Glossary
Key terms used across the course — short definitions, with a link to the lesson that teaches each one in full when it is ready.
- 2×2 contingency table
The four-cell table behind every effect measure: two groups (treatment/control) by two outcomes (event/no event). Risk, RR, RD, NNT and OR are all just different reads of these four numbers.
- Absolute risk reduction (ARR)
The actual drop in risk in percentage points (control risk − treatment risk) — the same thing as the risk difference. Anchored to how many people per hundred are really affected; NNT = 1 ÷ ARR.
- Absolute vs relative effect
The same result expressed two ways: a relative change (e.g. '50% lower') versus an absolute one (e.g. '1 percentage point lower'). Relative figures sound bigger; absolute ones show real impact and depend on the baseline risk.
- Adjustment (for confounding)
Statistical techniques that try to remove a confounder's influence after data is collected — used when randomisation isn't possible. Includes methods like propensity scores. (Full treatment in M11.)
Full lesson: M11 (coming soon)
- Allocation concealment
Hiding the next treatment assignment until a patient is irreversibly enrolled, so no one can steer who ends up in which group. Distinct from both randomisation and blinding.
- Attrition bias
A systematic tilt from who drops out. If the people who quit a study differ from those who stay, the survivors give a misleading answer.
- Baseline characteristics
A table comparing the trial groups before treatment begins, used to check that randomisation actually produced similar groups — especially important in smaller trials.
- Baseline risk
How common the event is without treatment (the control-group risk). The missing number that decides whether a relative reduction translates into a large or negligible absolute benefit.
- Bias
A systematic tilt built into how a study is designed, run, or measured, pushing the result away from the truth — distinct from random luck.
- Blinding
Keeping patients, carers, or assessors unaware of who received which treatment, so their knowledge can't tilt how outcomes are delivered or measured. The main defence against information bias.
- Bonferroni correction
The simplest fix for multiple testing: divide the 0.05 threshold by the number of tests, so the overall false-alarm rate stays at 5%. Run 10 tests, require p < 0.005 for each.
Full lesson: M3 (coming soon)
- Case series / case report
A description of one patient or a handful, with no comparison group. Can raise a first alarm, but cannot show cause. The bottom of the evidence ladder.
- Case-control study
An observational study that starts from the outcome — people with a disease (cases) versus without (controls) — and looks backward at past exposure. Efficient for rare diseases.
- Censoring
When a patient is observed without the event up to some point — the study ends or they're lost to follow-up. Not missing data: it carries the truth 'survived at least this long,' which survival analysis uses.
- Central limit theorem
Averages tend toward a normal (bell) curve as the sample grows, almost regardless of the underlying data's shape.
- Chance (random error)
Misleading results that arise from pure luck, especially in small samples. Larger numbers reduce it. Distinct from bias, which a bigger sample won't fix.
- Clinical significance
Whether an effect is big enough to actually matter to patients — separate from statistical significance, which only asks whether it's likely real. Large samples can make trivial effects statistically significant.
- Cohort study
An observational study that enters people by their exposure and follows them forward in time to see who develops the outcome. Can be prospective or retrospective — the logic runs exposure → outcome either way.
- Comparator
What a new treatment is measured against — usually the current standard of care, not nothing. HTA almost always asks 'better than what we already do?'
- Confidence interval
A range of plausible values for the true effect: the estimate ± about 2 standard errors (for 95%). Its width shows the uncertainty; whether it excludes the no-effect value mirrors statistical significance.
- Confounding
A hidden third factor that influences both the supposed cause and the effect, creating a misleading link — like joggers living longer partly because they're also younger and richer.
- Cost-effectiveness threshold
A system's 'exchange rate' between money and health at the margin — the price of one QALY. Above it, a treatment displaces more health than it adds.
Full lesson: M7 (coming soon)
- Cox regression
The standard model used to estimate a hazard ratio from survival data, adjusting for other factors. It assumes proportional hazards. (More in M11.)
