NotebookGlossary
Glossary
Terms from the slides, each with a one-line meaning and links to the topics that teach it.
412 terms
#
- 2x2 contingency table (DTA layout)
- Rows = index test (+/−); columns = reference standard (Yes/No); TP top-left, FP top-right, FN bottom-left, TN bottom-right; row totals = total test +/−; column totals = total disease/normal.
- 2x2 table notation
- A, b, c, d as laid out on p.30.
- 3+3 dose escalation design
- Cohorts of 3 at pre-set dose levels; escalate, expand or step down by number of DLTs.
- 68–95–99.7 rule
- Mean ± 1/2/3 SD covers 68/95/99.7% of observations in a normal distribution.
- 95% confidence interval
- Mean ± 1.96 × SE: the range likely to contain the population mean (or difference), calculated from the estimate and its standard error.
A
- Absence of evidence ≠ evidence of absence
- P > 0.05 doesn't prove no difference; beta error is possible.
- Absolute reduction in percentage points
- E.g. 23% → 13% = "10% reduction".
- Accuracy
- Nearness of a measurement to the true value.
- Adaptive trial design
- Pre-specified changes based on accumulating data (e.g. wrong SD assumptions, dropping arms); beyond module scope.
- Adjusting for measurement error
- Generally needs multiple measurements: regression calibration, simulation extrapolation (SIMEX), other advanced methods. Beyond the module's scope.
- Adjuvant treatment
- Additional therapy after the main treatment (surgery); chemo, RT, monoclonal antibody.
- Advantages of surrogates
- Quicker and cheaper; shorter trials; can be non-invasive.
- Allocation concealment
- Sequence hidden from recruiter, e.g. opaque, sequentially numbered, sealed envelopes held by an independent party.
- Allocation concealment vs blinding
- Concealment: next allocation (would be); blinding: received allocation (were).
- Alpha error (type I error, false positive)
- Rejecting the null hypothesis when it is actually true: concluding there is a difference when there is none.Also: Alpha error (type I, false positive); Type I error (α, false positive)
- Alternate hypothesis
- The opposite of the null (test) hypothesis, e.g. "the difference is at least the MCID".Also: Alternate (alternative) hypothesis
- Analysis differs with more than 2 groups
- Different statistical tests are used.
- Analytical study
- Finds the association between predictors and outcomes.
- Antagonistic interaction
- Combined effect is less than the sum but greater than either individual effect (13 pp: <15, >10, >5).
- Area under the ROC curve (AUC)
- The example shows 0.6494. Not explained on the slides.
- Arithmetic mean
- Sum of the values ÷ number of observations; affected by extreme values.
- Assessing random error
- Alpha via statistical tests/p-values (assuming the predictor is the only difference and the data are error-free); beta via sample size calculations (later lecture).
- Association
- A relation between two variables: knowing one helps determine the other. For binary outcomes it is measured by the risk ratio, odds ratio and risk difference.Also: Association / measures of association
- Attrition bias
- Eligible participants excluded after initial inclusion; equivalent to bias due to missing data.
B
- Balanced ("just right") sample size
- Keeps error risks at the desired level and avoids waste and delay.
- Berkson error model
- When neither the true value nor replicates are available, assign everyone in a subgroup the subgroup average (e.g. factory nicotine monitors).
- Beta error (type II error, false negative)
- Not rejecting the null hypothesis although there is a real difference, usually because wide CIs overlap.Also: Beta error (type II, false negative); Type II error (β, false negative)
- Bias due to confounding
- Bias that occurs when one or more prognostic variables also predict the intervention or exposure received, distorting the estimated association.Also: Confounding
- Bias due to deviations from intended interventions (performance bias)
- Systematic differences in the care provided between groups that deviate from the intended interventions.
- Bias due to missing data
- Later follow-up missing for included individuals, or individuals with missing intervention/confounder data excluded.
- Bias in classification of interventions
- Bias from differential or non-differential misclassification of intervention/exposure status.Also: Bias in classification of interventions (RCT)
- Bias in measurement of outcomes
- Assessors aware of intervention status; different methods by group; or measurement error related to intervention.Also: Bias in measurement of outcomes / blinding outcome assessors
- Bias in selection of the reported result
- Authors choose which results to report based on the findings (outcome measures, time points).
- Binary (dichotomous) variable
- Nominal variable with only two possibilities (alive/dead, disease present/absent).
- Binomial distribution
- Distribution of a binary outcome; parameters n and p ("probability of success").
- Binomial vs Poisson framing of adverse events
- People with ≥1 event (binomial) vs total number of events (Poisson).
- Bland–Altman plot
- Differences vs averages of two measurements. Points within mean difference ± 2 SD of the differences mean minimal measurement error.
- Blinding
- Participants and providers unaware of the group to which participants were randomised.
- Box plot
- Box = quartiles, line = median, whiskers, outliers shown as points.Also: Boxplot
- Bright-line threshold problem
- Decisions shouldn't rest only on p ≤ 0.05; consider study design and measurement error.
C
- Case series vs cohort
- Functional definition: a cohort compares ≥2 exposure groups; otherwise it is a case series.
- Case-control study
- Start with cases (with outcome) and controls (without); compare past exposure.
- Case-control vs cross-sectional
- Distinguished by starting point and by timing of exposure (past vs same time).
- Case-control vs retrospective cohort
- Distinguished by starting point: cases/controls vs outcome-free people.
- Cases and controls
- People with the outcome (cases) and without the outcome (controls).
- Categorical variable
- Answers "What type?"; nominal or ordinal.
- Categorising quantitative data
- Use clinically meaningful thresholds and count values in each range, then summarise as a proportion.
