STROBE is the 22-item checklist journals expect observational studies to follow — cohort, case-control, and cross-sectional. This walks the items that reviewers check first, grouped by what they protect, with a self-check for each.
STROBE — the Strengthening the Reporting of Observational Studies in Epidemiology statement — is a 22-item reporting checklist ↗ maintained through the EQUATOR Network ↗. It does not tell you how to design a study; it tells you what a reader needs in order to judge one. Many journals now require a completed STROBE checklist at submission and check the manuscript against it, so a missing item is a concrete, citable reason to return the paper before review. For studies built from routinely-collected health data, the RECORD extension adds items specific to that data source.
The items below are grouped by the question they answer. This is a reader's tour, not a substitute for the official checklist — download the current version from the STROBE site and complete it line by line.
Say the design out loud, in the title or abstract (item 1). “A retrospective cohort study of…” A reviewer wants to know what they are reading before the first result.
State objectives and any pre-specified hypotheses (item 3). An analysis framed as confirmatory should have been specified in advance; one that was not is exploratory, and should say so.
Give the setting, eligibility, and how participants were selected (items 5–6). Dates, locations, inclusion and exclusion criteria, and — for case-control and matched designs — how controls or matches were chosen.
Self-check. Can a reader reconstruct exactly who entered your study and who could not? If the eligibility rule lives only in your head, it is a missing item.
Define every variable, including confounders and effect modifiers (item 7). The exposure, the outcome, and the covariates you adjusted for — each with its definition and, where relevant, its diagnostic criteria.
State data sources and measurement methods (item 8), and note any difference in how groups were assessed — a classic source of information bias.
Describe efforts to address bias (item 9) and justify the study size (item 10). Even a retrospective study should say why the available sample is adequate to the question, not merely convenient.
Self-check. For a cohort study, the sharpest question is which confounders you could not measure. If diagnosis, severity, or comorbidity were unavailable, the reader must be told — an unadjusted association is not a causal one. This is exactly the gap flagged in our colectomy cohort case study.
Item 12 is where most observational studies are won or lost, and it has sub-parts worth reading as separate obligations: describe the methods including confounding control; explain how subgroups and interactions were handled; state how missing data were addressed; describe any sensitivity analyses. Silence on missing data — a cohort that starts at 400 and ends at 280 with no account of the 120 — is one of the most common reasons a reviewer distrusts a result.
Self-check. Write one sentence each for: your confounding strategy, your missing-data handling, and your treatment of multiple comparisons. If any sentence is “we did not address this,” that is a finding you want to surface yourself, not have a reviewer surface for you. See responding to Reviewer 2's statistics comments for how to phrase those concessions.
Account for every participant (item 13). Numbers at each stage — eligible, included, analyzed, lost — ideally as a flow diagram. Report adjusted estimates with confidence intervals (item 16), not bare p-values, and make clear which confounders each model adjusts for.
Key results, limitations, interpretation, generalizability (items 18–21). The limitation an observational study cannot skip is the direction of potential bias: which way would residual confounding push your estimate? A conclusion that reaches past association into cause is the single most common over-claim in this design.
Self-check. Put your abstract's conclusion beside your primary estimate and its interval. If the interval includes effects you would not call equivalent, do not summarize it as “no difference.” An imprecise null is a question, not an answer.
Walk the official checklist line by line before submission, not after revision — it is published in advance, which makes a missing item the easiest rejection reason to prevent. A structured pre-submission review does this alongside the statistics: RigorMD maps an observational manuscript to STROBE (and RECORD, where the data are routinely-collected), item by item, as one of six scored domains, and pairs it with a deterministic forensic layer that recomputes the reported numbers and two independent engine appraisals of whether the conclusions are calibrated to the evidence. It flags; it does not certify, and it does not replace a statistician's input on design.
See a full sample report on a STROBE/RECORD cohort →, read how the engine works, or review pricing — the pre-submission review is $25. For other designs, see the CONSORT, TRIPOD, and PRISMA walkthroughs.