Full lesson: M11 (coming soon)
- Cross-sectional study
An observational snapshot measuring exposure and outcome at a single point in time. Shows what travels together, but never which came first.
- Diagnostic 2×2 table
The four-cell table crossing test result (positive/negative) with disease status (present/absent): TP, FP, FN, TN. Every diagnostic statistic — sensitivity, specificity, PPV, NPV, LR — reads from these four cells.
- Distribution
The pattern of how often each value occurs across many measurements — the shape data makes when you stack it up. Read it for where values centre and how widely they spread.
- Effectiveness
How well a treatment works in routine, real-world practice — ordinary patients, imperfect adherence, measured against the care people already receive. What HTA needs to know.
- Efficacy
How well a treatment works under ideal, controlled conditions — selected patients, perfect adherence, expert centres. What a regulatory trial is built to show.
- EMA
European Medicines Agency — the regulator deciding whether a medicine may be sold in the EU. Permission to sell is separate from a decision to fund.
- Explanatory trial
A trial run under tightly controlled, ideal conditions to ask whether a treatment can work — maximising internal validity, often at the cost of real-world generalisability. (More in M11.)
Full lesson: M11 (coming soon)
- External validity (generalisability)
How well a study's result carries over to patients outside the trial — older, sicker, or otherwise different from its sample. A true result can still fail to generalise.
- Extrapolation (of survival)
Projecting a survival curve beyond the trial's follow-up, when most events haven't occurred yet — necessary for economic models, and one of the most contested assumptions in an appraisal. (More in M8.)
Full lesson: M8 (coming soon)
- FDA
Food and Drug Administration — the US medicines regulator. Like the EMA, it judges safety and efficacy, not whether a health system should pay.
- Forest plot
A chart showing several studies' estimates as points with their confidence intervals as horizontal lines, against a 'no effect' line — the standard way meta-analyses display and combine results. (Full lesson coming in M4.)
Full lesson: M4 (coming soon)
- Hazard
The instantaneous rate of an event among those who have survived to that moment — a 'speed' of events right now, not a cumulative total. The building block of the hazard ratio.
- Hazard ratio (HR)
A single-number summary comparing the event rate (hazard) between two groups over time; HR < 1 means benefit. It states a relative rate, not extra time or extra survivors, and assumes a constant effect (proportional hazards).
- Hierarchy of evidence
A ranking of study designs by how well they guard against being fooled — anecdote at the bottom, systematic reviews of trials at the top. A starting presumption of strength, not a verdict.
- ICER
Incremental cost-effectiveness ratio — the extra cost of a treatment divided by the extra health it gives, i.e. its price per QALY. Compared against the threshold to judge value.
Full lesson: M7 (coming soon)
- Incidence
How often new cases appear over a period of time — the rate at which people develop a condition. What a cohort, following people forward, can measure. (Full treatment in M3.)
Full lesson: M3 (coming soon)
- Information (measurement) bias
A systematic tilt from how things are measured. If one group is measured differently from another, the comparison is distorted. Recall bias is one variety.
- Intention-to-treat (ITT)
Analysing trial participants in the group they were randomised to, regardless of what they actually did — preserving randomisation's protection against bias. (Full treatment later, in M4.)
Full lesson: M4 (coming soon)
- Internal validity
How confident we can be that a study's result is true for the people actually in it — that the effect is real and not an artefact of bias or confounding.
- Kaplan-Meier curve
A step-shaped survival curve: it drops at each event in proportion to the patients still at risk, and only marks (doesn't drop at) each censoring. The standard way to display time-to-event data.
- Likelihood ratio (LR)
How much a test result updates the odds of disease: LR+ = sensitivity ÷ (1 − specificity); LR− = (1 − sensitivity) ÷ specificity. LR+ > 10 or LR− < 0.1 makes a meaningful difference. Does not depend on prevalence.