- Causal association
- Exposure determines the outcome (vs a mere relation).
- Causal inference
- Inferring that the exposure causes the outcome; strengthened by temporal sequence.
- Causes of beta error
- Small n gives wide CIs, so samples from different populations overlap.
- Central Limit Theorem
- With large n, sample means are approximately normal and centred at μ, whatever the original distribution.
- Chi-square test for trend / Cochran-Armitage trend test
- Unpaired ordinal (few categories), binary responses, 2 groups.
- Choosing a threshold
- By trial and error, or (more easily) with an ROC curve.
- Choosing summary measures by distribution
- Normal: mean and SD; non-normal: median and quartiles.
- CI for a proportion
- P ± 1.96 × ASE (SAS also reports the exact limits).
- CI interpretation
- About the population mean, not 95% of individual values; the sample mean is not the true mean.
- CI of difference vs CIs of each sample
- For continuous data, non-overlapping 95% CIs of the two samples mean p < 0.05. This is not reliable for nominal categorical data (SE under null vs alternative), and is slightly off when SDs differ greatly.
- CI of the difference and significance
- If the CI of the difference excludes the test-hypothesis value, the difference is significant; the CI also shows its magnitude.
- CI overlap as a chance check
- Overlapping 95% CIs mean the samples could come from the same population.
- Classification of research studies
- Clinical/preclinical studies are observational, quasi-experimental or experimental.
- Cluster RCT
- Whole groups (e.g. schools) are randomised rather than individuals.
- Cluster-adjusted analysis
- Analysis "takes into account clustering".
- Clustered bar chart
- Bars for an outcome's categories grouped side by side within each exposure group.
- Co-interventions / time-varying confounders in ITS
- Other changes at the same time (antibiotic, disinfectant) can explain the trend change.
- Cochran-Mantel-Haenszel test
- Unpaired ordinal (few categories, multiple responses), and paired ordinal (few categories), 2 groups.
- Cohen's kappa
- Measure of inter-rater reliability (agreement beyond chance; the formula is not on the slides). The example value 0.52 is computed "by software".
- Cohort multiple RCT
- A cohort is used to run repeated RCTs by randomly offering interventions to eligible members; aims to improve recruitment.
- Cohort study
- Group specified at outset, without outcome, exposure measured, followed over time.
- Combining concealment and blinding
- Pharmacist-prepared, sequentially numbered drug containers.
- Compatibility interpretation
- Smaller p means greater incompatibility of the data with the test hypothesis, if assumptions hold; it may point against the null or the assumptions.
- Confidence interval width / margin of error
- Higher with too few samples.
- Confirmation bias
- Unconsciously interpreting, favouring and recalling information that supports one's beliefs.
- Confounding variable
- A prognostic variable (other than the one of interest) that is also related to the exposure/intervention.
- Consequence of surrogate misuse
- Wrong conclusions; harmful interventions despite surrogate improvement.
- Consequences of too few samples
- More false positive and false negative conclusions; wide confidence intervals.
- Consequences of too many samples
- Wasted human, financial and time resources; delayed results and adoption.
- Context of assessment
- The same methods assess accuracy (test vs reference standard), precision (repeat measures of the same field/sample) and sampling error (different fields/samples from the same subject).
- Context-dependence of variable roles
- The same factor can be a confounder in one question and a mediator in another (physical activity).
- Contingency table
- Cross-tabulated frequencies of two categorical variables.
- Continuous variable
- Takes any value within a given range (height, weight, age).
- Control group
- Another intervention or no active treatment.
- Control-selection bias
- Source of controls can bias results; same-population or multi-source controls preferred over hospital-based.
- Conventional alpha
- 0.05 common; 0.01 rare (may suit multiple hypotheses); 0.10 and 0.001 rarer.
- Conventional beta
- 0.20 and 0.10 (power 0.80 and 0.90).
- Correlation coefficient
- Measures association between two quantitative variables (r = 0.17 for height vs weight).
- Correlation coefficients as agreement measures (limitation)
- They measure association and ignore any consistent difference between observations, which can lead to wrong conclusions.
- Counted (discrete) variable
- Counts within a specified time or area; non-negative (children, infections, heart attacks).
- Counted vs ordinal distinction
- Counts have equal intervals (8−6 = 6−4); ordinal grades do not (muscle strength).
- Cross-over RCT
- Every participant receives both intervention and control in sequence.
- Cross-sectional / cohort / case-control studies
- Named as the major categories of analytical observational study (not defined).
- Cross-sectional study
- Exposure and outcome measured at a certain point or period of time.
D
- Data
- Factual information used as a basis for reasoning; in research, the values a variable takes for each subject/material.
- Data dictionary and coding
- Each variable has a type, a description and numeric codes for categories (e.g. 0 = No, 1 = Yes).
- Decimal places
- Number of digits to the right of the decimal point.
- Decision-driven directional question (Examples 1, 2)
- "at least as good as" (A cheaper and safer), or "better by at least 5%" (A costlier and riskier).
- Degrees of freedom
- The single parameter of the t-distribution; df = n − 1 for one sample.Also: Degrees of freedom (n − 1)
- Dependence on prior information
- Required inputs vary by outcome type and hypothesis.
- Descriptive study
- Describes distributions of predictors and outcomes in a population.
- Descriptive vs analytical observational studies
- Analytical studies compare ≥2 groups differing by exposure; descriptive studies do not.
- Design effect
- The factor by which the sample size is inflated in cluster designs; depends on average cluster size and ICC.
- Diagnostic test accuracy (DTA)
- Accuracy of a diagnostic test (lab, radiology, clinical exam, histopathology, endoscopy) against the reference standard.
