Glossary of evidence-appraisal terms

Absolute risk is the actual probability of an event in a group, for example 2 in 100 people over a year. It contrasts with relative risk, which compares groups. Absolute figures show the real-world size of a risk or benefit, which relative percentages can exaggerate.

Absolute risk reduction (ARR) is the plain difference between the event rate in the control group and the event rate in the treated group. If 10 percent of untreated people have an event versus 6 percent of treated people, the absolute risk reduction is 4 percentage points.

Accelerated approval is a regulatory pathway that lets a drug be marketed based on a surrogate endpoint judged reasonably likely to predict clinical benefit, before that benefit is directly confirmed. It trades earlier access for greater uncertainty, with confirmatory studies required afterward.

An adaptive trial design is a clinical trial that lets researchers change specified features (like sample size, treatment arms, or randomization odds) partway through, based on planned interim looks at the accumulating data. Every possible change is defined in advance, so the trial stays statistically valid.

Age standardization adjusts disease or death rates so that populations with different age structures can be compared fairly, by referring both to one common standard age distribution rather than to their own. It removes the distortion that arises because age itself strongly affects most health outcomes.

Allocation concealment hides the upcoming group assignment from the people enrolling participants, until the moment of assignment. It stops recruiters from steering certain patients toward or away from a group. It protects the randomization process itself and differs from blinding, which applies after assignment.

The alpha level is the threshold, chosen before analysis, at which a result is called statistically significant; it is the type I error rate the study is willing to accept, commonly 0.05.

A rule that decides how much of a study's total false-positive budget (its alpha) may be spent at each interim look at the data. It lets a trial peek early without inflating the overall chance of a spurious result.

Analytical validation confirms that a test or measurement reliably and accurately captures the quantity it targets, such as a biomarker's concentration. It asks whether the assay measures what it claims to, separate from whether that quantity is clinically useful.

Ascertainment bias is systematic error arising from the way cases, outcomes, or data are identified, so that some are more likely to be captured than others. The result is a distorted picture that does not reflect the true underlying pattern.

Attributable risk is the extra risk of an outcome in an exposed group that can be ascribed to the exposure, calculated as the risk in the exposed minus the risk in the unexposed. It is the harm-side counterpart of absolute risk reduction and, unlike relative risk, is expressed in absolute terms.

Attrition bias is systematic error introduced when participants who drop out or are lost to follow-up differ from those who remain, and the loss is unequal or related to outcome. Intention-to-treat analysis limits it but does not fully cure it.

The average difference in outcome if everyone in a population received the treatment versus if no one did. It is the target quantity that many causal studies set out to estimate.

Baseline risk is how often an outcome occurs without the treatment or exposure being studied, usually measured in the control or comparison group. It is the starting point against which any effect is judged. The same relative change means far more when baseline risk is high than when it is low.

A trial that tests a single targeted treatment across several different diseases that share a common molecular feature, such as a specific genetic mutation, rather than sharing an organ of origin.

A number that weighs how well the observed data fit one hypothesis versus another, showing which the evidence favors and by how much. Unlike a p-value, it can support the null hypothesis as well as the alternative.

Bayesian inference is an approach to statistics that combines a prior probability with the observed data to produce a posterior probability, expressing results as probability distributions for the quantity of interest.

Blinding, or masking, keeps people in a trial unaware of who received which treatment. It can cover participants, clinicians, outcome assessors, and analysts. Blinding reduces bias from expectations and behavior, so that reported differences reflect the treatment rather than knowledge of the assignment.

A simple way to control false positives when many hypotheses are tested at once: divide the significance threshold by the number of comparisons. It makes each individual test harder to pass.

The Bradford Hill criteria are a set of considerations, such as strength, consistency, temporality, and dose-response, used to judge whether an observed association is likely to be causal. They are aids to reasoning, not a checklist that proves causation.

The Brier score measures how close predicted probabilities land to what actually happened, rewarding predictions that are both confident and correct. Lower is better, with zero being perfect.

The c-statistic (concordance statistic) measures how well a prediction model separates people who have the outcome from those who do not. It is the probability that a randomly chosen person with the event has a higher predicted risk than one without. 0.5 means chance; 1.0 means perfect ranking.

Calibration measures whether a model's predicted probabilities match observed reality. A well-calibrated model that predicts a 30 percent risk for a group should see the event actually occur in about 30 percent of them. It is distinct from discrimination, which is about ranking cases correctly.

Calibration in the large asks a simple question: on average, does a model's predicted risk match the actual event rate in the group? If it predicts ten percent but twenty percent have events, it is miscalibrated in the large.

The calibration slope measures whether a risk model's predictions are too extreme or too timid. A slope of one is ideal; a slope below one means high predictions run too high and low predictions run too low.

A detailed description of a single patient's presentation, treatment, and outcome. It documents one instance rather than measuring an effect across a group.

A descriptive report of the course and outcomes of a group of patients who share a diagnosis or received a similar treatment, with no comparison or control group.

A case-control study starts with the outcome, comparing people who have a disease (cases) with similar people who do not (controls), then looks backward at past exposures. It is efficient for rare diseases and reports an odds ratio rather than a direct risk.

In a setting with competing events, the rate at which one particular type of event happens among people still event-free at that moment. It treats other competing events as simply removing people from the at-risk group.

Censoring happens in survival studies when a participant's exact time to the event of interest is unknown because the study ended, they dropped out, or they were lost to follow-up before it occurred. We only know they stayed event-free up to their last contact.

Certainty of evidence is a rating of how much confidence you can place in an estimate of effect, commonly graded as high, moderate, low, or very low. It reflects how likely further research is to change the conclusion, not how large or how significant the effect is.