- Log-rank test
A statistical test for whether two survival curves differ overall. Commonly paired with the hazard ratio. (More in M3.)
Full lesson: M3 (coming soon)
- Logistic regression
A statistical model for yes/no outcomes that naturally produces odds ratios — which is part of why odds ratios are so common in the literature. (More in M11.)
Full lesson: M11 (coming soon)
- Marketing authorisation
A regulator's decision that a medicine may be sold — based on efficacy and safety, not on cost or value. Permission to sell, not a decision to fund.
- Mean (average)
The arithmetic average — the balance point of a distribution. Fully informative for symmetric data, but on skewed data it gets dragged toward the long tail and can mislead.
- Median
The middle value when data is lined up in order — half above, half below. More honest than the mean for skewed data like costs or survival, because outliers don't drag it.
- Median survival
The time at which the survival curve crosses 50% — half the patients have had the event by then. Preferred over the mean (survival is skewed); reported as 'not reached' if the curve never falls to 50%.
- MeSH
Medical Subject Headings — the controlled vocabulary used to tag records in MEDLINE, so a search can match a paper by its assigned tag even when its text used different words.
- meta-analysis
The statistical step of combining results from multiple studies into a single pooled estimate. Optional, and always comes after selection.
Full lesson: M4 (coming soon)
- Multiple comparisons (multiplicity)
Testing many hypotheses at once — many outcomes, subgroups, or time points — which multiplies the chance of a false alarm. Run enough tests and 'significant' results appear from pure noise.
- narrative review
An expert summary of a field based on the author's own judgement of which studies matter, with no pre-declared rule for inclusion.
- Normal distribution (bell curve)
A symmetric, bell-shaped distribution that appears throughout nature, described completely by two numbers: its mean and its standard deviation. Follows the 68–95–99.7 rule.
- Null hypothesis
The default 'nothing is going on' assumption a study tests against — no effect, no difference, any gap is just chance. Statistics works by trying to embarrass it, not prove the alternative.
- Number needed to treat (NNT)
How many patients you must treat to prevent one event: NNT = 1 ÷ risk difference. Small NNT = powerful; large NNT = many treated per benefit. The key bridge from effect size to cost.
- Observational study
A study that watches what happens without assigning treatment — patients (or doctors) choose. Useful and often large, but open to confounding because the groups differ from the start.
- Odds
Events divided by non-events (not by everyone). Distinct from risk (events ÷ everyone): a 1% risk is odds of about 1-to-99. The basis of the odds ratio.
- Odds ratio (OR)
The ratio of the odds of an event in two groups: (a/b)÷(c/d). Close to relative risk when events are rare, but always further from 1 than RR — it overstates the effect more as events get common. The only effect measure a case-control study can give.
- Opportunity cost
The health a fixed budget was already producing elsewhere, given up to fund something new. In HTA the real cost of a treatment is this displaced health — not the money spent.
- p-hacking
Trying many analyses of the same data — different outcomes, subgroups, cut-offs — and reporting only those that reach p < 0.05. Usually not fraud, but it manufactures false 'findings' from noise.
- p-value
The probability of getting a result at least as extreme as the one observed, if there were truly no effect. How often chance alone would fake your result — not the probability the null is true.
- Performance & detection bias
Subtler tilts in how a treatment is delivered (performance) or how outcomes are looked for and recorded (detection), often tied to lack of blinding. Part of formal risk-of-bias grading. (More in M4.)
Full lesson: M4 (coming soon)
- PICO
A four-slot frame for any clinical question — Population, Intervention, Comparator, Outcome. How HTA turns a vague 'does it work?' into a question evidence can actually answer.
- Placebo
An inactive dummy treatment made to look identical to the real one, so patients (and often clinicians) can't tell who got what. The device that makes blinding possible. Its surgical equivalent is a sham procedure.
