Plain-English explanations, worked examples and the questions worth asking when a research term appears in a headline.
24 terms · Updated 27 Sep 2026
Written and source-checked by AI. No human expert review. Sources & access notes
24 of 24 terms
All scenarios below are invented for education. They do not describe a trial or estimate anyone's treatment benefit. Read the linked guide for a fuller explanation.
The chance of a specified event in a defined group over a stated period.
Invented example: 12 of 400 people have an event during one year. The observed one-year risk is 3%.
Keep in mind Always ask: which event, which people and how long?
The comparison-group risk minus the intervention-group risk for the same outcome and period. A positive value indicates fewer events with the intervention.
Invented one-year risks of 6% and 4% differ by two percentage points: 20 fewer events per 1,000 people.
Keep in mind Percentage points and percentage reductions use different denominators.
Keeping the next treatment assignment unknown until a participant has been enrolled and assigned.
Imagine a recruiter who can see the next assignment and delays enrolling someone. Concealment is intended to prevent that opportunity.
Keep in mind This protects assignment; blinding concerns knowledge after assignment.
A systematic distortion of a result caused by how research is designed, conducted, analysed or reported.
Imagine a symptom survey that loses most participants whose symptoms worsened. The remaining responses may present a misleading picture.
Keep in mind A large sample does not automatically remove bias.
An objectively measured biological characteristic used to indicate a process, disease state or response to an intervention.
In a fictional experiment, researchers measure a protein concentration before and after an intervention. That is a measurement, not yet a claim about how people feel.
Keep in mind A useful biomarker is not automatically a validated surrogate for a clinical benefit.
Keeping specified people unaware of assigned treatment to limit effects of that knowledge on care or outcome measurement.
Imagine assessors scoring anonymised scans without knowing the treatment group. This does not tell us whether participants or treating staff were also blinded.
Keep in mind Look for who was blinded and how.
Random assignment of groups, such as clinics or villages, rather than separate assignment of every participant.
Imagine 30 clinics assigned to two appointment systems. People within a clinic share that assignment and may share other influences.
Keep in mind Ask whether the analysis accounts for grouping; people within a cluster are not automatically independent observations.
An endpoint combining several specified types of outcome into one measure. Its counting rule must be stated.
A fictional trial measures time to the first of two events: an urgent visit or a hospital admission. Fewer first events could mainly reflect fewer urgent visits.
Keep in mind Inspect each component. A result for the composite does not establish the same effect on every component.
An interval expressing statistical uncertainty around an estimate under the analysis assumptions. A valid 95% procedure captures the true parameter in 95% of hypothetical repetitions.
Invented estimate: a two-point improvement, with an interval from a one-point worsening to a five-point improvement. The estimate alone hides that uncertainty.
Keep in mind It does not mean 95% of patients improve, and it does not account for every source of bias.
Mixing the effect of an exposure or treatment with the influence of other causes of the outcome that differ between comparison groups.
Imagine people in one exercise programme are younger and healthier at entry. Better later outcomes could partly reflect those starting differences.
Keep in mind Adjustment only helps to the extent that relevant factors are identified, measured and modelled appropriately.
A comparison of event hazards: the instantaneous event rates among people still at risk. A single reported ratio summarises the comparison using a particular time-to-event model.
Invented report: HR 0.80. Before calling that “20% fewer people affected,” look for the event proportions, follow-up and model assumptions.
Keep in mind A hazard ratio is not a risk ratio or an absolute probability.
Differences between studies. These may involve participants, methods or interventions; statistical heterogeneity concerns variation in effect estimates beyond sampling error.
Imagine three studies of the same exercise programme delivered at different intensities. Their shared label does not establish an identical intervention.
Keep in mind A heterogeneity statistic does not explain why studies differ.
Analysis by original random assignment, regardless of later treatment use, aiming to include every randomised participant.
Imagine someone switches groups after assignment. Under this principle their outcome stays with the originally assigned group.
Keep in mind The label does not resolve missing outcomes; check exclusions.
Statistical combination of results from multiple studies to estimate an effect. The studies and effect measures must support a meaningful comparison.
Imagine combining four independent trials of the same question. A follow-up paper from one trial should not quietly become a fifth independent trial.
Keep in mind Pooling increases neither relevance nor reliability automatically.
For a specified comparison, outcome and time, the reciprocal of an absolute risk reduction. It expresses how many people would need treatment for one additional beneficial outcome, on average.
Invented complete one-year follow-up: a risk falls from 6% to 4%. The reduction is 0.02, so NNT is 1/0.02 = 50 over one year.