Channeling bias is a distortion in drug studies where prescribers steer certain patients (often sicker, higher-risk, or those who failed other treatments) toward a specific medication. Because those patients differ in prognosis before treatment starts, comparisons with other drugs can wrongly credit or blame the drug for outcomes.

Clinical equipoise is a state of genuine uncertainty within the expert medical community about which treatment in a trial is better. It is the ethical condition generally required to justify randomly assigning patients to the arms of a study.

A clinical prediction model combines several patient characteristics into an estimate of the probability that someone has a condition now (diagnosis) or will develop an outcome later (prognosis).

Clinical significance is whether an effect is large enough to matter in real life, changing how a patient feels, functions, or fares. Unlike statistical significance, it asks about the magnitude and relevance of a result, not just whether it is distinguishable from chance.

Clinical validation establishes that a test's result corresponds to a particular clinical state, outcome, or disease in the intended population. It asks whether the measurement means something for patients, not merely whether the instrument reads accurately.

A cluster randomized trial randomly assigns whole groups (clusters), such as clinics, schools, or villages, rather than individual people, to the intervention or control arm. Everyone within a cluster gets the same assignment. It suits interventions delivered at a group level, or where randomizing individuals risks contamination between arms.

Extra care or treatment, beyond the intervention being tested, that one trial group receives more than the other. When it differs between arms, it can manufacture or mask an apparent effect that is not due to the treatment itself.

Cochran's Q is a statistical test for heterogeneity in a meta-analysis, checking whether the spread among study results is larger than chance alone would produce. A small p-value suggests real differences between studies exist.

A cohort study follows a group of people over time, comparing those with and without an exposure to see who later develops an outcome. Because exposure is measured before the outcome appears, cohort studies can establish that the exposure came first, though they do not randomly assign it.

Collider bias is a distortion that appears when you select or adjust for a variable that two other factors both affect. Filtering on that shared effect can create a fake link, or hide a real one, between factors that are not actually related in the wider population.

A competing risk is an event that prevents the outcome a study is tracking from ever happening. If researchers follow people to cancer death, a fatal heart attack is a competing risk: once it happens, that person can no longer die of cancer, so it must be handled specially.

A composite endpoint combines several individual outcomes into one, counting a participant as having an event if any component occurs. It boosts event counts and statistical power, but can blur whether the treatment affected the serious components or only the minor ones.

The probability that a trial will finish with a statistically significant result, calculated partway through using the data seen so far plus an assumption about the remaining effect. It guides interim decisions about futility.

A confidence interval is a range of values, computed from data, that is likely to contain the true effect. A 95 percent interval means that if the study were repeated many times, about 95 percent of such intervals would capture the real value. Its width reflects the precision of the estimate.

Confounding occurs when a third factor is linked to both the exposure and the outcome, creating a misleading association. It can make an exposure look harmful or protective when the real driver is something else the two groups differ on, such as age or smoking.

Confounding by indication is a bias in observational studies where the clinical reason a treatment was given is itself linked to the outcome. The apparent effect of the drug is mixed up with the effect of the condition that prompted it.

Consistency is agreement between the direct evidence (from head to head trials) and the indirect evidence (routed through a common comparator) for the same pair of treatments. When the two disagree beyond chance, the network is inconsistent and its combined estimates are suspect.

CONSORT (Consolidated Standards of Reporting Trials) is a checklist and flow diagram that sets out what the report of a randomized controlled trial should tell readers, from how patients were randomized to how many were analyzed.

A contour-enhanced funnel plot adds shaded regions marking levels of statistical significance to a standard funnel plot, helping distinguish publication bias from other reasons the plot might look asymmetric.

A counterfactual is the outcome that would have happened for the same person or group under a different exposure than the one they actually received. Causal effects are defined as the contrast between an observed outcome and its unobservable counterfactual.

Covariate adjustment is the statistical practice of accounting for other variables in an analysis so that the estimated effect of the factor of interest is separated from their influence.

A widely used regression method for time-to-event data that estimates how much each factor multiplies the rate of an event, without assuming a particular shape for the underlying risk over time. Its main output is the hazard ratio.

A credible interval is a Bayesian range that, given the data and a prior, contains the unknown parameter with a stated probability (say 95%). Unlike a confidence interval, you can literally say there is a 95% probability the true value lies inside it.

A cross-sectional study measures exposure and outcome in a population at a single point in time, like a snapshot. It is useful for estimating how common something is (prevalence), but because it captures everything at once, it usually cannot show which came first.

A randomized trial in which each participant receives two or more treatments in sequence, so each person serves as their own control. The order of treatments is randomized.

A crude rate is the overall rate of an outcome in a population, calculated without adjusting for the population's composition such as its age or sex mix. It reflects the true burden actually experienced but can be misleading when comparing populations that differ in structure.

Cumulative incidence is the proportion of a disease-free group that develops the outcome over a fixed follow-up period. You divide new cases by the number of people at risk at the start. It estimates an individual's average risk and always falls between 0 and 100 percent.

A cumulative meta-analysis repeats the pooled analysis each time a new study is added, in chronological order, showing how the combined estimate has evolved as evidence accumulated. It reveals when the answer became clear.

A data safety monitoring board (DSMB) is an independent group of experts that reviews accumulating trial data during a study to protect participants and judge whether it should continue, change, or stop.

Decision curve analysis judges whether a prediction model or diagnostic test is clinically useful, not just accurate. It plots net benefit across a range of threshold probabilities, weighing the gain from true positives against the harm from false positives, and compares the model to treat-all and treat-none strategies.

Detection bias is systematic error from measuring, looking for, or confirming an outcome more thoroughly in one comparison group than another. When the exposed or treated group is watched, tested, or scored differently, the recorded difference in outcomes partly reflects unequal detection rather than a real effect.