- Point estimate
A single best-guess number from a study (like '40% of patients responded') — an estimate of a hidden true value, not the truth itself. Incomplete without a measure of its uncertainty.
- Population vs sample
The population is the full, hidden truth (everyone the result is about); a sample is the handful actually studied. A study measures the sample to estimate the population.
- PPV & NPV
Positive predictive value: how often a positive test is truly positive. Negative predictive value: how often a negative test is truly negative. Both depend on prevalence — unlike sensitivity and specificity.
- Pragmatic trial
A trial run under everyday, real-world conditions to ask whether a treatment works in ordinary practice — favouring external validity over tight control. (More in M11.)
Full lesson: M11 (coming soon)
- Pre- and post-test probability
Pre-test probability is the chance of disease before testing — usually the prevalence or a clinician's prior estimate. A test result updates this to a post-test probability via likelihood ratios or the 2×2 table.
- Pre-specification (& pre-registration)
Declaring a trial's primary endpoint and analysis plan in advance — ideally registered publicly — so the main test keeps its honest false-alarm rate and 'findings' can't be fished out after seeing the data.
- precision
In literature searching, the fraction of retrieved records that are actually relevant (relevant ÷ all retrieved). Distinct from diagnostic specificity.
- Precision
How tightly an estimate is pinned down — higher precision means a smaller standard error and a narrower interval.
- Prevalence
How common a disease is in the population being tested. It doesn't affect a test's sensitivity or specificity, but it critically determines PPV — the same test gives mostly false positives when the disease is rare.
- PRISMA
Preferred Reporting Items for Systematic reviews and Meta-Analyses. A reporting standard consisting of a 27-item checklist and a flow diagram that accounts for every record from search to inclusion.
- Proportional hazards
The assumption behind a single hazard ratio: that the treatment's relative effect stays constant over time. Violated when survival curves cross, converge, or separate only late — and then one HR misleads.
- Publication bias
A distortion in the evidence base itself: studies with striking or positive results are more likely to get published than null ones, making a treatment look better than it is. (Detected with funnel plots, in M4.)
Full lesson: M4 (coming soon)
- QALY
Quality-adjusted life year — the unit HTA uses for 'health gained'. For now, read one QALY as one year of life in full health.
Full lesson: M6 (coming soon)
- Randomisation
Assigning participants to groups by chance, so the groups are alike in every respect except the treatment. The mechanism that lets a trial rule out confounding.
- Randomised controlled trial (RCT)
A study that assigns participants to treatment or control purely by chance, making the groups otherwise identical so nothing hidden can confound the comparison.
- Recall bias
A distortion where people with a disease remember past exposures more thoroughly than those without — a particular threat to case-control studies that rely on memory.
- Regression to the mean
The tendency of extreme measurements to drift back toward the average on remeasurement — a cousin of the winner's curse, and a frequent source of illusory 'improvements.'
- Relative risk (RR)
How many times the risk changes: treatment risk ÷ control risk. RR = 0.5 means half the risk. A ratio — it says nothing about how many people are affected without the baseline risk.
- Relative risk reduction (RRR)
The proportional drop in risk: how much smaller the treated risk is relative to control (e.g. 2%→1% is a 50% RRR). Blind to baseline — the same RRR can mean a huge or trivial absolute benefit.
- Restricted mean survival time (RMST)
The average time alive up to a set time point — an alternative survival summary that doesn't assume proportional hazards, increasingly used when a single hazard ratio can't be trusted.
- review protocol
A pre-registered plan for a systematic review that declares the question, search strategy, and inclusion rules before any results are seen. Registered on PROSPERO or a similar registry.
- Risk difference (RD)
The absolute change in risk: control risk − treatment risk, in percentage points. Unlike relative risk, it reflects how many people per hundred are actually affected.
- ROC curve
Receiver operating characteristic — a plot of sensitivity vs 1 − specificity as the diagnostic threshold moves. Shows the trade-off; the area under the curve (AUC) summarises how well a test separates cases from non-cases.
