Glossary

One line each

Epidemiology

Attributable fraction (exposed) — the share of disease among the exposed that is due to the exposure: (RR - 1) / RR.

Bias — a systematic error that does not shrink when you collect more data.

Case-control study — a study that starts from people with the disease and looks back at their exposure. Its usual output is an odds ratio. What that odds ratio estimates depends on how the controls were sampled.

Cohort study — a study that starts from exposure and follows people forward to see who develops the disease.

Confidence interval — the range of values compatible with the data at a stated level, usually 95%. It says nothing about whether the study was biased.

Confounding — a third variable that is associated with the exposure, causes the outcome, and is not on the path between them.

Cross-sectional study — a snapshot measuring exposure and outcome at the same moment. It cannot establish which came first.

Cumulative incidence — see Risk.

Differential misclassification — measurement error whose size depends on the other variable. Its direction is unpredictable.

Effect modification — see Interaction.

Healthy worker effect — employed people are healthier than the general population, so occupational exposures look safer than they are.

Incidence rate — new cases divided by person-time at risk. It has units and can exceed 1.

Information bias — error in how exposure or outcome was measured.

Interaction — the effect of the exposure genuinely differs between groups. It is a finding to report by stratum, not a nuisance to adjust away.

Mediator — a variable on the causal path from exposure to outcome. Adjusting for it removes part of the effect you are measuring.

Misclassification — putting a person in the wrong exposure or outcome category.

Non-differential misclassification — measurement error equally likely in both groups. For a yes/no exposure it pushes the estimate towards 1 on average. With more than two exposure categories it can push the other way, so do not treat it as a rule.

NNT, number needed to treat — how many people must be treated for one to benefit: 1 / abs(RD). Always report the time period with it.

NNH, number needed to harm — the same arithmetic when the exposure is harmful.

Odds — cases divided by non-cases: a / b. Not the same as risk.

Odds ratio — the cross product (a*d) / (b*cc). It approximates the risk ratio only when the outcome is uncommon.

Overadjustment — adjusting for a mediator, and so removing part of the real effect.

Person-time — the follow-up time each person actually contributed while still at risk, added up.

Population attributable fraction — the share of disease in the whole population that is due to the exposure. It depends on how common the exposure is.

Prevalence — the proportion of a population with the disease at a point in time. A snapshot, not a speed.

Randomised controlled trial — a study in which the investigator assigns the exposure at random. Randomisation balances confounders nobody thought of.

Recall bias — cases remember exposures differently from controls, because they have had reason to think about them.

Risk — new cases divided by the number at risk at the start, over a stated period. A proportion between 0 and 1.

Risk difference — risk_exposed - risk_unexposed. It measures impact.

Risk ratio — risk_exposed / risk_unexposed. It measures strength of association.

Selection bias — the people in the study differ from the people you meant to study, in a way related to both exposure and outcome. Ordinary adjustment does not fix it. Only a method that models how people were selected can, and only if you know enough about that selection.

Simpson’s paradox — a crude association that reverses when you stratify. Usually strong confounding, but a mediator or a collider can do the same thing, so find out which before you act on it.

Stratification — splitting the analysis by levels of a third variable, to see the association within each level.

R

Argument — a value you pass to a function, inside its brackets.

Assignment — <-, which puts a value into a name.

Console — the pane where you type R and get an answer at once.

CRAN — the archive R downloads packages from.

Data frame — R’s rectangle of data. A tibble is a data frame with better printing.

Function — a verb. It takes arguments and returns a value.

ggplot2 — the plotting package. A plot is data, an aes() mapping, and a geom_.

Logical — a TRUE/FALSE value. sum() of logicals counts; mean() gives a proportion.

NA — a missing value. It is contagious on purpose.

Native pipe |> — puts the value on its left into the first argument of the function on its right.

Object — anything you have given a name.

Package — a bundle of functions someone published. Install once, load every session.

Script — a file of R code you save and rerun. install.packages() does not belong in one.

Tibble — the tidyverse rectangle: one row per observation, one column per variable.

Tidyverse — the family of packages this course uses: dplyr, ggplot2, readr and others.

Vector — several values of the same type in a row, built with c().

Vectorisation — applying an operation to a whole vector at once, with no loop.

WebR — R compiled to run inside a web browser. It powers the 7 primers.

Working directory — the folder R treats as “here”. Find it with getwd().