The diagnostic odds ratio (DOR) is a single number summarizing how well a test separates people with a disease from those without it. It equals the odds of a positive result in diseased people divided by those odds in non-diseased people. Higher means better; 1 means useless.

The cutoff value of a continuous test result at or beyond which the test is called positive; where the threshold is set determines the trade-off between sensitivity and specificity.

A study method that estimates a treatment's effect by comparing the change over time in a group that got the treatment with the change over time in a group that did not. Subtracting one change from the other removes any fixed differences between the groups and any trend they shared.

A directed acyclic graph, or DAG, is a diagram of assumed causal relationships drawn as one-way arrows between variables with no feedback loops, used to decide which variables to adjust for. It makes causal assumptions explicit and shows which adjustments reduce bias and which create it.

A test or model's ability to assign higher risk scores to people who have the condition than to those who do not; it is usually summarized by the C-statistic or area under the ROC curve.

A dose-response relationship exists when the size or likelihood of an outcome changes in a consistent direction as the level, duration, or intensity of an exposure or treatment changes. Seeing more effect at higher doses, and less at lower doses, strengthens the case that the exposure genuinely causes the outcome.

The E-value is the minimum strength of association, on the risk-ratio scale, that an unmeasured confounder would need with both the exposure and the outcome to fully explain away an observed association, beyond the variables already adjusted for. Larger E-values signal findings more robust to hidden confounding.

The ecological fallacy is the error of assuming that a relationship seen in group-level averages also holds for the individuals within those groups. A correlation between two things across countries or regions does not mean the same two things are linked in any one person.

Effect modification happens when the size or direction of a treatment or exposure's effect on an outcome genuinely differs across subgroups, defined by a third variable such as age or sex. Unlike a bias, it is a real finding to report, not remove, and analysts detect it by stratifying or testing interaction.

Effect size is a measure of how large an effect or difference is, separate from whether it is statistically significant. It answers how much, using metrics such as a mean difference, a risk ratio, or standardized measures like Cohen's d, so results can be judged for real-world importance.

Efficacy is whether a treatment works under ideal, tightly controlled trial conditions; effectiveness is whether it works in everyday clinical practice with typical patients and imperfect adherence. A drug can prove efficacious in a strict trial yet deliver smaller real-world effectiveness once ordinary conditions apply.

Egger's test is a statistical check for funnel plot asymmetry, used to detect small-study effects such as publication bias in a meta-analysis. It quantifies whether smaller studies give systematically different results than larger ones.

An equivalence trial is designed to show that two treatments differ by no more than a small, pre-specified amount (the equivalence margin) in either direction. Instead of proving one is better, it aims to demonstrate the two are close enough to be considered practically interchangeable.

A precise statement of exactly what a trial is trying to measure: which treatment effect, in which patients, on what outcome, and how complications during the study are handled. It is the question, settled before you compute an answer.

The assumption that treated and untreated groups are comparable, so that the untreated group's outcome fairly represents what would have happened to the treated group had it gone untreated. Randomization creates it by design; observational studies must assume it after adjustment.

A randomized trial designed to test whether a treatment can work under ideal, tightly controlled conditions, using strict eligibility, close monitoring, and high protocol adherence. It measures efficacy rather than everyday effectiveness.

An external control arm is a comparison group drawn from outside the current study, such as past trials, registries, or health records, rather than from people randomized alongside the treated group. It lets researchers estimate a treatment's effect when a concurrent randomized control group is impractical, but it carries a higher risk of biased comparisons.

External validation tests a prediction model on data it has never seen, from a different place, time, or population, to check whether its performance holds up outside the setting where it was built.

External validity is the extent to which a study's findings hold beyond the specific participants, settings, and conditions studied. A trial can be internally sound yet still leave open whether its results transfer to other populations, care settings, or time periods.

A randomized trial that tests two or more interventions at once by randomizing participants to every combination of them, letting a single trial answer more than one question.

A graphical tool for turning a pre-test probability into a post-test probability: you draw a straight line from the pre-test probability through the test's likelihood ratio to read off the result.

The expected proportion of false positives among all the findings a study flags as significant. Controlling it accepts some errors in exchange for detecting more true effects.

The probability of making at least one false-positive claim across a whole set, or family, of statistical tests. Methods that control it aim to keep that combined risk at a chosen level.

A fixed-effect meta-analysis assumes every included study is estimating one single true effect and combines them by weighting each study mainly by its size. It treats all differences between study results as chance alone.

A forest plot is a graph that summarizes a meta-analysis. Each study appears as a box (its result) with a horizontal line (its confidence interval), sized by how much weight the study carries. A diamond at the bottom shows the pooled result combining all studies.

The number of patients whose outcomes would have to flip, from non-event to event, to turn a statistically significant trial result into a non-significant one. A small index means the finding rests on just a few events.

A funnel plot is a scatter graph used to check a meta-analysis for publication bias. Each study is a dot plotted by its effect size against its precision. Larger, more precise studies sit near the top; smaller ones scatter below. A roughly symmetric funnel shape is reassuring.

Generalizability is how well conclusions drawn from a study sample extend to the broader population it is meant to represent. It depends on who was sampled and how, so a result from a narrow or unrepresentative sample may not describe the wider group accurately.

An international ethical and quality standard for designing, running, recording, and reporting trials that involve people. Following it is meant to protect participants and keep the resulting data credible.

GRADE is a structured system for rating how much confidence to place in a body of evidence. It sorts certainty into four levels: high, moderate, low, or very low. It rates the evidence for each outcome separately, not a single study, and is used across systematic reviews and guidelines.

A trial design that plans a series of interim looks at the accumulating data, with pre-set rules that allow early stopping for benefit, harm, or futility while protecting the overall error rate.