- Sampling distribution
The spread of an estimate across many hypothetical repeats of the same study; its own SD is the standard error.
- Sampling variation
The wobble in a result caused purely by which individuals happened to land in your sample. It isn't an error — it's built into sampling, and it shrinks as the sample grows.
- search strategy
A PICO question translated into database language: each concept becomes a block, synonyms joined with OR within a block, blocks joined with AND between them.
- Selection bias
A systematic tilt from who gets into a study. If the people studied aren't representative, the result is skewed before any data is collected.
- Sensitivity
The fraction of truly diseased patients who test positive — a test's ability to find cases. A highly sensitive test misses few cases: good for ruling a disease out ('SnOut').
- Skewed distribution
A lopsided distribution with a long tail on one side — common for healthcare costs, length of stay, and survival time. The mean is pulled toward the tail, so the median often describes it better.
Full lesson: M5 (coming soon)
- Specificity
The fraction of truly healthy patients who test negative — a test's ability to clear the innocent. A highly specific test rarely raises false alarms: good for ruling a disease in ('SpIn').
- Standard deviation (SD)
A measure of how spread out individual values are around their average — the width of the distribution. Acts as a unit of distance from the mean. Distinct from the standard error.
- Standard error
How much an estimate (e.g. a mean) would wobble from one repeat of a study to the next — the SD of the sampling distribution. SE = SD / √n.
- Standard of care
The treatment patients would otherwise receive today — the care a new technology must be compared against, not a placebo or an outdated option.
- Statistical power
A study's ability to detect a real effect if one exists — the chance it comes back 'significant' when the drug truly works. Driven mainly by sample size. Low power makes 'not significant' uninformative.
- Statistical significance
A result is called 'statistically significant' when its p-value falls below a chosen threshold (by convention 0.05) — i.e. it would be fairly unusual by chance alone. A flag, not a verdict, and not a measure of importance.
- Subgroup analysis
Looking for an effect within a slice of the trial population (e.g. older patients, a biomarker group). Hypothesis-generating, never confirmatory — and a prime source of multiplicity when many subgroups are examined.
- Surrogate outcome
A short-term, easy-to-measure marker (like a blood-sugar reading) used as a stand-in for an outcome patients actually care about (like avoiding heart attacks). Whether the stand-in holds up is its own question.
Full lesson: M6 (coming soon)
- Survival analysis
Methods for outcomes measured as time-to-event (survival, time to relapse) rather than yes/no. Built to handle patients observed for different lengths and those who haven't yet had the event.
- systematic review
A review that applies a pre-specified, reproducible rule for finding and selecting studies, so the set of included studies is fixed before any results are known.
- Test statistic
A number summarising how far the data sits from 'no effect,' measured in standard errors (e.g. z = difference ÷ SE). The bigger it is, the more extreme the result — and the smaller the p-value.
- Test–treat pathway
The clinical flow: test, interpret, treat (or don't). HTA models this pathway end-to-end — what the test costs, how it changes treatment decisions, and what those decisions cost and deliver. (More in M9.)
Full lesson: M9 (coming soon)
- Type I error (false alarm)
Concluding there's an effect when there really isn't — a false positive. How often it happens is set by the significance threshold (0.05 means a 1-in-20 false-alarm rate).
- Type II error (miss)
Failing to detect a real effect that's genuinely there — a false negative. Usually the result of low power, mostly from too small a sample.
- Variance
The average squared distance of values from the mean; its square root is the standard deviation.
- Winner's curse
The tendency of small studies that reach significance to overstate the true effect — because only an exaggerated, lucky result could clear the bar with few patients. A big reason early findings shrink on replication.
- Z-score
How many standard deviations a value sits from the mean — a way to express 'how typical or surprising' any single value is, on a universal scale. (Full treatment later in M3.)
Full lesson: M3 (coming soon)