Common science and health reporting problems

What to take away

• Replace broad claims with the studied population, exposure or intervention, comparison, outcome, and period.
• Give absolute numbers and baseline risk whenever relative change appears.
• Treat peer review, trial registration, and regulatory approval as different events.
• Investigate outcome switching, selective reporting, and scientific-image anomalies without making unsupported misconduct claims.
• Correct the headline and distributed graphics whenever they preserve an error removed from the body.

Health reporting errors often arise when a limited research result is made shorter, broader, and more certain at every stage. These recurring patterns help editors identify the missing qualification and choose a specific repair.

1. Association becomes cause

An observational study finds that people reporting one behavior also report an outcome more often. The story says the behavior caused the outcome. Confounding, reverse causation, selection, or measurement error may offer other explanations.

Repair: Use association language and explain the strongest plausible alternative. Reserve causal verbs for a design and analysis that support them.

2. Cell or animal results become a human treatment

A molecule changes cells in a dish or an intervention changes a measure in mice. The headline says a cure is near. Human dose, delivery, metabolism, effectiveness, and harm remain unknown.

Repair: Name the model and measured result. State the next research step instead of predicting clinical success.

3. Relative risk hides the baseline

"Risk fell 50 percent" can describe a fall from 2 in 100 to 1 in 100 or from 2 in 100,000 to 1 in 100,000. The same relative figure carries very different practical scale.

The National Institutes of Health guide to understanding health risks explains absolute risk as the chance of an outcome over a period and shows how a large-sounding relative reduction can correspond to a small absolute change. It also advises considering whether the people behind a statistic resemble the person to whom the risk is being applied.

Repair: Give event counts, denominators, baseline risk, absolute change, relative change, time horizon, and uncertainty.

4. A surrogate becomes a patient benefit

A study changes a biomarker, scan, test score, or intermediate endpoint. The story says patients live longer, avoid hospitalization, or feel better even though those outcomes were not measured.

Repair: Name the endpoint. Explain whether it is validated to predict the claimed patient outcome and how much uncertainty remains.

5. Statistical significance becomes clinical importance

A tiny difference clears a statistical threshold in a large sample. The story calls the effect meaningful without comparing it with measurement error, treatment burden, cost, harms, or a clinically relevant threshold.

Repair: Report the effect size and interval. Ask what magnitude would matter to patients or practice.

6. No statistical significance becomes no effect

A small or noisy study cannot distinguish its estimate from zero under the selected test. The story says the intervention has no effect. The data may also be compatible with benefit or harm.

Repair: Describe the estimate, interval, sample size, and what effects the study could reasonably detect.

7. The sample becomes everyone

Participants from one clinic, age band, ancestry group, or high-risk population are generalized to all patients. Eligibility exclusions can remove people with common conditions or medications.

Repair: Put the population in the first result sentence. Identify exclusions that materially limit use elsewhere.

8. A subgroup becomes the main finding

The primary result is uncertain, but one of many subgroups has a favorable result. The story presents that subgroup as proof without checking whether it was prespecified, adequately powered, or supported by an interaction test.

Repair: Label exploratory results, report the number of comparisons, and require replication before practice claims.

9. Changed outcomes are invisible

The registration named one primary outcome, while the paper emphasizes another measure or time point. The change may have a sound explanation, but readers cannot assess it when the story consults only the paper.

Repair: Compare the earliest plan, amendments, results record, and publication. Ask for the reason and timing of any material change.

10. "Peer reviewed" becomes "proven"

Peer review can identify problems and improve a paper. It does not reproduce every analysis, guarantee the data, erase uncertainty, or settle disagreement. Published papers can be corrected or retracted.

Repair: Treat review status as one fact about the publication process. Check methods, wider evidence, data access, conflicts, and current journal notices.

11. Trial registration becomes approval

A product appears in a registered clinical study. Publicity calls it approved, validated, government tested, or endorsed. A registry documents study information; it does not issue a product approval decision.

Repair: Check the exact regulator record, product category, indication, dose, manufacturer, jurisdiction, and date.

12. "FDA approved" loses its object

A facility may be registered, an ingredient may be present in another approved product, or a device may have a different regulatory status. The article applies "approved" to the advertised product and use without evidence.

Repair: Quote the regulator's exact decision and intended use. Avoid extending a decision to a different formulation, use, or claim.

13. Funding is either ignored or treated as disproof

Sponsor control can affect design, analysis, reporting, and publication. Yet a financial tie alone does not establish that a result is false.

Repair: Report who funded, designed, analyzed, held the data, and controlled publication. Evaluate methods and replication separately.

14. A scientific-image anomaly becomes a misconduct verdict

Repeated features, sharp boundaries, unusual contrast, or inconsistent labels can warrant questions. A compressed figure, assembly error, or legitimate processing can also create appearances that need original data and context.

The U.S. Office of Research Integrity's guidance on scientific image samples says image discrepancies can support further examination, but authentication requires original data and an anomaly alone does not establish intent or a finding of misconduct. This distinction protects both the research record and the fairness of reporting.

Repair: Preserve the figure and version, request raw files and processing history, contact authors and journal, seek qualified review, and describe only what the evidence establishes.

15. A press release supplies the strongest verb

The paper says "associated," while the release says "prevents" and the headline says "protects." Repetition can make the widened claim look sourced even though every version traces to the same institution.

Repair: Map each verb to the paper's design and results. Quote promotional language only when reporting on the promotion itself.

16. The body is fixed but the headline survives

An editor adds a limitation deep in the article while leaving a causal headline, distorted chart, and definitive social card unchanged. Most readers still receive the unsupported version.

Repair: Apply corrections and qualifications to every surface. The shortest format needs the clearest boundary, not the broadest claim.

Rapid repair table

Problem phrase	Safer reporting move
"causes" from an observational study	State association and alternatives
"cuts risk 60%"	Add baseline and absolute change
"safe" after a short trial	State observed harms and follow-up
"works in humans" after animal research	Name the model and result
"FDA approved study"	Separate registry and product status
"proof of fraud" from a figure anomaly	Report the discrepancy and investigation status

Common questions

Is cautious language always more accurate?

No. Vague hedging can obscure a well-supported result. Use the strongest precise language the design and evidence justify.

Must every story discuss every study limitation?

No. Include limits that could change the reader's interpretation of the central claim, then point to detailed methods where useful.

Is industry-funded research unusable?

No. Funding and control are bias risks to disclose and examine. Study quality, data access, consistency, and independent replication still require direct assessment.

Should a newsroom wait for a journal investigation to mention an image issue?

Not always. It can report a documented discrepancy and responses when newsworthy, but it must not convert an open question into a misconduct finding.