Hazard ratio (HR) compares how quickly an event happens in two groups over time. It is the ratio of their hazard rates, the instantaneous chance of the event among those still at risk. An HR of 1.0 means no difference; below 1.0 means slower events in the treated group, above 1.0 means faster.

The healthy-user effect is a confounding pattern in observational studies where people who take a preventive treatment or attend screening tend to be healthier and more health-conscious than those who do not. Their better outcomes may reflect that lifestyle rather than the treatment.

Heterogeneity is variation among the studies combined in a meta-analysis, differences in their results beyond what chance alone would explain. High heterogeneity signals the studies may differ in populations, treatments, or methods, making a single pooled estimate harder to interpret.

The Hosmer-Lemeshow test is a classic check of whether a risk model's predictions match observed outcomes across groups of patients sorted by predicted risk. A large p-value is read as reassurance of good calibration.

I-squared is the percentage of the total variation across studies in a meta-analysis that reflects genuine differences between studies rather than chance. Higher values mean the study results are more inconsistent with one another.

Immortal time bias is a distortion in cohort studies where a stretch of follow-up is guaranteed event-free because of how exposure was defined. A person must survive long enough to receive a treatment, so that waiting time gets wrongly credited to the treated group, faking a survival benefit.

Imprecision describes how much random error clouds an effect estimate, usually shown by a wide confidence interval. When the interval is wide enough to span meaningfully different decisions, certainty in the evidence is rated down for imprecision.

Imputation is the practice of filling in missing values with estimated ones so that participants with incomplete records can still be included in an analysis.

Incidence is the rate at which new cases of a condition appear in a population over a set period. It counts only people who develop the condition during that time, among those who were previously unaffected and at risk. It measures how fast a condition is arising, unlike prevalence, which counts existing cases.

An incidence rate is the number of new cases of an outcome divided by the total person-time at risk. Unlike a proportion, it has units of cases per unit of time (for example, per 1000 person-years) and is not bounded between 0 and 1.

An incidence rate ratio is the incidence rate in one group divided by the incidence rate in another, showing how many times faster new cases arise in the exposed group. A value of 1 means the rates are equal; above 1 means the exposed group develops the outcome faster.

Inconsistency is unexplained variation in results across the studies pooled for an outcome. When effects point in different directions or differ widely in size without a clear reason, certainty in the combined estimate is rated down.

The diagnostic test whose accuracy is being evaluated in a study, measured by comparing its results against the reference standard.

Indirectness is the gap between the evidence you have and the question you are asking, for example when the studied patients, treatments, comparisons, or outcomes differ from the ones you care about. It is one of the reasons certainty in a body of evidence gets rated down.

An individual participant data meta-analysis gathers the raw, person-level data from each included study and reanalyzes it centrally, instead of pooling the summary results each study published. This allows consistent analyses and a reliable look at how effects vary by patient characteristics.

Information bias is systematic error that arises when an exposure, outcome, or other variable is measured or classified inaccurately. Unlike random measurement noise, it distorts results in a consistent direction.

When people leave a study before an event for reasons tied to their underlying risk, so those still being followed no longer represent everyone. It quietly biases survival estimates.

An instrumental variable is a factor used in observational research that influences whether people get a treatment but affects the outcome only through that treatment, and shares no common cause with it. Researchers use it to estimate causal effects while reducing distortion from unmeasured confounding.

Integrated discrimination improvement summarizes how much adding a new predictor to a model widens the gap in predicted risk between people who do and do not have the outcome. It complements net reclassification improvement.

Intention-to-treat analyzes trial participants in the group they were randomized to, regardless of whether they finished, switched, or followed the treatment. By preserving the original random groups, it keeps randomization's protection against bias and estimates the effect of assigning the treatment, which reflects how it tends to perform when prescribed rather than under perfect adherence.

Something that happens after treatment starts and changes how a patient's outcome should be read, such as stopping the assigned drug, switching treatments, or taking rescue medication. How these events are handled is central to defining a trial's estimand.

An interim analysis is a planned examination of trial data before enrollment and follow-up are complete. Independent monitors compare treatment groups partway through to decide whether to continue, stop early for clear benefit or harm, or halt for futility, using pre-specified statistical rules that guard against being misled by early, unstable results.

Internal validity is the degree to which a study's design and conduct support a genuine cause-and-effect link between the intervention and the measured outcome, rather than the result being produced by bias, confounding, or chance. High internal validity means you can trust the finding within that study.

A design that tracks an outcome at many time points before and after an intervention, then looks for a shift in the level or slope of the trend at the moment the intervention began. The pre-intervention trend stands in for what would have happened otherwise.

A technique that reweights each participant by the inverse of their probability of receiving the treatment they actually got, creating a pseudo-population in which treatment is unrelated to the measured confounders. Comparisons in that reweighted population estimate the treatment effect.

The Kaplan-Meier estimate is a method for charting the probability that an event has not yet happened over time, such as survival. It handles censored subjects (those still event-free when follow-up ends) and produces the familiar stepped survival curve that drops at each observed event.

A survival method that picks a fixed time point after baseline, then compares outcomes only among people still event-free at that moment, grouped by their status up to it. It avoids crediting a group for time before an exposure could act.

Last observation carried forward (LOCF) is a simple imputation method that replaces a participant's missing later value with their most recent measured one.

Lead-time bias is an illusion of longer survival that comes purely from diagnosing a disease earlier, without changing when the person actually dies. Screening moves the diagnosis date forward in time, so measured survival from diagnosis looks longer even when the true outcome is unchanged.

Length-time bias is the tendency of screening to preferentially detect slowly progressing disease, because slow cases spend longer in a detectable phase and are more likely to be caught. This makes screen-detected disease look more survivable than it is.

