Clinical Research Writing for a Public Health Audience: Population Claims, Study Design and Reporting

Overstatement in public health writing is usually unintentional. It arises when a well-conducted study in a defined sample is reported in population language, for example “the burden of this condition is rising” from a study that counted cases, or “screening should be expanded” from a cross-sectional association. Both statements describe something the paper did not measure. The remedy is not vagueness but the matching of each claim to the design, the reporting of rates alongside counts, and the inclusion of uncertainty intervals in the sentences that state the conclusion.

Implications of a population-level audience

A clinical reader seeks to know whether to take a particular action for an individual patient. A public-health reader seeks to know what a result implies for a group of people, most of whom were not in the study. The second question introduces three requirements that clinical writing often leaves implicit.

The first is a denominator. Any population claim is a claim about a numerator over a defined population at risk, over a defined period. If a paper reports 340 cases, the immediate question of the public-health reader is: out of how many, over what period and in what population. If these cannot be specified, the study is a case series and should be described as such.

The second is generalisability stated as a property of the sampling, not as an aspiration: who could have entered the study, who did, and who is therefore represented by the estimate 19. A hospital-based sample provides information about people who reach that hospital, and writing as though it describes the surrounding population is the most common form of overreach in this literature.

The third is uncertainty carried through to the conclusion. Population estimates are often imprecise, and the interval is frequently omitted from the sentence that states the implication. If a conclusion holds only at the upper bound, the appropriate conclusion is that the data are compatible with a range of implications.

Distinction between counts and rates

This distinction is among the most important to report correctly, and the GBD 2021 nervous-system analysis provides an unusually clear illustration. The GBD 2021 Nervous System Disorders Collaborators, writing in The Lancet Neurology, reported that between 1990 and 2021 global DALY counts for disorders affecting the nervous system rose 18.2%, while age-standardised death rates fell 33.6% and age-standardised DALY rates fell 27.0% 28. Age-standardised prevalence was almost unchanged, at a 1.5% change.

These four figures, covering one period, move in apparently contradictory directions, yet all four are correct. The count rose because the world’s population grew and aged, so that more people lived into the ages at which these conditions occur 28. The age-standardised rates fell because, at any given age, the risk of dying from or being disabled by these conditions decreased. Prevalence changed little. A paper citing only the first figure describes a worsening crisis; a paper citing only the third describes substantial progress; both are selective.

The corresponding writing principle is that a reported trend should be identified as a count, a crude rate or an age-standardised rate, and that any difference in direction should be stated and explained. A reader who finds the count in the abstract and the rate in the supplement is likely to conclude that the more striking figure was selected. The same principle applies to geography: a crude rate difference between two countries with different age structures largely reflects the difference in age structure.

Correspondence between claims and study design

Overreach in a public-health manuscript can almost always be traced to a specific sentence in which the claim exceeds the design. Knowing in advance which claims each design can support allows the discussion to be drafted against that list rather than written freely and reduced afterwards.

A cross-sectional study establishes prevalence and association at one time. It cannot establish incidence, direction of effect, or what would happen if an exposure changed. A case-control study can estimate an odds ratio efficiently for a rare outcome, but it cannot give prevalence or absolute risk unless the sampling fraction is known. A cohort study establishes incidence and temporal order and supports relative and absolute risk, but residual confounding remains possible. A trial supports causal inference about the intervention in the population randomised, which is rarely the general population; external validity therefore becomes the principal question 15. Routine surveillance or registry data supports population trends, with changes in ascertainment and coding over time as the persistent threat.

This does not preclude discussion of implications. It requires the implication sentence to state what would have to be true for the implication to follow. “If the association is causal and the exposure distribution in the wider population resembles ours, a shift of this size would correspond to…” is a legitimate and informative sentence. “Reducing the exposure would prevent…” from a cross-sectional study is not.

STROBE as a drafting framework

Von Elm and colleagues’ STROBE statement, in BMJ, provides a checklist of 22 items for reporting observational studies, of which 18 apply to all three designs it covers — cohort, case-control and cross-sectional — and 4 are design-specific 19. The practical implication is that almost the entire checklist applies to a study of any of the three designs, so the argument that STROBE does not fit a particular design is rarely tenable 12,19.

Used at the end of drafting, STROBE becomes a burden: a checklist to be mapped onto a finished manuscript, which usually reveals items that can no longer be reported. Used at the outset, it is a list of the questions a reader will ask and therefore serves as a drafting outline. The items of greatest importance in a population study concern setting and eligibility, the sources and methods of selection, the handling of confounders, the treatment of missing data, the flow of participants from eligible to analysed, and whether sensitivity analyses were performed 19. These are also precisely the items that cannot be added retrospectively.