- Dichotomising a quantitative test
- Pick a threshold; below = Negative, at or above = Positive; then build the 2x2 table.
- Difference in means / medians
- Effect measure for a quantitative outcome; mean or median chosen by the distribution.
- Difference-in-Differences
- Compare within-group before-after changes between treated and untreated groups.
- Different observation periods per cluster
- An analysis consideration in stepped-wedge.
- Differential co-treatment / supportive care
- E.g. extra antibiotics, nutrition, monitoring or early treatment in one arm only.
- Differential measurement methods
- Outcome assessed differently between groups (lighting; accurate scale vs home weighing).
- Differential misclassification
- Related to the outcome or perceived risk of the outcome; likely to cause bias.
- Direction ambiguity in cross-sectional data
- It may be unclear which variable is exposure and which outcome (obesity/video-gaming).
- Distribution of data
- Graph (or function) showing how often values or intervals occur.
- Dose-limiting toxicity (DLT)
- Drug-related toxicity that limits dose.
- Dose-response study
- A multi-arm use case that asks "does effect increase with dose".
- Dose–response pattern (implied)
- RR/OR increase steadily across ordered exposure levels.
- Double dummy design
- Each arm gets its active drug plus a placebo matching the other arm's route or form.
E
- Effect size vs effect measure
- "Effect size" = SMD (Cohen's d) specifically. RR, OR, RD, MD and SMD are all "effect measures".
- Effect-of-interest dependence
- Whether a deviation causes bias depends on whether the ITT or the per-protocol effect is wanted.
- Empirical
- Capable of being verified or disproved by observation or experiment.
- Equivalence and equivalence margin
- You must pre-define what counts as equivalent (a margin) because identical results are near impossible. P-value calculations differ.
- Equivalence trial
- Tests that treatments are equivalent within a zone.
- Ethical comparison design
- Replace an unethical "do vs don't" comparison with an ethical active comparison (app reminder vs standard advice).
- Evidence of surrogate failure
- FDA 2008–2012 PFS/DFS approvals: only about 15% improved overall survival (13/36 insufficient evidence, 18/36 no improvement). Of those 18, HRQoL data for 7: 2 worse, 4 no difference, 1 mixed.
- Excel T.DIST.2T
- Returns the two-tailed p-value from |T| and df.
- Excess kurtosis
- Kurtosis − 3 (often reported as "kurtosis" by software); a CI that excludes 0 means non-normal.
- Expected vs desired proportion
- Expected comes from a previous study; desired is a clinically meaningful change.
- Experimental study
- Researcher applies an intervention and examines its effect; all experimental studies are analytical.
- Exploratory (unplanned) analysis
- Should not strongly influence conclusions; deviations must be explained.
- Exploratory vs confirmatory trial
- Exploratory usually needs no adjustment; confirmatory needs adjustment for doses/regimens of the same drug, with no consensus for distinct treatments.
- Exposure
- The predictor whose association with the outcome is studied; it defines the groups compared. Strictly an exposure is not determined by the researcher, unlike an intervention.Also: Exposure vs intervention; Exposure vs outcome; exposed vs unexposed
- Exposure odds (case-control)
- E/(N−E) in cases and in controls.
- External validity (generalisability)
- How far results apply beyond the people studied; a non-representative sample, or exclusions applied to both groups, limit it without biasing the comparison.Also: Representative sample / external validity; External validity; External validity / generalisability
F
- Factorial efficiency
- Without interaction, the same participants answer both questions (300 vs 600 or 450).
- Factorial RCT (2×2)
- Two-stage randomisation to factor 1 present/absent and factor 2 present/absent, giving 4 cells.
- Factors for choosing a test
- Data type, distribution, number of groups, paired vs independent.
- Feasibility sample size rule of thumb
- Formal calculation may be inappropriate; about 24–50 patients; surrogate outcomes may be used.
- Feasibility study
- Assesses whether a full study is possible and necessary, and how; estimates design parameters.
- Feasibility vs pilot distinction
- They are conceptually different (Reality check 4).
- Finding the MID
- Multiple methods; a research study in its own right; search the literature for the MID of a measurement tool.
- Fisher's exact / Chi-square test
- Unpaired nominal outcome, 2 groups (Chi-square for 3+). Caution with small expected counts.
- Fleiss kappa
- Kappa-type statistic for any number of observers; used for binary variables with multiple observers and (per the slide) ordinal variables with few categories; interpreted like Cohen's kappa.
- Frequency → probability distribution
- Frequencies convert to probabilities; each probability is between 0 and 1.
- Frequency table
- Frequency, percent, cumulative frequency and cumulative percent per category.
G
- Geometric / harmonic mean
- Other kinds of mean (named only; not covered).
- Gold standard
- A test that gives the true value. Rare in clinical medicine.
- Grouping data / equal class intervals
- Group values to avoid too many categories; intervals must be similar or the shape misleads.
- Groups
- Samples that differ in the predictor variable.
H
- Histogram
- Normal data look symmetric and bell-shaped; minor deviations are acceptable.Also: Histogram (normality check)
- Hypothesis
- A tentative assumption made to draw out and test its consequences; "an assumption that can be tested".
I
- ICC and CCC
- Overcome the correlation limitation. Roughly similar results, but many versions cause confusion.
- Ideal surrogate scenario
- The surrogate is the only causal pathway from disease to clinical outcome, and the intervention acts on that pathway.
- Identifying interaction by regression
- Add an interaction term and test its significance.
- Incidence rate
- New cases per unit of person-time: r/(n × t).
- Index test
- The test whose accuracy is being measured.
