The STROBE statement

Using the wrong checklist is a visible mistake, and journals do notice. Match it to the design rather than to the subject area.

Service evaluation and audit sit outside all of these, because they are not research. If you are not certain which category your project falls into, that question has to be settled before the reporting standard matters at all.

The checklist follows the shape of a paper, which is why it works well as a drafting tool and badly as a final tick-box exercise.

Several items have lettered sub-parts. Item 12 runs from 12a to 12e on its own, which is why the checklist is 22 items but around 30 separate things to answer.

Most items are satisfied by writing down what you did. Four are different, and they are the ones a statistical reviewer reads first.

Item 9, bias. Not a generic limitations paragraph. It asks what you did about specific, named sources of bias. For most observational work the honest answer involves self-selection, who was lost along the way, and how exposure was measured. Naming the bias and saying nothing was done about it is a legitimate answer. Leaving it blank is not.

Item 10, study size. Authors write "all available records were used", which is a description rather than a justification. If the sample was determined by availability, say so and discuss what that means for precision. If a calculation was done, give it. Reviewers notice the difference.

Item 12c, missing data. Say what was missing, how much, and what you assumed about why. Complete-case analysis is a choice with consequences, not a neutral default, and "there was no missing data" in a real-world dataset invites scepticism.

Item 16a, estimates. Unadjusted and adjusted estimates, both, with confidence intervals, plus which confounders were adjusted for and why they were chosen. The last part is where papers are thinnest, because the choice is often inherited rather than reasoned.

Most items are satisfied by describing what you did. These four are where papers are thin, and they are the first ones a statistical reviewer reads. The difference between a weak and a strong answer is rarely more than three sentences.

Item 9, efforts to address bias

Weak: "Potential sources of bias are acknowledged in the limitations."

Strong: "Three sources of bias were anticipated. Participants self-selected into the cohort, so baseline characteristics were compared with national registry figures and are reported in Table 1. Exposure was self-reported, so a validation sub-study of 120 participants compared responses with pharmacy records. Loss to follow-up was 18%, and those lost were compared with completers on all baseline variables."

The strong version names each bias and says what was done. "Nothing could be done about this one" is an acceptable answer; silence is not.

Item 10, how the study size was arrived at

Weak: "All available records were included."

Strong: "All 1,840 eligible records in the study period were included; no sample size calculation was performed as the analysis was exploratory. With 212 events this gives approximately 80% power to detect a hazard ratio of 1.4, and the study is underpowered for smaller effects."

The weak version describes what happened. The strong version says what that means for what the study can detect, which is the actual question.

Item 12c, how missing data were addressed

Weak: "Cases with missing data were excluded."

Strong: "Outcome data were complete. Among covariates, body mass index was missing for 14% and smoking status for 9%. Missingness was associated with age and site, so multiple imputation by chained equations was used with 30 imputations, including the outcome in the imputation model. Complete-case analysis is reported as a sensitivity analysis in Appendix 2 and gave materially similar estimates."

Per variable, with the assumed mechanism and a sensitivity analysis. See the guide to missing data for why the choice matters more than the amount.

Item 16a, estimates and confounders

Weak: "Results were adjusted for confounders."

Strong: "Unadjusted and adjusted estimates are both reported in Table 3. Confounders were selected a priori from a directed acyclic graph and comprised age, sex, deprivation quintile and baseline severity. Variables on the causal pathway were deliberately excluded."

The last sentence matters. Adjusting for a mediator removes part of the effect you are trying to measure, and a reviewer who sees a plausible mediator in the adjustment list will ask.

Item 13 asks for numbers at each stage with reasons. Observational studies often skip the diagram on the grounds that it is a trial convention. Reviewers increasingly disagree, because the same question applies: who was eligible, who ended up in the analysis, and what happened to the difference.

A flow that reconciles

records in the source database: 12,480
excluded, outside the study period: 3,160 → 9,320
excluded, under 18: 540 → 8,780
excluded, prevalent cases at baseline: 1,210 → 7,570
excluded, no follow-up recorded: 430 → 7,140
included in the analysis: 7,140

Each exclusion is a separate line with its own count and reason, and the running total is shown. A single sentence saying 7,140 of 12,480 were analysed tells a reader nothing about whether the exclusions were reasonable.

The exclusion that deserves a sentence of its own

Removing 1,210 prevalent cases is correct for a study of incidence, since someone who already has the condition cannot develop it. Removing 430 with no recorded follow-up is a different matter: that is a potential source of selection bias, because people with no follow-up may differ systematically from those with some. Compare them on baseline variables and report it.

Filled in after the paper is drafted, the checklist records what you happened to write. Used at the protocol stage it changes what you collect.

Item 10 asks how the study size was arrived at, which is a question worth confronting before collection rather than after. Item 12c asks how missing data were handled, which is easier when you planned for missingness. Item 9 asks what you did about bias, and almost everything that can be done about bias has to be done before data exists. A checklist completed at the end can only describe; completed at the start it can prevent.

This matters because completed checklists are sometimes presented as evidence of quality, including by authors who should know better.

The value is real but narrow: it makes a study transparent enough to be judged. Someone still has to do the judging, which is what an independent statistical review is for.

Several extensions adapt STROBE for particular fields, including STROBE-ME for molecular epidemiology, STROBE-NI for neonatal infection, RECORD for studies using routinely collected health data, and STREGA for genetic association studies.

RECORD is the one most often relevant in UK applied health research, since it covers work built on routine data such as primary care records or hospital episode statistics. It adds items on how the database was used, how the study population was identified from codes, and what the data was originally collected for, which is exactly where routine-data studies tend to be vague.

A large share of UK applied health research runs on data collected for something other than research: primary care records, hospital episode statistics, prescribing data. STROBE was not written with that in mind, and RECORD is the extension that fills the gap.

The code list point is the one most often skipped and the one that most affects whether anyone can replicate the study. A paper that defines its outcome as "diagnosed heart failure" without saying which codes that covers has not described its outcome.