A likelihood ratio tells you how much a diagnostic test result shifts the odds that a condition is present. The positive LR compares the chance of that result in people with the condition versus without it. Values far above 1 raise probability; values near 0 lower it.

A living systematic review is a systematic review kept continually up to date, with new studies incorporated on a frequent schedule rather than in a single fixed snapshot. It suits fast-moving fields where the evidence changes quickly.

Loss to follow-up is when participants enrolled in a study drop out or cannot be reached before their final outcome is measured, so their results are missing. If those who leave differ systematically from those who stay, the missing data can distort the study's conclusions.

A model, usually fit with inverse probability weighting, built to estimate the effect of treatments that change over time when time-varying confounding is present. It recovers the effect a sustained treatment strategy would have had.

A set of methods that split a total effect into the part carried through an intermediate variable (the indirect effect) and the part acting by other routes (the direct effect). It answers how, or through what, an exposure changes an outcome.

A mediator is a variable that lies on the causal pathway between an exposure and an outcome, carrying part or all of the exposure's effect. Adjusting for a mediator removes the portion of the effect that travels through it.

Mendelian randomization uses genetic variants as natural stand-ins for an exposure to test whether that exposure causes an outcome. Because gene variants are set at conception and largely independent of lifestyle, they can reduce confounding and reverse causation that trouble ordinary observational studies.

A meta-analysis is a statistical method that combines results from multiple studies into a single pooled estimate. By merging data, it can produce a more precise effect estimate than any single study, but its reliability depends entirely on the quality of the studies included.

Meta-regression examines whether study-level characteristics, such as average dose or mean patient age, help explain the differences in effects seen across studies. It fits a regression relating those characteristics to each study's result.

The minimal clinically important difference (MCID) is the smallest change in an outcome that patients or clinicians would regard as meaningful, rather than merely detectable.

Missing data are values a study intended to collect but did not obtain, such as outcomes for participants who dropped out or skipped a visit.

Modified intention-to-treat (mITT) is an analysis that starts from all randomized participants but excludes a pre-defined subset, such as those who never received any treatment or had no post-baseline data.

Multiplicity is the problem that arises when a study runs many statistical tests at once. Each test carries its own chance of a false positive, so testing many outcomes, subgroups, or time points inflates the overall probability that at least one "significant" result is a fluke.

An n-of-1 trial is a controlled experiment in a single patient. That one person cycles repeatedly between a treatment and a comparator (often placebo), usually in randomized, blinded order, with outcomes measured in each period, to find which works best for them individually.

A natural experiment is a study that exploits a naturally occurring event or policy that assigns exposure in a way that approximates randomization, letting researchers estimate causal effects without a planned trial. Its credibility rests on the assignment being unrelated to the outcome's other causes.

An outcome the exposure could not plausibly cause, examined to see whether a study's methods produce a spurious association anyway. If the exposure appears linked to something it cannot affect, hidden bias or confounding is probably at work.

How likely a negative test result is in people who have the condition compared with people who do not; it equals (1 minus sensitivity) divided by specificity.

Negative predictive value is the chance that someone who tests negative truly does not have the condition. It is the proportion of negative results that are true negatives. Like PPV, it depends on how common the disease is: NPV tends to be high when the condition is rare in the tested group.

Net benefit is a single number that puts the value of finding true cases and the cost of false alarms on the same scale, so you can judge whether acting on a test or model does more good than harm at a chosen risk threshold.

A measure of whether adding a new marker to a risk model correctly shifts people across risk categories: upward for those who later have the event and downward for those who do not.

Network meta-analysis compares three or more treatments at once by combining direct evidence from head-to-head trials with indirect evidence, where treatments are linked through a shared comparator. It can estimate and rank options even when some pairs were never tested against each other.

A study approach that includes only people starting a treatment for the first time and follows them from that start, rather than mixing in people already established on it. This avoids biases that arise from studying long-term survivors of a treatment.

Node-splitting is a check in network meta-analysis that separates the direct evidence for a treatment comparison from the indirect evidence and tests whether they agree. A meaningful disagreement flags local inconsistency in the network.

A non-inferiority trial tests whether a new treatment is not meaningfully worse than an existing one, rather than better. Researchers set a pre-specified "non-inferiority margin," the largest acceptable loss of benefit, and conclude non-inferiority only if the new treatment stays within that margin.

The null hypothesis is the default claim that there is no real effect or no difference, for example that a treatment works no better than placebo. Statistical tests assess how well the data fit this assumption; a small p-value gives reason to doubt it.

Number needed to harm (NNH) is the number of people who must receive a treatment or exposure for one additional person to experience a specified harm, over a given time. It is the reciprocal of the absolute risk increase: an absolute increase of 2 percentage points gives an NNH of 50.

Number needed to screen is how many people must go through a screening program to prevent one bad outcome, such as one death from the targeted disease, over a given period.

Number needed to treat (NNT) is the number of people who must receive a treatment for one additional person to benefit, over a given time. It is the reciprocal of the absolute risk reduction: an absolute reduction of 5 percentage points gives an NNT of 20.

The odds ratio compares the odds of an outcome in one group to the odds in another. A value of 1 means no difference. Above 1 means the outcome is more likely in the exposed or treated group; below 1 means less likely. It is common in case-control studies and logistic regression.

A trial in which both participants and investigators know which treatment each person is receiving; there is no blinding. Randomization may still be used to assign treatment.

The optimal information size is the number of participants or events a body of evidence would need to reliably detect a given effect, roughly the sample a single adequately powered trial would require. Falling short of it is a reason to rate evidence down for imprecision.

Optimism is the gap between how well a model appears to perform on the same data used to build it and how well it truly performs on new data. Models almost always look better on their home turf than they really are.

