort/biobank, a two-sample MR combining exposure and outcome GWAS summary statistics, multivariable MR, MR with summary data from consortia (e.g., a journal manuscript, a triangulation paper supporting a drug-target or biomarker causal claim, or an MR component embedded in a larger evidence package). Decision rule for choosing the right STROBE family member: use plain STROBE for a generic observational cohort/case-control/cross-sectional study; use STREGA when the focus is reporting a genetic association study (genotyping, HWE, population stratification) that is not using genotypes as instruments for causal inference; use RECORD / RECORD-PE when the study is built on routinely collected health data (claims/EHR/registries) — those govern data provenance, not genetic instruments; and use STROBE-MR specifically when germline variants are deployed as instrumental variables to estimate a causal effect. The distinction is the analytic intent: genetics-as-instrument (STROBE-MR) versus genetics-as-association (STREGA) versus routine-data-provenance (RECORD). STROBE-MR is a reporting guideline for journal publication and scientific transparency; it is not itself an FDA/EMA submission template, though MR evidence is increasingly cited in regulatory and HTA causal-triangulation arguments and a STROBE-MR-compliant report is the expected substrate for that use.

What it requires

— Beyond the generic STROBE items (title/abstract, structured introduction with a pre-specified hypothesis, methods, results, discussion, funding), STROBE-MR enforces MR-specific reporting that maps onto the three instrumental-variable assumptions: (1) Relevance — report how genetic instruments were selected, the source GWAS, genome-wide significance and clumping thresholds, instrument strength (F-statistics, variance explained R²), and steps taken against weak-instrument bias. (2) Independence (exchangeability) — report control for population stratification (ancestry, principal components, restriction to a single ancestry), and assortative mating / dynastic (parental-genotype) effects where relevant. (3) Exclusion restriction (no horizontal pleiotropy) — report the biological plausibility of the variant–exposure pathway and the battery of sensitivity analyses that probe the no-pleiotropy assumption: MR-Egger (intercept and slope), weighted median, weighted mode, MR-PRESSO, leave-one-out, and heterogeneity statistics. STROBE-MR additionally requires explicit reporting of the data structure (one-sample vs two-sample; degree of sample overlap between exposure and outcome GWAS, which biases two-sample estimates toward the confounded observational association), the estimand and its scale (per-unit or per-SD change in the genetically-proxied exposure — a lifelong-exposure contrast that is not the same as a clinical-intervention effect), harmonization of effect alleles across data sources, and the software/packages and versions used. In RWE terms these are the field's analogues of data fitness for use, time-zero/estimand specification, confounding control, and quantitative sensitivity / bias analysis — adapted to the genetic-instrument setting.

When NOT to use — limitations and common misapplications

— STROBE-MR is a reporting checklist, not a risk-of-bias instrument, not a quality score, and not a substitute for the IV assumptions themselves. Concrete failure modes: (1) Wrong extension — using plain STROBE (which has no items for instrument strength, pleiotropy, or sample overlap) for an MR study, or using STROBE-MR for a genetic-association study that should use STREGA, or for a routine-data observational study that should use RECORD/RECORD-PE. (2) Checklist completeness mistaken for causal validity — a fully STROBE-MR-compliant paper can still report a biased estimate: transparent reporting of weak instruments, residual pleiotropy, or substantial sample overlap does not fix those problems, it only surfaces them. Completing the checklist does not make the MR estimate causal. (3) Treating the MR estimate as a drug effect — MR estimates a lifelong genetically-proxied exposure contrast; reporting it as if it were the effect of a short-term clinical intervention is a misinterpretation the discussion items exist to prevent. (4) Checklist-as-theater — ticking items while leaving instrument-selection criteria, F-statistics, sample-overlap fraction, or the pleiotropy-sensitivity suite vague defeats the purpose; the value is the substantive disclosure, not the page count. (5) Using it as an appraisal tool for someone else's study — to grade MR evidence, pair the report with a critical-appraisal/ risk-of-bias framework (e.g., ROBINS-style or MR-specific appraisal); STROBE-MR tells you what should have been reported, not whether the study is at low risk of bias.

How it maps to this catalog

— In this repo, STROBE-MR's substantive requirements correspond to these implementing concepts a reader or reviewer can reason against: - The instrument-and-causal engine (the relevance / exclusion-restriction items): instrumental-variables-pharmacoepi-rwe implements the IV logic — instrument strength, the exclusion restriction, and weak-instrument bias — that STROBE-MR's genetic-instrument items demand. - The causal-structure justification (the independence / no-pleiotropy items): dags-backdoor-criterion-drug-studies formalizes the directed-acyclic-graph reasoning that distinguishes a valid instrument from a pleiotropic or confounded one. - The estimand discipline (what the per-SD genetic contrast actually estimates): estimands-ate-att-intercurrent-events-rwe supplies the language for stating the target estimand and why the MR contrast differs from an interventional effect. - The sensitivity / quantitative-bias spine: e-value-sensitivity-analysis, empirical-calibration-negative-controls-rwe, and quantitative-bias-analysis-toolkit-rwe are the catalog homes for the "how robust is this to unmeasured violations?" reporting STROBE-MR requires (the MR field's MR-Egger / leave-one-out analogues live conceptually alongside these). - External validity of the genetic estimate: generalizability-transportability-external-validity-rwe frames whether a single-ancestry MR estimate transports to the target population. - Sibling reporting guidelines (to pick the right one): strobe (the parent), and the genetic- association vs routine-data distinctions captured by the STREGA and RECORD/RECORD-PE entries in this `guidelines` set. These are the lens for the wrong-extension failure mode above, not items of STROBE-MR itself.

Applied note (claims/EHR/registry RWE)

Most STROBE-MR studies run on biobank or consortium GWAS data, but the extension is directly relevant to RWE built on linked biobank–EHR/claims resources (e.g., UK Biobank, All of Us, or biobank-linked administrative cohorts). When MR is run inside such a resource, the report must still satisfy the data-fitness items RWE reviewers expect: which linked source supplied the phenotype (an algorithm-defined outcome from EHR/claims carries the same misclassification and PPV concerns as any RWE outcome — see the catalog's outcome-algorithm and phenotype concepts), how ancestry and relatedness were handled, and what fraction of the exposure and outcome samples overlap. An MR estimate offered to support a drug-target causal claim in an HTA or regulatory triangulation argument should be reported to STROBE-MR and its underlying real-world phenotypes documented to the standard the rest of this catalog enforces — the two are complementary, not interchangeable.