- Inputs for a binary outcome
- Α, β, control-group proportion, expected or desired intervention proportion.
- Inputs for a continuous outcome
- Α, β (or power), difference in means, variance/SD of both groups.
- Intention-to-treat (ITT) effect
- Benefit/harm of starting (being assigned to) a treatment.Also: Intention-to-treat effect
- Interaction (between interventions)
- The effect of one intervention depends on whether the other is given.
- Interaction alters power in factorial trials
- If unrecognised, leads to wrong conclusions; more participants needed.
- Interpreting a hypothesis test
- Reject the null and accept the alternate if p is below the pre-set threshold (and groups differ only in treatment/time). The CI approach is recommended as more clinically relevant.
- Interquartile range (IQR)
- Upper − lower quartile; not influenced by extreme values.
- Interrupted Time Series
- Compare the outcome trend before vs after an intervention.
- Interventions applied in conjunction
- Giving one does not prevent giving another; prerequisite for a factorial design.
- Intra-cluster correlation coefficient (ICC)
- How similar outcomes are within a cluster; must be included in cluster analysis.
J
- Just noticeable difference
- The smallest change that can be perceived (e.g. less than 0.25 °C is not noticed); noticing a change does not mean a change in decision.
K
- Kolmogorov–Smirnov test
- Normality test where a low p-value means non-normal; can give false negatives, so do not rely on it alone.
- Kurtosis
- Mainly about outliers or extreme values, not peak sharpness; normal = 3.
L
- Lack of adherence
- Participants stop treatment, or the researcher does not deliver it as planned (intentionally, unintentionally, or by failure); responsibility doesn't change the bias assessment.
- Landis–Koch kappa bands
- Poor (<0), Slight (0–0.20), Fair (0.21–0.40), Moderate (0.41–0.60), Substantial (0.61–0.80), Almost perfect (0.81–1.00).
- "Large" sample size
- Above 30 (enough for some, e.g. binomial) or 60 according to others.
- Lifetime prevalence
- Proportion who have ever had the characteristic at some point in life.
- Limits of the 1.96 multiplier
- Use t for small n and for differences in means; transform non-normal data; medians need a different method; use software.
- Link between sample size and hypothesis testing
- Sample size calculations are "closely linked with hypothesis testing".
- Log-rank test
- Time-to-event outcome, any number of groups.
- Longitudinal study
- Same individuals followed over time.
- Loss to follow-up
- Participants lost during follow-up; absent in cross-sectional designs.
- Lower alpha for multiple hypotheses
- E.g. several interventions vs inactive treatment.
M
- Mann-Whitney U test (= Wilcoxon rank-sum test) / Kruskal-Wallis test
- Unpaired non-normal or ordinal-multi or count outcomes, 2 / 3+ groups.
- Marginal totals
- Row and column sums along the margins of a contingency table.
- Matched case-control study
- Controls matched to cases on confounders; only known confounders are addressed.
- Maximum tolerated dose (MTD)
- Dose at which 0 or 1 of 6 patients have a DLT.
- McNemar test / Cochran Q-test
- Paired binary nominal outcome, 2 groups / 3+ groups.
- Mean vs median skew rule
- Mean > median suggests positive skew; mean < median suggests negative; with exceptions.
- Mean/median comparison
- Summary measure for quantitative outcomes between groups (cross-sectional, cohort, RCT).
- Measurement error
- Error in accuracy, precision or sampling (or all three). It contributes to random and systematic error rather than being a separate entity.
- Measurement error in confounders
- Error in even one confounder distorts the adjusted contribution of the other confounders, even those measured without error.
- Measures of effect: absolute vs relative
- The difference between exposed and unexposed in the outcome, expressed as an absolute or a relative measure.
- Mechanism vs association (confounding hint)
- A significant difference doesn't confirm the proposed mechanism. Other behaviours (e.g., social distancing) may explain it.
- Median
- Middle of the data, with 50% of values on each side; robust to extremes but ignores some information.
- Mediator variable
- A third variable in the causal path from exposure to outcome.
- Method choice for binary outcomes
- Accuracy: Sn, Sp and other DTA measures. Precision/sampling error: % agreement, Cohen's kappa.
- Methods for quantitative measurement error
- Bland–Altman; Pearson/Spearman correlation; ICC; CCC; SD; RSD; CV.
- Minimal important difference (MID)
- Smallest difference patients or proxies perceive as important (beneficial or harmful) that would prompt a change in management.
- Missing at random vs not missing at random
- Maximal collection limits data not missing at random; use appropriate methods for data missing at random.
- Mitigations when blinding is impossible
- Common protocol with objective criteria; second blinded team.
- Mixed retrospective-prospective cohort
- Retrospective cohort whose follow-up continues into the future.
- Mode
- Most common value; descriptive only, not used in analysis.
- Moderator variable
- The exposure–outcome association differs across values of a third variable; the exposure does not depend on it.
- Multi-armed parallel RCT
- More than two concurrent arms; needs a larger sample than two arms.
- Multiple-testing adjustment in multi-arm trials
- Whether alpha or sample size should be adjusted for several comparisons; described as "controversial".
- Multiplication rule for independent events
- P(A and B) = P(A) × P(B), here for draws with replacement.
N
- Named QE designs
- Instrumental variables, regression discontinuity, ITS, DiD, fixed effects (named only).
- NC3Rs Experimental Design Assistant
- Stepwise visual tool for designing animal experiments.
- Negative interaction
- "Combination treatment is less effective than at least one of the individual treatments" (3 pp, less than 10 pp).
- Negative predictive value
- The proportion of negative index-test results that are true negatives: TN/(FN + TN).