Quietly changing which outcomes a trial reports compared with what was planned, for example promoting a secondary outcome to primary because it looked better. It inflates the chance that a highlighted result is a false positive.

Overdiagnosis is the correct detection of a condition that would never have caused symptoms or harm during a person's lifetime. The finding is technically accurate, but identifying it brings no benefit and can lead to treatment of something that would have stayed silent.

Overfitting happens when a model learns the noise and quirks of its development data rather than the real signal, so it looks accurate at home but performs worse on new patients.

A p-value is the probability of getting results at least as extreme as those observed if the null hypothesis (no real effect) were true. It measures how surprising the data are under that assumption. It does not give the probability that the hypothesis itself is true or false.

A per-protocol analysis includes only participants who followed the trial's rules closely, completing the assigned treatment without major deviations. It estimates the effect under ideal adherence, but by dropping non-compliers it can break randomization's balance and overstate benefit. It is usually read alongside intention-to-treat.

The treatment effect you would see if everyone actually followed the trial's protocol as intended. It is a target of estimation, distinct from the per-protocol analysis that naively tries to measure it.

Performance bias is systematic error from differences in the care or attention that comparison groups receive, apart from the intervention being tested. Blinding of participants and staff is the main defense against it.

Person-time is the total amount of time that all participants in a study are observed and at risk for the outcome, added together. It forms the denominator of an incidence rate, so ten people followed for one year and one person followed for ten years both contribute ten person-years.

Clinical trial phases are the ordered stages of testing a treatment in humans: phase I checks safety and dose in small groups, phase II explores preliminary efficacy, phase III compares against standard care in larger populations, and phase IV monitors safety after approval. Each phase answers a different question and gates the next.

A placebo is an inactive comparison, such as a dummy pill or sham procedure, made to resemble the real treatment. It lets a trial separate a treatment's specific effect from improvements caused by expectation, attention, or the natural course of illness. Comparing treatment to placebo isolates the true added benefit.

A trial that evaluates several treatments against a shared common control under one master protocol, with treatments allowed to enter or leave over time as evidence accumulates.

A pooled estimate is the single combined result, such as a summary risk ratio or mean difference, produced by statistically averaging the individual studies in a meta-analysis. Each study is weighted, usually by its precision.

The population attributable fraction (PAF) is the proportion of an outcome in a whole population that would be avoided if a harmful exposure were removed, assuming the exposure causes the outcome. It depends on both how strongly the exposure raises risk and how common the exposure is.

How many times more likely a positive test result is in people who have the condition than in people who do not; it equals sensitivity divided by (1 minus specificity).

Positive predictive value is the chance that someone who tests positive actually has the condition. It is the proportion of positive results that are true positives. Unlike sensitivity and specificity, PPV depends heavily on how common the disease is in the tested group, so it falls sharply when the condition is rare.

A condition for valid causal comparison requiring that every type of person in the study has a nonzero chance of receiving each treatment being compared. If some group could only ever receive one treatment, its effect there cannot be estimated from the data.

Post-market surveillance is the ongoing monitoring of a drug or device's safety and performance after it is approved and in general use. It can detect rare or delayed harms that pre-approval trials were too small or too short to reveal.

The updated probability that a person has the condition after their test result is taken into account, obtained by combining the pre-test probability with the test's likelihood ratio.

In Bayesian analysis, the posterior probability is the updated probability of a hypothesis or parameter after the prior has been combined with the observed data.

A way of defining causal effects as the comparison between the outcome a person would have under one treatment and the outcome the same person would have under another. Because only one can ever be observed, the other is a counterfactual.

A randomized trial designed to measure how well a treatment works in everyday clinical practice, using broad eligibility, routine settings, and usual care as the comparison. It aims to inform real decisions rather than isolate a mechanism.

Pre-registration means publicly recording a study's hypotheses, outcomes, and analysis plan in a time-stamped registry before data are collected or examined. It locks in what researchers said they would test, so readers can check whether the published paper matches the original plan rather than a story shaped after seeing the results.

The probability that a person has the condition before their test result is known, estimated from the prevalence in similar people plus the individual's own symptoms and risk factors.

In a random-effects meta-analysis, a prediction interval gives the range in which the true effect of a new, similar study is expected to fall. It combines the uncertainty in the average effect with the real variation between studies, so it is wider than the confidence interval around the pooled estimate.

A predictive biomarker forecasts whether a patient is likely to benefit, or suffer harm, from a specific treatment. It helps choose therapy, unlike a prognostic factor, which only forecasts the general course of disease.

Prevalence is the proportion of a population that has a condition at a given time. It counts existing cases (both old and new) divided by everyone in the group. Because it reflects how common something is right now, it depends on both how often the condition arises and how long it lasts.

The primary endpoint is the single main outcome a trial is designed and powered to measure, chosen in advance to answer its central question.

In Bayesian analysis, the prior probability is the probability assigned to a hypothesis or parameter before the current study's data are taken into account.

PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) is a checklist and flow diagram for reporting how a systematic review was searched, screened, and synthesized.

A trial that is randomized and open-label for patients and treating doctors, but has outcomes judged by assessors who do not know the assignment. PROBE stands for Prospective Randomized Open Blinded Endpoint.

A prognostic factor is a patient or disease characteristic that predicts the likely course of an illness, regardless of which treatment is given. It tells you what tends to happen, not what to do about it.

A propensity score is each person's estimated probability of receiving the treatment given their measured baseline characteristics. In observational studies, researchers use it to compare treated and untreated people who looked equally likely to be treated, mimicking randomization to reduce confounding from those measured factors.

The core requirement of a Cox regression: the hazard ratio between groups stays constant over the whole follow-up period. In plain terms, one group's relative risk versus the other does not grow or shrink as time passes. If it does, a single reported hazard ratio can mislead.

