Glossary

Plain-language definitions of the terms used across the site. Effect measures here are computed as described in the methodology.

Prevention & outcomes

Primary prevention: Treating people who do not yet have the disease (e.g. statins in someone with no known cardiovascular disease) to stop a first event. Baseline risk is lower, so the absolute benefit per person is usually smaller and the number-needed-to-treat larger.
Secondary prevention: Treating people who already have the disease or have had an event (e.g. statins after a heart attack) to prevent recurrence. Baseline risk is higher, so the same relative effect yields a larger absolute benefit and a smaller number-needed-to-treat.
Composite outcome (e.g. MACE): A single endpoint combining several events—"major adverse cardiovascular events" might pool cardiovascular death, heart attack, and stroke. Composites increase event counts and power but can be driven by the least severe component.
Surrogate outcome: A stand-in marker (e.g. LDL cholesterol, a lab value) used instead of a clinical event. Useful but not always predictive of how patients actually fare.
Balancing measure: In quality-improvement work, an outcome tracked to make sure an intervention isn't causing unintended harm (e.g. checking that pain scores don't rise when IV opioids are reduced).

Study design

Randomized controlled trial (RCT): Participants are randomly assigned to treatment or control, which balances known and unknown confounders. The strongest design for establishing that a treatment causes a benefit.
Double-blind: Neither participants nor investigators know who received the treatment, reducing bias in how outcomes are reported and assessed.
Intention-to-treat (ITT): Analyzing everyone in the group they were randomized to, regardless of whether they actually took the treatment or dropped out. Preserves randomization and gives a conservative, real-world estimate of effect.
Per-protocol: Analyzing only participants who completed the treatment as assigned. Can exaggerate efficacy because it discards non-adherent patients who may differ systematically.
Observational study (cohort, case-control): Compares groups that were not randomized—following exposed vs unexposed people (cohort) or working backward from those with vs without an outcome (case-control). Subject to confounding; reported effects are usually statistically "adjusted."
Pre-post / quality improvement (QI): Measures an outcome before and after an intervention in the same setting, without a randomized control group. Useful for implementation work but vulnerable to secular trends, so it cannot firmly establish causation.
Meta-analysis: Statistically combining results from multiple studies of the same question to produce a single pooled estimate with greater precision.

Effect measures

Absolute risk reduction (ARR): The plain difference in event rates: control risk minus treatment risk. If 4% of controls and 3% of treated patients have an event, ARR = 1 percentage point.
Relative risk / risk ratio (RR): Treatment risk divided by control risk. RR = 0.75 means a 25% lower risk with treatment; RR = 1 means no difference; RR > 1 means higher risk.
Relative risk reduction (RRR): The percent change in risk, (1 − RR) × 100. A relative reduction can sound large even when the absolute benefit is small, which is why we show both.
Odds ratio (OR): The ratio of the odds of an event between groups. For uncommon outcomes it approximates the risk ratio; for common outcomes it overstates the effect.
Hazard ratio (HR): A risk ratio that accounts for when events happen over follow-up—the relative rate of an event at any given moment.
Rate ratio: The ratio of event rates per unit of exposure time (e.g. doses per patient-day after vs before an intervention). Used for exposure/count outcomes.
Mean difference (MD) / standardized mean difference (SMD): For outcomes measured as a quantity rather than an event (days of diarrhea, a symptom score, a lab value): the mean difference is the plain difference in averages between groups, in the original units; the standardized mean difference (Cohen's d / Hedges' g) divides by the spread so results measured on different scales can be combined. The no-effect value is 0, and an SMD of roughly 0.2/0.5/0.8 is small/moderate/large.
Number needed to treat (NNT): How many people must be treated to prevent one additional bad outcome, = 1 ÷ ARR. Lower is better. A large NNT can still be worthwhile if the outcome prevented is severe.
Number needed to harm (NNH): How many people must be treated for one additional person to be harmed. The mirror image of NNT.

Statistical concepts

Confidence interval (95% CI): The range of values compatible with the data; loosely, where the true effect plausibly lies. If a risk-ratio CI crosses 1, the result is not statistically significant.
P-value: The probability of seeing a result this extreme if there were truly no effect. Conventionally p < 0.05 is called "significant," but it says nothing about the size or importance of an effect.
Random-effects model: A pooling method that assumes the true effect varies between studies and accounts for that variation. We use the DerSimonian–Laird random-effects model.
Heterogeneity (I²): How much study results differ beyond chance. I² near 0% means studies broadly agree; high I² (say > 75%) means they disagree, and a single pooled number should be read cautiously.
Forest plot: The chart showing each study's effect (a square, sized by its weight) with its confidence interval (a horizontal line), a vertical "no-effect" line at 1, and a diamond for the pooled estimate.
Continuity correction: A small adjustment (adding 0.5 to each cell) applied when a study has zero events in a group, so a risk ratio and its variance can still be computed.