- Nested case-cohort study
- Random sample of the cohort drawn first, then divided into cases and controls; allows several outcomes.
- Nested case-control study
- Cases at end of cohort follow-up compared with randomly selected non-cases from the same cohort.
- NNT / NNH
- 1 / RD. NNT: treated per one fewer bad outcome (beneficial intervention). NNH: treated per one more bad outcome (harmful).
- No interaction (additive)
- The effect of chemo is the same with or without RT; combined effect = sum of individual effects (15 pp).
- No previous study → feasibility study
- Run one to gather the inputs.
- No safe missingness threshold
- No level of missing outcome guarantees no bias.
- No SD needed for a binary outcome
- The variance is calculated from the proportions.
- Nominal variable
- Categories with no specific order (blood group, eye/skin colour, gender).
- Non-differential misclassification
- Unrelated to the outcome (random); usually biases towards the null (false negatives).
- Non-inferiority margin
- The acceptable difference, usually below the MCID / clinically acceptable difference.
- Non-inferiority trial
- Tests that one treatment is not worse than another by more than an acceptable margin.
- Non-research meanings of "bias"
- Dictionary sense (prejudice/preconceived opinion); electrical sense (voltage applied to a device) is a different meaning.
- Normal (Gaussian) distribution
- Family of symmetric bell-shaped curves; many biological measures; bell shape alone is not proof.
- Normal approximation
- The normal distribution approximates others under certain conditions (e.g. binomial with n > 30).
- Normal curve overlay
- Fitted normal density drawn over the histogram, to judge normality.
- Normal vs non-normal distribution
- Counted variables are usually non-normal. Determines parametric vs non-parametric.
- Null (test) hypothesis
- The hypothesis actually tested: no (pre-specified) difference between the groups. In this module it is framed around the MCID, e.g. "the difference is less than the MCID".Also: Null hypothesis (Hulley); Null hypothesis (Myors); Test hypothesis (Greenland); Null hypothesis
O
- Objective outcome criteria
- Decrease measurement bias when blinding assessors is hard.
- Observational study
- Researcher takes a passive role in making measurements on study subjects.
- Observed vs desired difference rule
- Use the observed difference only if it exceeds the MID; designing around an unimportant difference is illogical.
- Odds
- Number with the event ÷ number without it (r/(n − r), or a/b in a 2×2 table); can be any positive number.Also: Odds (summary measure); Odds (in a group)
- Odds ratio (OR)
- Odds in exposed ÷ odds in unexposed.
- One-sample t (sample mean vs known population mean)
- Compares a sample mean with a known population mean: T = (μ − x̄)/(s/√n), converted to a p-value with df = n − 1.
- Only one outcome per case-control study
- Case-control studies one outcome; cohort and cross-sectional can study more.
- Open allocation
- A visible sequence lets recruiters selectively enrol, causing confounding (liver-cancer example).
- OR to RR conversion
- RR = OR / [1 + control risk × (OR − 1)].
- OR–RR divergence
- The OR moves further from the RR as the control-group risk (outcome frequency) rises.
- Ordinal exposure with a reference category
- Each exposure level is compared with a chosen reference level (RR/OR = "Reference" for that level).
- Ordinal variable
- Ordered categories (cancer grade, pain intensity).
- Ordinal variables with many categories
- Treat them as quantitative.
- Other outcomes and designs
- More complex; seek expert advice; EQUAL-STATS.
- Outcome
- The event or characteristic whose occurrence is compared between exposure groups.
- Outcome / intervention
- The basis for judging whether an intervention works. "Intervention" is preferred to "treatment" because it covers prevention.
- Outcome variable
- The variable storing each participant's outcome data.
- Outlier
- Extreme value lying beyond the whiskers (labelled on the plot).
P
- P-hacking / data dredging / selective reporting
- Reporting only significant analyses makes p-values uninterpretable. Pre-plan analyses.
- P-value (ASA)
- Probability, under a specified model (the test hypothesis plus valid data collection), of a data summary at least as extreme as observed.
- P-value ≠ P(hypothesis true) or P(chance alone)
- A p-value is about the data given a hypothesis, not about the hypothesis or explanation.
- Paired data and paired-profiles plot
- The same people measured twice; a line joins each person's two values, plus a mean line.
- Paired t-test / repeated measures ANOVA
- Paired normal continuous outcome, 2 / 3+ groups.
- Paired vs independent (unpaired) samples
- Paired = repeated measures in the same person, matched individuals, twins, split body-part; independent = separate unrelated groups.
- Parallel design pros/cons
- Commonest and simple; needs many RCTs; cannot assess interactions.
- Parallel trend assumption
- Without treatment, the treated group's average change would equal the control group's.
- Parameters
- Variables inside a distribution's formula; usually 1–2 in medicine (mean, or mean and SD).
- Parameters a feasibility study estimates
- SD of outcome, patient willingness, clinician willingness, number eligible, follow-up, response and adherence rates.
- Parametric and non-parametric tests
- Parametric tests are used for normally distributed data; non-parametric tests make fewer distributional assumptions but are more prone to false negatives.Also: Parametric tests; Parametric vs non-parametric tests
- Patient-reported outcome measures
- The patient is the outcome assessor, so an unblinded patient introduces bias.
- Pearson vs Spearman
- Pearson if the variables are normally distributed; Spearman if not.
- Per-protocol effect
- Benefit/harm of adhering to a treatment.
- Percentage agreement
- (both Yes + both No) / total observations.
- Period prevalence
- Proportion with the characteristic at any point during a period (often past 12 months).
- Person-time units
- Rate expressed per 1000 person-years or person-months, depending on the units of t.