Any departure from what a trial's protocol specified, such as a missed visit, a dosing error, or enrolling someone who did not fully meet the criteria. Frequent or serious deviations can distort results and signal problems in how the trial was run.

Publication bias is the tendency for studies with positive, striking, or statistically significant results to be published more often than those with null or negative findings. As a result, the published literature can overstate an effect compared with all the research actually conducted.

The step of creating the unpredictable order in which participants will be assigned to trial groups, for example with a computer random number generator. Doing it soundly is the foundation of randomization; concealing it is a separate safeguard.

A random-effects meta-analysis assumes the true effect varies from study to study and estimates the average of a distribution of effects rather than a single value. It gives relatively more weight to smaller studies than a fixed-effect model does.

A randomized controlled trial assigns participants to treatment or comparison groups by chance. Randomization tends to balance known and unknown differences between groups, so any outcome gap can more credibly be attributed to the intervention. It is the strongest single study design for testing whether a treatment causes an effect.

Real-world evidence (RWE) is clinical evidence about a treatment's benefits or risks that comes from analyzing real-world data, information collected during routine care rather than in a controlled trial. Sources include electronic health records, insurance claims, disease registries, and data from apps or wearable devices.

Recall bias is a systematic error that occurs when the accuracy or completeness of participants' memories differs between the groups being compared. It is most damaging in case-control and other studies that ask people to remember past exposures.

The best available method for deciding whether the target condition is truly present, against which a new diagnostic test is compared when estimating its accuracy.

An observational study built on a registry, a structured collection of data on people who share a given disease, exposure, device, or procedure. Patients are observed as they are treated rather than assigned to groups at random.

Regression is a family of statistical methods that model how an outcome changes with one or more explanatory variables, estimating each variable's association with the outcome while holding the others fixed.

A method used when a treatment is assigned by a cutoff on a continuous score, such as a lab value or an age. People just below and just above the threshold are treated as nearly identical, so comparing their outcomes isolates the treatment's effect near the cutoff.

Regression to the mean is the tendency for extreme measurements to move closer to the average when repeated. A very high or low value often reflects part chance, so the next reading tends to be less extreme, even with no treatment or real change.

Relative risk (RR) is the ratio of the chance of an outcome in one group to the chance in a comparison group. An RR of 1.0 means no difference; below 1.0 means the outcome is less common in the first group, and above 1.0 means it is more common.

Relative risk reduction (RRR) is the proportion by which a treatment lowers the risk of an outcome compared with the control group. If risk falls from 10 percent to 6 percent, that is a 4-point absolute drop but a 40 percent relative reduction, because 4 is 40 percent of the original 10.

Reporting bias is the distortion of the evidence base that occurs when whether and how results are shared depends on their direction or statistical significance rather than their scientific merit. It is an umbrella term whose forms include publication bias, when whole studies go unpublished, and selective outcome reporting, when specific results within a study are dropped or downplayed.

The average event-free time a group accumulates up to a chosen cutoff, found by measuring the area under the survival curve out to that time point. It expresses survival results in plain units like months rather than as a ratio.

When the outcome actually causes the exposure rather than the other way around, making a real association point in the wrong direction. Often the developing disease changes a behavior or a measured factor before it is diagnosed.

A judgment about whether flaws in how a study was designed, run, or reported could have pushed its results away from the truth. It asks not whether a study is biased, but how much we should worry that it might be.

Cochrane's structured tool for judging risk of bias in randomized trials. It walks a reviewer through set questions across several domains and yields a rating of low risk, some concerns, or high risk of bias.

A structured tool for judging risk of bias in non-randomized studies of interventions, such as cohort comparisons of one treatment against another. Its name stands for Risk Of Bias In Non-randomized Studies of Interventions.

An ROC curve plots a test's true-positive rate against its false-positive rate across every possible cutoff. The area under the curve (AUC) condenses that plot into one number: the probability the test ranks a random positive case above a random negative one. 0.5 is chance; 1.0 is perfect.

A phase before randomization in which candidates receive a placebo or the active drug, so the trial can screen out poor adherers or non-responders before the real comparison begins. It can sharpen a trial's power but narrow who the results apply to.

A secondary endpoint is an additional outcome a trial measures beyond its primary one, used to describe further effects but not usually the basis for the trial's main conclusion.

Selection bias arises when the people included in a study, or retained in it, differ systematically from the population the results are meant to describe. Because of how participants were chosen or who dropped out, the sample gives a distorted picture that does not generalize accurately.

Sensitivity is how well a test detects people who truly have the condition. It is the proportion of people with the disease who correctly test positive. A test that is 95 percent sensitive misses 5 percent of true cases (false negatives). When sensitivity is high, a negative result helps rule the condition out.

A sensitivity analysis repeats an analysis under different reasonable assumptions or after excluding certain data to see whether the main conclusion still holds. If the result barely changes, the finding is considered robust.

A modeling technique that deliberately pulls extreme estimates toward a central value or overall average, trading a little bias for more stable, reliable predictions. It counters the tendency of small or noisy groups to produce exaggerated results.

Simpson's paradox is when a trend that appears in every subgroup of the data reverses or vanishes once the subgroups are combined. It typically happens because a lurking variable is unevenly distributed across the groups being pooled.

Small-study effects are the tendency for smaller studies in a meta-analysis to report larger or more favorable treatment effects than larger ones. They often show up as asymmetry in a funnel plot.

Specificity is how well a test correctly clears people who do not have the condition. It is the proportion of disease-free people who correctly test negative. A test that is 90 percent specific wrongly flags 10 percent of healthy people (false positives). High specificity means a positive result is fairly convincing.