Three specific practices are recommended. The participant flow should be reported numerically, from the population approached to the number in each analysis, because a public-health reader reconstructs the denominator from it. The handling of missing data should be stated rather than leaving a complete-case analysis implicit. At least one sensitivity analysis should be reported, because an estimate presented as though it had no alternative specification appears fragile. The Pubrica page on reporting guidelines explains how to select the appropriate guideline when a design is a hybrid.

Scoping of claims at sentence level

The practical task is narrower than it may appear. A manuscript usually contains only four or five sentences in which a population-level claim is made: the final sentence of the abstract, the opening of the discussion, the paragraph on implications, the conclusion and, in some cases, the title. These are the sentences to review.

Three questions should be asked of each. To what population does the sentence refer, and was that population sampled? What would have to be true, beyond what was measured, for the sentence to hold? Does the sentence remain valid if the estimate lies at the lower bound of its interval rather than at the point estimate? The third question most often necessitates revision, and the revision is usually a qualification rather than a retraction.

A conclusion scoped in this way is not weaker in practice. It is more difficult to challenge, it can be quoted without being misrepresented, and it indicates to a policy reader what further evidence would be needed, which is more useful than a recommendation the data cannot support. It also removes one of the structural objections underlying most manuscript rejections, which cluster around the question and the design rather than the prose 6,10.

Required declarations in public health submissions

Population research tends to involve more parties than a single-centre clinical study (multiple sites, routinely collected data, government or NGO funding, and data custodians who are not authors), so the declaration requirements are more demanding. The ICMJE recommendations are the baseline most journals apply.

Authorship rests on four criteria, all of which must be met; in multi-site population work this typically excludes some individuals who contributed substantially and who must therefore be acknowledged instead. Late decisions on authorship cause genuine disputes, so authorship should be settled at the protocol stage 1. Where a trial is involved, registration must be in a registry meeting a minimum 24-item data set, and since 1 July 2018 manuscripts reporting clinical trial results must include a data sharing statement 1. Data sharing is particularly consequential in this field, because datasets often cannot be shared in raw form for reasons of confidentiality; this is an acceptable response, provided that it is stated explicitly rather than the statement being left blank.

Two further declarations are specific to population work. The ethical approval and the legal basis for using routinely collected or linked data should be stated, including whether consent was waived and by whom. The role of a funder with a policy interest should be described in sufficient detail for a reader to judge independence: who designed the analysis, who had access to the data and who could have prevented publication. Wager and colleagues’ Global Publication Survey in BMJ Open examined awareness and enforcement of publication guidelines among publication professionals; its broad lesson is that written policies and actual practice diverge, so the declaration carries weight precisely because it is verifiable 11.

Reporting implications without predicting uptake

The implications paragraph is where public-health papers most often substitute aspiration for evidence. “These findings will inform policy” and “this tool can be implemented in routine practice” are predictions, and the base rate for such predictions is unfavourable. Churruca and colleagues, in Health Expectations, provide an illustration from the measurement literature: of 315 patient-reported outcome measures identified, 270 (85.7%) were used in subsequent research, but only 51 (16.2%) had any online report of clinical or policy use, and 45 had no evidence of subsequent use at all. The field publishes considerably more than it adopts 4.

A stronger implications paragraph does three things. It names the specific decision on which the evidence bears (a screening interval, an eligibility threshold or a resource allocation) rather than referring to policy in general terms. It states what else a decision-maker would require, which is usually cost, acceptability, workforce feasibility and harms, and which the study is very unlikely to have measured. It also identifies the next study rather than the next policy, which is both more accurate and more likely to be acted on by the readers of these sections.

Where grading the strength of the evidence is appropriate, the GRADE framework described by Guyatt and colleagues in BMJ and in the Journal of Clinical Epidemiology provides a vocabulary for stating that an estimate is of low certainty without dismissing it, and for separating the quality of evidence from the strength of a recommendation 14,18. Maintaining that separation addresses most of what the implications paragraph requires.

Reporting of unmeasured factors

Public-health studies rely on data collected for other purposes, which inevitably contain gaps. The question is not whether to acknowledge them but where to place them so that they strengthen rather than undermine the paper.

