STRATOS (STRengthening Analytical Thinking for Observational Studies)

What it is

— STRATOS (STRengthening Analytical Thinking for Observational Studies) is an international collaborative initiative that produces methodological guidance documents — not a single unified checklist — for the statistical analysis of observational studies. It was founded by Willi Sauerbrei, Patrick Royston, and colleagues and formally introduced in a 2014 Statistics in Medicine paper. STRATOS is organised into nine topic groups (TGs), each producing technical guidance on a different analytic dimension: TG1 — Selection of variables and functional forms; TG2 — Missing data; TG3 — Measurement error and misclassification; TG4 — Study design (matching); TG5 — Causal inference, estimands, and target parameters; TG6 — High-dimensional data; TG7 — Survival analysis; TG8 — Prediction modelling (overlaps with TRIPOD/PROBAST); and TG9 — Meta-analysis of individual participant data. Each TG publishes tutorials, primers, and peer-reviewed papers in statistics and biomedical journals. Unlike STROBE (which tells you what to report about your analysis) or GRADE (which evaluates evidence certainty), STRATOS tells you how to do the analysis correctly — it fills the gap between the reporting checklist and the textbook. The initiative is maintained at stratos-initiative.org and publications span BIOM, Statistics in Medicine, American Journal of Epidemiology, and BMJ.

When to use

— Reference STRATOS when you need authoritative, peer-reviewed guidance on how to conduct a specific analysis in an observational study, beyond what STROBE or RECORD require you to report. Concrete use cases: (1) Variable selection and functional forms (TG1) — when building a multivariable model and choosing between a priori variable sets, stepwise selection, or penalised regression; or when deciding whether to model continuous variables linearly or with splines/fractional polynomials. (2) Missing data (TG2) — when designing an imputation strategy (multiple imputation vs. single imputation vs. complete-case analysis) and checking whether the missing-at-random assumption holds. (3) Measurement error (TG3) — when exposure or covariate measurement error in claims, EHR, or survey data is likely to bias effect estimates. (4) Causal inference (TG5) — when translating a PICO into a causal diagram (DAG), selecting an estimand, or using propensity scores, g-estimation, or instrumental variables. (5) Survival analysis (TG7) — when choosing between Cox regression, accelerated failure time models, or competing-risk models and reporting assumptions. (6) High-dimensional data (TG6) — when hdPS, LASSO, or machine-learning variable selection is used in a claims or EHR study. Decision rule: STRATOS is a methods guidance resource, not a reporting checklist; use it alongside (not instead of) STROBE, RECORD-PE, HARPER, or TRIPOD — those checklists require disclosure of what was done, STRATOS explains how to do it well.

What it requires (checklist domains)

— STRATOS does not enforce a single checklist; instead, each topic group issues guidance with recommended analytical practices and common pitfalls. Across TGs the guidance collectively enforces these analytical disciplines: TG1 (Variable selection and functional forms): pre-specification of the analysis strategy before looking at outcome-predictor associations; avoiding data-driven selection of covariates without penalty or regularisation; and modelling continuous predictors with flexible forms (restricted cubic splines or fractional polynomials) rather than arbitrary categorisation. TG2 (Missing data): characterising the missing-data mechanism (MCAR, MAR, MNAR) before choosing a strategy; using multiple imputation under MAR with a correctly specified imputation model (including the outcome); and conducting sensitivity analyses under MNAR assumptions. TG3 (Measurement error): quantifying the likely direction and magnitude of exposure-measurement error; using regression calibration, SIMEX, or probabilistic bias analysis where error is non-differential but substantial; and reporting sensitivity analyses. TG5 (Causal inference): explicit specification of the target estimand (ATE, ATT, or per-protocol); use of directed acyclic graphs (DAGs) to guide covariate selection and identify confounders vs. mediators; and reporting the assumptions required for causal interpretation. TG7 (Survival analysis): checking proportional-hazards assumptions; considering competing risks; and using complementary log-log or Nelson-Aalen estimates rather than Kaplan-Meier when hazards are not proportional. TG8 (Prediction): guidance that parallels TRIPOD/PROBAST on calibration, overfitting, and internal validation.

When NOT to use — limitations and common misapplications

— (1) Confusing STRATOS with a reporting guideline — STRATOS is a methods guidance initiative, not a checklist authors submit with manuscripts; journals require STROBE or RECORD, not a STRATOS sign-off. (2) Using STRATOS in isolation without specifying which TG — citing "STRATOS" without specifying the relevant topic group is vague; be specific (e.g., "we followed STRATOS TG2 guidance on missing data"). (3) Treating STRATOS as a mandatory standard — unlike TRIPOD or PRISMA, STRATOS guidance documents are educational outputs; they do not carry the same journal-submission or regulatory-submission weight as formal reporting checklists. (4) Over-applying TG1 to causal studies — STRATOS TG1 (variable selection) guidance is primarily framed for prediction modelling; causal observational studies should use DAG-guided pre-specified covariate sets per TG5, not algorithmic selection. (5) Missing the TG5 / estimands connection — many pharmacoepidemiologists use STRATOS TG5 informally for DAG reasoning but do not formally cite it; aligning with ICH E9(R1) estimand thinking (ich-e9-r1 in this catalog) provides a complementary and increasingly regulatory- accepted framework. (6) Assuming STRATOS replaces biostatistical peer review — STRATOS guidance documents are written for a wide audience; complex analytic challenges in specific study types still require biostatistical co-authorship and review.

How it maps to this catalog

— STRATOS topic-group guidance directly supports the implementation of several core catalog methods: TG1 (variable selection / functional forms) underpins rigorous propensity-score-methods-psm-iptw construction and the covariate specification step in any multivariable model. TG2 (missing data) implements the missing-data discipline required by multiple-imputation-longitudinal-rwe and contributes to missing-data-pattern-table-rwe (the reporting side). TG3 (measurement error) connects to quantitative-bias-analysis-toolkit-rwe (probabilistic bias analysis accounts for differential and non-differential measurement error) and to claims-outcome-algorithm-ppv-sensitivity-rwe (misclassification of outcomes is a form of measurement error). TG4 (design/matching) informs matching-estimators-rwe and the design choices in active-comparator new-user studies (active-comparator-new-user). TG5 (causal inference) directly operationalises the DAG discipline underlying dag-framework and provides the estimand-selection complement to estimands-ate-att-intercurrent-events-rwe. TG6 (high- dimensional data) provides the methods basis for high-dimensional-propensity-score-hdps-rwe. TG7 (survival analysis) informs time-to-event modelling across catalog concepts including accelerated-failure-time-models, competing-risks-methods, and kaplan-meier-methods. TG8 (prediction) bridges to tripod and probast in this catalog. The reporting obligations that STRATOS guidance helps fulfil are captured by strobe, record-pe, and harper — the companion reporting checklists in the guidelines section of this catalog.