- Pie chart
- Circle split into slices proportional to category share; used for a categorical variable.
- Pilot sample size
- No special calculation; based on recruitment period and main-study sample size.
- Pilot study
- A mini-version of the main study, checking that components work together; training; go/no-go; external or internal.
- Placebo
- Inert comparator identical in route, colour, shape, taste, texture; its choice can cause bias.
- Placebo and nocebo effects
- Effects of expectation (positive or negative).
- Point prevalence
- Proportion with the characteristic at a specific point in time.
- Poisson distribution
- For rare counted variables; parameter = events per unit time.
- Policy-maker vs researcher conflict
- The motivation for stepped-wedge (implement vs evaluate).
- Positive predictive value
- The proportion of positive index-test results that are true positives: TP/(TP + FP).
- Power
- The chance of detecting a real difference: power = 1 − β.
- Power calculation
- The false negative error for a given sample size, observed difference and variability, at the desired alpha.
- Precision
- Agreement between a set of replicate or repeat measurements. It says nothing about accuracy.
- Preclinical sources of bias
- Incubator conditions, housing, batches, discarded cultures, differential nutrition/monitoring, unblinded assessment of cultures or animals.
- Predictor variable
- The variable assumed to influence the outcome (an exposure or intervention); the study tests whether it changes the outcome variable.Also: Predictor variable / outcome variable
- Prevalence
- Proportion of people with the disease or outcome at a point or period of time (r/n).
- Probability density function (PDF)
- Formula for a continuous distribution (term sometimes used for both).
- Probability mass function (PMF)
- Formula for a discrete probability distribution.
- ProbOnto
- Knowledge base of about 150 distributions, their PDFs, parameters and alternative parameterisations.
- Prognostic variable
- "Variable that determines the outcome".
- Proportion / percentage
- Number with the characteristic or event ÷ total; × 100 for a percentage. A proportion can be a risk or a rate.Also: Proportion (percentage)
- Prospective
- Collect information that happens now and in future.
- Prospective cohort
- Data collection starts now and participants are followed forward.
- Purpose of randomisation / baseline imbalance
- Reduce confounding by balancing confounders between arms.
Q
- Q-Q plot
- Normal data fall roughly on a straight line against normal quantiles.
- Qualitative interaction (direction)
- The effect of one treatment changes direction in the presence of the other.
- Qualitative research
- Research based on information that cannot be easily measured (opinions, feelings).
- Quantitative interaction (magnitude)
- The effect changes in size but not direction.
- Quantitative research
- Research based on information that can be shown in numbers.
- Quantitative variable
- Answers "How much?"; continuous or counted.
- Quartiles / percentiles
- Values that split sorted data: the lower quartile (Q1) is the 25th percentile, the median the 50th and the upper quartile (Q3) the 75th.Also: Median and quartiles
- Quasi-experimental study
- Another category with five major designs: instrumental variables, regression discontinuity, interrupted time series, difference-in-differences, fixed effects (named only).Also: Quasi-experimental designs
R
- RAA tool / clinical relevance of preclinical research
- Assessing the clinical relevance of animal studies (Gurusamy 2021) and cell lines (Gillet 2013).
- Random allocation / assignment
- Assigning participants to arms by chance.
- Random error (chance variation)
- Error due to chance: samples from the same population differ purely by chance (4 heads and 6 tails from a fair coin). Alpha and beta errors are both random errors.Also: Random error (sampling variation); Chance (sampling) variation; Random error
- Random measurement error
- No pattern between groups. The errors sum to zero, so the average is unaffected but variability rises. You need a larger sample size, or accept a higher beta error.
- Random vs systematic sampling error
- No pattern in how exposed and unexposed groups are measured, vs a systematic difference between them (e.g. BP in the morning vs the evening).
- Randomised controlled trial (RCT)
- The main type of experimental study design; participants are similar at entry.Also: Randomised controlled trial
- Range
- Highest − lowest value; a measure of variation.Also: Range (highest/lowest)
- Rare disease / outcome and design choice
- Case-control is best; cross-sectional is impractical; prospective cohort and RCT are inefficient.
- Rate
- Proportion of events over a specified time (e.g. mortality rate = deaths per year).
- RCT strengths and weaknesses
- Best for causal association; expensive and inefficient for rare outcomes.
- RCT vs cohort difference
- Random assignment; a well-designed prospective cohort can minimise all other biases.
- Reasons for poor surrogate accuracy
- Alternate intervention pathways; alternate disease pathways; surrogate change not translating into equivalent clinical change.
- Recall bias
- People with the outcome (e.g. cancer) remember past exposure more than people without it.Also: Recall bias / bias in classification of exposure
- Reduced between-person variation leads to smaller sample size (cross-over)
- Within-person comparison.
- Reducing beta error
- Shrink SE by increasing n (main route in biology) or reducing SD. Choosing similar subjects reduces SD but costs external validity. In manufacturing, reduce variability.
- Reference standard
- The most accurate test available. The best way to determine whether a subject has the target condition.
- Regression modelling
- Recommended for paired data comparisons (footnote 1), ordinal few-category 3+ groups, and multiple-response nominal paired data. Easier with normal data.
- Relative standard deviation / coefficient of variation
- RSD = SD / mean × 100. Below 5% means minimal measurement error (interpret in context). Conventions differ: fraction vs %, and whether the sign is kept.
- Research fraud
- Data deliberately altered to favour one intervention (vs unconscious bias).
- Research misconduct
- Selective reporting can count as misconduct and cause harm (Excel trial scandal).
- Research question to hypothesis
- A good hypothesis states the sample studied, the predictor and the outcome.