Three gaps recur. Ascertainment change over time: where a trend is based on routinely collected data, part of the trend may reflect diagnostic practice rather than disease. Authors should state what changed over the period (coding systems, screening programmes, diagnostic criteria) and, where possible, bound the effect with a sensitivity analysis restricted to a period with stable definitions. Unmeasured confounding: authors should name the specific confounder that was not measured and the direction in which it would shift the estimate, rather than writing “residual confounding cannot be excluded”, which conveys no information to the reader 19. Differential coverage: population data are usually least complete for the groups with the highest burden, so an inequality estimate is often conservative to an unquantifiable degree. Stating this is more informative than presenting the estimate as neutral.

None of these is a reason not to publish. Each is a reason to write a limitation specific enough to guide the next study. A weak limitations paragraph lists generic threats; a strong one states what the next design would have to do differently, which is what a reader undertaking follow-up work requires. Where the gap is statistical rather than substantive, it should be raised with the team responsible for biostatistics before the analysis is finalised rather than at the writing stage.

Key points

Application to your study

Frequently asked questions

@misc{a_guide_to_clinical_research_writing_o_2026,
  author       = {{Pubrica Academy}},
  title        = {Clinical Research Writing for a Public Health Audience: Population Claims, Study Design and Reporting},
  howpublished = {Pubrica},
  year         = {2026},
  month        = {oct},
  url          = {https://pubrica.com/academy/medical-writing/a-guide-to-clinical-research-writing-on-public-health-for-clinical-researchers/},
  note         = {Accessed: <span data-cite-acc="iso">2026-10-15</span>}
}

POPULATION-CLAIM AUDIT
Manuscript: ______________________   Design: ______________________

PART A - ONCE PER MANUSCRIPT

A1. DENOMINATOR
    Numerator (what was counted): ______________________
    Population at risk: ______________________
    Period: ______________________
    [ ] All three stated in the methods, in these terms
    [ ] If any is missing, the paper is described as descriptive and the
        conclusion says so

A2. SAMPLING AND WHO IS REPRESENTED
    Sampling frame: ______________________
    Who could have entered but did not: ______________________
    One sentence naming the limit of inference: ____________________
    [ ] The title does not imply a wider population than was sampled

A3. TRENDS AND RATES
    For each trend reported:
      Figure: __________  Type: count / crude rate / age-standardised rate
      Does another type of the same figure point the other way?  Y / N
      [ ] If Y, both are reported and the difference is explained
    [ ] No crude-rate comparison between populations with different age
        structures is presented as a risk difference

A4. STROBE (22 items; 18 apply to all three designs)
    [ ] Setting, locations and relevant dates
    [ ] Eligibility criteria and sources/methods of selection
    [ ] Each variable defined, including diagnostic criteria
    [ ] Confounders named and the handling method stated
    [ ] Missing data: how many, and how handled (not left implicit)
    [ ] Participant flow, numerically, from approached to analysed
    [ ] At least one sensitivity analysis reported
    [ ] Design-specific items for my design addressed: ______________

A5. DECLARATIONS
    [ ] Authorship: all four ICMJE criteria met by every author
    [ ] Substantial contributors who do not qualify are acknowledged
    [ ] Trial registration, with registry and number (24-item data set)
    [ ] Data sharing statement drafted - including the reason if the data
        cannot be shared
    [ ] Ethical approval and legal basis for routine or linked data use
    [ ] Funder's role: who designed, who had data access, who could have
        stopped publication

PART B - RUN ON EACH POPULATION-CLAIM SENTENCE
  Locate them: last line of the abstract; first line of the discussion;
  the implications paragraph; the conclusion; the title.

  Sentence: ____________________________________________________
  B1. Which population does it refer to? ______________________
      Was that population sampled?  Y / N
  B2. What must be true, beyond what was measured, for it to hold?
      ____________________________________________________________
      [ ] Those conditions are stated in the sentence or the one after it
  B3. Does it still hold at the LOWER bound of the interval?  Y / N
      [ ] If N, rewritten as a range of compatible implications
  B4. Does it recommend an action?  Y / N
      If Y, which of these did the study measure?
        magnitude at population scale ____  cost ____  acceptability ____
        feasibility ____  harms ____
      [ ] Unmeasured inputs are named as what a decision-maker still needs
  B5. Rewritten sentence:
      ____________________________________________________________

PART C - LIMITATIONS, MADE SPECIFIC
  Unmeasured confounder I lack: ____________  Likely direction: ________
  What changed in ascertainment over the period: ________________________
  Groups with poorest data coverage: ____________________________________
  What the next study would have to do differently: ____________________