Spectrum bias occurs when a diagnostic test's measured sensitivity and specificity shift because the study sample does not reflect the full range of patients who would be tested in practice. Studying obviously sick people against clearly healthy controls makes a test look more accurate than it truly is.

Standard deviation measures how far individual values in a dataset typically fall from their mean, expressed in the same units as the data.

Standard error estimates how much a summary statistic, such as a sample mean, would vary from one sample to another; it reflects the precision of the estimate, not the spread of the data.

A standardized mortality ratio (SMR) is the number of deaths observed in a study group divided by the number expected if that group had the death rates of a reference population, matched on factors like age and sex. A value of 1 means as many deaths as expected, above 1 means more, and below 1 means fewer.

Statistical power is the probability that a study will detect an effect of a specified size when that effect truly exists; it equals one minus the type II error rate.

Statistical significance means a result is unlikely to have arisen by chance under the null hypothesis, usually judged by a p-value below a preset threshold such as 0.05. It signals that an effect was detected, but says nothing about how large, important, or real-world meaningful that effect is.

A cluster randomized design in which every cluster starts in the control condition and then crosses over to the intervention at a randomly assigned point in a staggered sequence, until all clusters have received it.

A stopping rule is a pre-specified statistical threshold that determines when a trial may be ended early, whether for clear benefit, harm, or little chance of a useful answer.

A study protocol is the written plan that specifies a trial's objectives, design, eligibility criteria, outcomes, and analysis before the data are collected.

A quantity used in competing-risks analysis that connects directly to the cumulative probability of a specific event, keeping people who had a competing event in the at-risk pool rather than removing them. It underlies the Fine-Gray model.

Subgroup analysis splits a study's participants into groups (by age, sex, disease severity, and so on) and estimates the treatment effect within each. It asks whether the intervention works differently for different kinds of people, rather than reporting a single average effect for everyone in the trial.

A summary of findings table is a compact table that presents, for each main outcome of a review, the size of the effect, the number of people studied, and the certainty of the evidence. It is meant to convey the bottom line at a glance.

A trial designed to detect whether one treatment is better than a comparator, such as another drug or a placebo. Its goal is to show a difference in favor of the new treatment.

SUCRA is a single number, from zero to one, that summarizes how highly a treatment ranks against all the others in a network meta-analysis. A value near one means the treatment tends to be among the best; near zero, among the worst.

A surrogate endpoint is a substitute measure, such as a lab value or scan finding, used to stand in for an outcome that matters directly to patients, like survival or symptoms. It is easier or faster to measure but does not always predict the real outcome.

Survivorship bias is a distortion that arises when analysis includes only the subjects, records, or units that made it past some selection point, while those that dropped out, failed, or died are silently missing. Conclusions then reflect the survivors rather than the whole.

A technique that builds a comparison for a single treated case, such as a country or hospital, by weighting several untreated units so their combined pre-treatment history matches the treated one. Divergence afterward estimates the treatment's effect.

A systematic review answers a focused question by searching for, appraising, and summarizing all relevant studies according to an explicit, pre-planned method. Its transparent and reproducible process is designed to reduce bias, unlike an informal review that cites only a convenient or favorable selection of studies.

A framework for analyzing observational data by first writing down the randomized trial you would ideally run, then emulating each of its features (eligibility, treatment start, follow-up, outcome) as closely as the data allow. It disciplines the analysis against common time-related biases.

Tau-squared is the estimated variance of the true effects across studies in a random-effects meta-analysis, measured on the same scale as the effect. Its square root, tau, is the estimated standard deviation of those true effects.

Temporal validation checks a prediction model on patients from a later time period than those used to build it. It sits between reusing the development data and fully external testing in a new place.

Confounding that changes during follow-up, where a variable is both affected by past treatment and influences future treatment and the outcome. Standard adjustment fails here because the confounder is also a step on the causal path.

Transitivity is the assumption behind indirect comparison: that studies comparing A to B and B to C are similar enough in their patients and settings that B can serve as a shared bridge to compare A and C. If the trials differ in ways that affect the outcome, the bridge is unreliable.

A way of defining a trial's target question that counts every patient's outcome regardless of what they did after assignment, including stopping the drug. It is the estimand that most closely matches a classic intention-to-treat analysis.

Trim and fill is a method that estimates how many studies may be missing from an asymmetric funnel plot, then imputes them to gauge how much publication bias might have shifted the pooled result. It is a sensitivity check, not a correction of the truth.

TRIPOD is a reporting checklist that tells authors what to include when they publish a clinical prediction model, so readers can judge how it was built and validated. It is the prediction-model counterpart to CONSORT for trials.

A type I error is rejecting a null hypothesis that is actually true, that is, concluding there is an effect when there is none. It is a false positive.

A type II error is failing to reject a null hypothesis that is actually false, that is, missing a real effect. It is a false negative.

An umbrella review is a review of existing systematic reviews and meta-analyses on a broad topic, summarizing what those reviews collectively found rather than analyzing primary studies directly. It sits one level above a systematic review.

Variance is the average of the squared distances between each value and the mean; it is the square of the standard deviation and measures how spread out a set of values is.

Verification bias distorts a diagnostic test's accuracy when whether a patient gets the reference (gold-standard) test depends on the index test result. It typically inflates sensitivity and deflates specificity.

The Will Rogers phenomenon is an illusion in which survival appears to improve in every disease stage without anyone actually living longer, simply because better tests reshuffle patients between stages. It is also called stage migration.

A way to compare two groups on several ranked outcomes at once by pairing up patients and deciding who wins on the most important outcome first, then less important ones. The win ratio is wins for treatment divided by wins for control.

A single summary of a test's performance at a chosen cutoff, calculated as sensitivity plus specificity minus one, ranging from 0 for a useless test to 1 for a perfect one.