- Residual confounding
- Confounding that remains despite matching, adjustment or other steps.
- Retrospective
- Collect already existing information, e.g. review of records.
- Retrospective cohort
- Cohort identified from existing data in the past and followed to the present.
- Retrospective exposure ascertainment
- Exposure collected from past records or memory; where misclassification mainly occurs; rare in experiments.
- Reverse causality
- The outcome causes a change in exposure (heart attack reduces physical activity).
- Risk
- Number with the event ÷ everyone in the group (r/n, or a/(a + b) in a 2×2 table); always between 0 and 1.
- Risk difference (RD)
- Risk in exposed − risk in unexposed (sign shows direction).
- Risk ratio (RR)
- Risk in exposed ÷ risk in unexposed.
- Risk-of-bias tools
- Matching each study type to its tool (RoB 2, ROBINS-I, QUADAS-2, QUIP, PROBAST, SRQR, ROBIS, AGREE-II, SYRCLE, QHES).Also: Risk-of-bias assessment tools; Risk-of-bias tools for design
- Risk↔odds conversion
- Odds = Risk/(1−Risk); Risk = Odds/(1+Odds).
- ROC curve
- A plot of sensitivity (y) against 1 − specificity (x) across thresholds.
- Rounding convention
- Commonly 2 decimal places and/or 3–4 significant digits.
- Rounding of intermediates (implied)
- Final ratios should come from unrounded intermediate values.
S
- Same-population framing
- Comparing groups tests whether the samples come from the same population.
- Sample and population
- Samples are studied to draw conclusions about populations; the sample mean and SD are the best estimates of the population values but are not identical to them.Also: Sample vs population (estimates); Sample-to-population inference
- Sample size calculation
- "Calculate the minimum sample size required to ensure that the false positive and false negative errors are at or below the level that is desired".
- Sample size for a binary outcome (illustrative)
- 23% vs 10%, α = 0.05, power 0.80 gives 142 per group, rounded to 150.
- Sample size vs power calculation
- Expected/desired inputs at the design stage vs observed inputs after the data are collected.
- Sampling error
- Measurement error arising from how sampling is done (which microscope field, which tissue part, when/where the blood is drawn, time of day for BP). It can be random or systematic.
- Scatterplot matrix
- Grid of all pairwise scatter plots, with histograms on the diagonal.
- Scenario (sensitivity) analysis / multiple imputation
- Analyse under different assumed outcomes (favourable/unfavourable) or use multiple imputation.
- SD / SE / 95% CI alongside mean or median
- Occasionally reported with the central value.
- SD as a measure of measurement error
- Also called RMSE/RMSD. Hard to interpret alone.
- SD vs mean skew rule of thumb
- Minimum 0: SD > mean/2 is likely skewed, SD ≥ mean highly likely; otherwise use D = mean − lowest (or highest − mean).
- Selection bias
- Eligible participants, initial follow-up time, or outcome events related to both intervention and outcome are excluded (selectively, from one group).
- Selection bias vs missing-data bias
- Before inclusion (selection) vs after inclusion (missing data).
- Sensitivity
- The proportion of people with the target condition whom the index test calls positive: TP/(TP + FN).
- Serial cross-sectional study
- Repeated cross-sectional surveys at intervals; not a cohort or longitudinal study.
- Seven bias domains reduced in RCTs
- Confounding, selection, classification of interventions, deviations, missing data, outcome measurement, reported-result selection.
- Seven domains of bias (ROBINS-I)
- Confounding; selection of participants; classification of interventions; deviations from intended interventions; missing data; measurement of outcomes; selection of reported result.
- Sham surgery and its ethics
- Surgical placebo; ethically debated; even harder in physiotherapy and psychology trials.
- Shape of the binomial depending on p
- P = 0.2 gives a right-skewed shape (n = 10); p = 0.5 gives a symmetric bell shape.
- Significant digits
- All digits, ignoring the decimal point, leading zeros and some trailing zeros.
- Simple bar chart
- One bar per category, height = frequency.
- Simple before-after study
- Single-group pre/post comparison, confounded by general improvement over time.
- Simple randomisation / random sequence generation
- Coin-toss-like sequence generated in advance (random number tables, now computers).
- Skewness
- Asymmetry about the mean; a longer right tail is positive skew, a longer left tail is negative skew.
- SnNout
- With high sensitivity, a negative result rules OUT the target condition.
- Software context
- Tables and graphs come from SAS; students do not have to produce graphs.
- Sources of inputs, continuous outcome
- Conventional α/β; difference from a previous study (observed) or desired; SD from a previous study.
- Specificity
- The proportion of people without the target condition whom the index test calls negative: TN/(FP + TN).
- Split-part design
- The person is the cluster and their parts (e.g. teeth) are the units of analysis.
- SpPin
- With high specificity, a positive result rules IN the target condition.
- Stacked bar chart
- One bar per group, divided into category segments; bar height = group total.
- Standard deviation (SD)
- Measure of variation in the sample (no formula given).
- Standard error (SE)
- The SD of sample means, estimated from one sample as SD/√n; it measures the precision of an estimate and shrinks with a smaller SD or a larger n.Also: Standard error of the mean
- Standard normal distribution
- Normal distribution with mean 0 and SD 1.
- Standardised mean difference
- Mean difference ÷ standard deviation.
- Statistical significance
- A p-value below a pre-set threshold, usually 0.05.Also: Statistical significance (p ≤ 0.05)
- Statistical test
- Testing the (test) hypothesis; finding out whether an observed difference is due to chance.
- Statistical vs practical (clinical) significance; effect size
- P doesn't measure effect size or importance.
- Stepped-wedge cluster RCT
- All clusters eventually get the intervention, in a randomised order.