Keep in mind Keep the time frame attached; the number does not identify who benefits.
A study observing exposures, care or outcomes without investigators assigning the exposure or treatment under investigation.
Imagine comparing existing clinic records for people who chose different exercise programmes. The comparison did not begin with random allocation.
Keep in mind An association needs careful assessment of alternative explanations before a causal claim.
The odds of an event in one group divided by those in another. Odds compare events with non-events, rather than with all people.
Invented risks of 20% and 10% give odds of 20/80 and 10/90. The odds ratio is 2.25, while the risk ratio is 2.
Keep in mind Calling an odds ratio a risk ratio can overstate a relative risk comparison when events are common.
Under a specified model, including its null hypothesis, the probability of a test statistic at least as extreme as the one observed.
Invented report: p = 0.03. It still needs an effect estimate, uncertainty interval, outcome definition and account of how the analysis was selected.
Keep in mind It is not the probability that the hypothesis is true, or that chance alone caused the result.
Total time contributed by people under observation. Event counts divided by this time produce a rate.
Invented example: 80 people each observed for three years contribute 240 person-years. That is not the same study design as observing 240 people for one year.
Keep in mind A rate per person-year is not directly a percentage of people affected.
A study using chance to assign participants to intervention and comparison groups, then comparing outcomes.
Imagine a computer assigning 600 volunteers to two appointment systems. Random assignment concerns the comparison, not whether volunteers represent every patient.
Keep in mind Randomisation helps causal inference; it does not guarantee perfect conduct, equal groups or universal applicability.
The event risk in one group divided by the event risk in the comparison group, using the same outcome and period.
Invented one-year risks of 4% and 6% give a risk ratio of about 0.67: roughly a one-third relative reduction, but a two-percentage-point absolute difference.
Keep in mind The ratio alone does not reveal the starting risk.
A label often applied when a statistical test passes a chosen threshold, commonly p < 0.05. It is not a measure of practical importance.
Imagine a very large study finding a tiny change in a symptom score. A small p-value cannot tell you whether that change is noticeable or worthwhile.
Keep in mind A result above the threshold also does not establish no effect.
A substitute endpoint used in place of a direct measure of how people feel, function or survive. Its ability to predict benefit needs evidence in the relevant setting.
In a fictional trial, a laboratory measure improves. That alone does not demonstrate fewer symptoms, hospital admissions or deaths.
Keep in mind Biological plausibility and proven prediction of clinical benefit are different claims.
A structured effort to answer a defined question by using explicit methods to find, select, appraise and synthesise relevant studies.
Imagine a review specifying eligible trials before searching, then explaining which were excluded and why. A collection of favourite papers does not perform the same task.
Keep in mind Read the search date, methods and limitations. A systematic review may contain no meta-analysis.
Try a shorter word, an abbreviation such as NNT, or another subject.
Show all terms →These are original short explanations checked by AI against the listed institutional and methods references. Examples and reading prompts were written for this glossary. Sources support the concepts; they do not describe the invented scenarios. This is an educational reference, not a clinical review or a complete statistics textbook.
Relevant definitions and sections on risk differences, odds and rates; selected web sections consulted, not a full handbook review · Accessed 27 Sep 2026
Sections on confidence intervals and number needed to treat; selected web sections consulted, not a full handbook review · Accessed 27 Sep 2026
Sections on allocation, blinding, intention-to-treat and bias domains; selected web sections consulted, not a full handbook review · Accessed 27 Sep 2026
Sections on meta-analysis and variation between studies; selected web sections consulted, not a full handbook review · Accessed 27 Sep 2026
Overview and purpose of systematic reviews; selected web sections consulted, not a full handbook review · Accessed 27 Sep 2026
Explanation of confounding and non-randomized comparisons; selected web sections consulted, not a full handbook review · Accessed 27 Sep 2026
Official indexed web text; direct retrieval was unavailable during this check · Accessed 27 Sep 2026
Official course overview; video lectures not reviewed · Accessed 27 Sep 2026
Official dictionary definition · Accessed 27 Sep 2026
Official indexed dictionary definition · Accessed 27 Sep 2026
Official dictionary definition; paired with Cochrane time-to-event methods · Accessed 27 Sep 2026
Official web sections defining biomarkers and surrogate endpoints · Accessed 27 Sep 2026
Official podcast transcript, particularly the composite-endpoint discussion; audio not reviewed · Accessed 27 Sep 2026
Official 2016 announcement PDF, including the six principles on page 2 · Accessed 27 Sep 2026