- Stepped-wedge sample size inputs
- Design effect, number of steps, clusters per step, people per cluster.
- Stratification / minimisation
- Measures used to limit chance baseline imbalance.
- Stratified graphs
- The same graph drawn separately per subgroup (e.g. by sex).
- Student's t-distribution
- Symmetric bell shape with heavier tails than the normal.
- Study size ↔ precision; systematic error ↔ accuracy
- Large studies give precise results, small studies imprecise ones. Few systematic errors give accurate results, many give inaccurate ones.
- "Success" terminology / q = 1 − p
- Success means the event of interest, even a bad one; q = probability of failure.
- Summaries for ordinal data
- Few categories: summarise like nominal data. Many categories: summarise like discrete quantitative data.
- Summarising multi-category variables
- Proportion in each category (nominal with more than 2 categories, and ordinal).
- Summary statistics
- A few numbers summarising a data set, usually location plus variability.
- Summary table ("Table 1")
- Groups as columns; n (%) for categorical and mean (SD) for quantitative variables.
- Surrogate analogy (football)
- Possession, shots on target and corners vs the match result.
- Surrogate as therapeutic target: two conditions (Prentice)
- Good agreement between surrogate and clinical outcome, AND a change in the surrogate produces an equivalent change in the clinical outcome.
- Surrogate outcome / endpoint / biomarker
- An indirect outcome used as a substitute to predict an important clinical outcome. Usually lab or radiology based.
- Surrogate outcome for sample size
- Used to show the intervention "has the potential to work".
- Survival curves by exposure
- Probability of remaining event-free over follow-up time, plotted per group.
- SUTVA
- A unit's treated outcome is the same whatever the assignment mechanism and independent of other units' treatment.
- Synergistic interaction
- Combined effect is greater than the sum of individual effects (18 > 15 pp).
- Systematic error (bias)
- Error from the way research is designed or performed, so that the difference between groups is estimated wrongly and one answer is favoured; a deviation from the truth. A CI is valid only if there is none.Also: Systematic error; Systematic error (bias) / internal validity
- Systematic measurement error
- Differs between exposed and unexposed groups. The errors do not sum to zero, so averages shift, estimates are biased and conclusions are wrong.
T
- T → normal as n grows
- The difference becomes negligible beyond about 30–40 observations.
- Table orientation
- Exposure on the rows, outcome on the columns.
- Tail areas / mechanism of the p-value
- The p-value is the area beyond the observed value, in both tails for a two-sided test, of the distribution centred on the hypothesised value.
- Target condition
- The disease the test diagnoses.
- Temporal trend (temporal sequence)
- Knowing that the outcome followed the exposure.
- Test-selection table
- Maps outcome type × normality × pairing × number of groups to a test (V7).
- The four accuracy × precision combinations
- Accurate+precise, inaccurate+precise, accurate+imprecise, inaccurate+imprecise.
- Timing
- Sample size is calculated at the design stage, before recruitment.
- Total-subject comparison across designs
- 2 separate trials = 600; 3-arm = 450; 2×2 factorial = 300.
- TP / FP / FN / TN
- Test-positive & diseased / test-positive & not diseased / test-negative & diseased / test-negative & not diseased.
- Transformation / back-transformation
- Convert non-normal data to normal, analyse, then convert results back to the original scale.
- True value
- The actual value of the quantity being measured. The bullseye in the target analogy.
- True value unmeasurable but replicates possible
- E.g. daily fat intake: destructive measurement, food not homogenised, so sampling the food introduces error.
- Two-armed parallel RCT
- One pool randomised into two concurrent groups, intervention and control.
- Two-sided vs one-sided p-value
- Two-sided when either direction matters (the default). One-sided "may be appropriate" when only one direction would change the decision.
- Two-way scatter plot
- Each participant plotted at (x, y) for two quantitative variables, optionally with a fitted line.
U
- Unconfoundedness assumption
- No confounding, required in observational studies.
- Units consistency
- All inputs must be in the same units.
- Univariate vs adjusted analysis
- Crude comparison vs an analysis that takes group differences (e.g. age) into account; results can differ.
- Univariate vs multivariate distribution
- One variable vs several variables with their relationships (multivariate not covered).
- Unpaired t-test (with Welch correction) / ANOVA
- Unpaired normal continuous outcome, 2 / 3+ groups. Welch if SDs are assumed different.
V
- Variable types (data types)
- Variables are quantitative (continuous or counted/discrete) or categorical (nominal, binary or ordinal); the type decides which summaries, graphs and tests to use.Also: Variable classification by data type; Data types; Data types recap
- Variance
- Square of the SD.
W
- Weak evidence near 0.05
- A p-value near 0.05 on its own is weak evidence against the null.
- When a hypothesis is needed
- Only when comparing two or more groups.
- Which effect to report
- Report ITT in all situations; per-protocol matters for toxicity/complications and when motivation is needed to continue.
- Why power calculations matter
- Observed difference and variability may differ from what was expected, so check that the false negative risk is still acceptable.
- Why risk is not computable in case-control
- Researcher sets the case:control ratio; source population and time period unknown.
- Wilcoxon signed-rank test / Friedman test
- Paired non-normal or ordinal-multi or count outcomes, 2 / 3+ groups. Signed-rank ≠ rank-sum.
Z
- Z-score (standard score)
- Z = (x − μ)/σ; use the sample mean and SD as estimates.
- Zero-cell correction (+0.5)
- Add 0.5 to every cell when a cell is zero, so ratios against that group are defined.
- Zone of equivalence
- E.g. ±2% when 2% is the clinically acceptable difference.