Glossary of terms

Number of observations or units included in the calculation. For a paired test, n usually means the number of complete pairs. Missing and excluded values are not counted.

A missing value is absent, not zero. An exclusion keeps the value in the table but leaves it out of the calculation.

The sum of included values divided by their number. Strongly affected by extreme values.

The middle of the ordered values: half are no greater, half no smaller. Usually less affected by extreme values than the mean.

Sample SD: √[Σ(xᵢ − x̄)² / (n − 1)]. It measures spread around the mean in the same units as the data.

Q1 and Q3 are the 25th and 75th percentiles. The interquartile range between them covers the middle half of the distribution.

SD divided by the mean, as a percentage. Used for positive quantities on a ratio scale with a meaningful zero.

SD / √n. For independent observations, it estimates the standard error of the mean, not the spread among individual values.

A 95% CI is built by a procedure that covers the true value in 95% of repeated samples when assumptions hold.

If the null hypothesis and test assumptions hold, this is the probability of a test statistic at least as extreme as the observed one. The test defines extremeness. It is not the probability of the hypothesis.

A threshold for a p-based decision, chosen in advance. If the null hypothesis is true and assumptions hold, a calibrated test rejects it falsely in no more than an α proportion of repetitions.

A number computed from data by a test, such as t, F, χ² or U. The p-value comes from this statistic's null distribution; the test determines how extremeness is measured.

A measure of the magnitude and direction of a difference or association, expressed in original units or on a standardized scale. It describes how large the effect is, not whether it matters in practice.

For a specified model, design, α and alternative, the probability that the test rejects H₀. It depends on effect size, variability and sample size. It is not the probability that a hypothesis is true.

Standardized mean differences for different designs. Hedges g divides the mean difference by pooled SD and applies a small-sample correction, matching the Student test. Bonett δ divides by the square root of the average group variance and does not assume equal variances, matching the Welch test. Paired Cohen d_z divides the mean paired difference by the SD of paired differences.

In one-way ANOVA, η² is the share of total sums of squares associated with the factor. ω² is usually a less biased estimate of the population share and Linza truncates it at zero. η²p uses effect and error sums of squares, so it depends on the model. ε² is an effect estimate for Kruskal-Wallis.

For independent groups, this is P(A > B) - P(B > A); ties contribute equally in both directions. Linza shows it for Mann-Whitney and for each pair after Kruskal-Wallis. The Dunn adjusted p-value uses pooled ranks, while this pairwise effect is calculated from the two groups' observations. For matched Wilcoxon, it is (W+ - W-) / (W+ + W-), using ranks of nonzero paired differences. Positive values favor the first group or condition. Zero means balanced ranks, not proof that there is no effect or that the distributions are equal.

For many parametric tests, normality concerns model errors within groups; for a paired t-test, it concerns pair differences. Shapiro-Wilk is a diagnostic, not proof of normality, and can have low power with small n.

The test assesses whether group spreads are compatible with a common variance. A non-significant p-value does not prove equal variances or confirm other model assumptions.

A p-value recalculated by the selected procedure for a prespecified family of hypotheses. Compare it with α to control the error rate that the procedure targets.

Pairwise or control-versus-group comparisons with a multiplicity adjustment. Choose a procedure for the hypothesis family; comparisons may be planned or post hoc. Linza provides Tukey, Dunnett, Bonferroni, Sidak and Dunn procedures.

Degrees of freedom parameterize the reference t, F or χ² distribution used for inference. For a paired t-test, df = n − 1, where n is the number of complete pairs; Welch degrees of freedom are approximate.

Measurements linked to the same unit or matched by design, for example before and after. A paired test analyses differences within complete pairs rather than treating values as independent.

In ANOVA, the ratio of the effect mean square to the error mean square. A larger F is less compatible with the null hypothesis of no effect; interpretation depends on the design and test assumptions.

The expected change in Y for a one-unit increase in X in linear regression, holding other predictors constant. Its sign gives the direction of the association.

In linear regression with an intercept, R² is the share of Y variability explained by the model relative to the mean. A high R² alone does not show that the model is correct.

The difference between observed Y and the value predicted by the model. Residual plots can reveal nonlinearity, changing spread and other model problems.

On a fitted curve, the dose or concentration at which the response is halfway between the estimated lower and upper plateaus, not necessarily 50%. IC50 is the corresponding level for inhibition.

The steepness parameter of an S-shaped dose-response curve. The larger its magnitude, the sharper the transition.

The odds of an event in one group divided by the odds in another. OR = 1 means equal odds for this comparison; it says nothing about other groups or outcomes.

Observation ended without the event: we only know the event had not happened by that time.

The number of units still under observation immediately before a given time, for which the event has not yet occurred.

A test comparing whole time-to-event curves. It does not say which pair of groups differs.

When the effect of one factor depends on the level of another.

The repeated ANOVA assumption of equal variances of differences between conditions. When violated, the Greenhouse-Geisser correction is used.

A unit-specific shift in a model's intercept. For repeated data, it accounts for dependence among measurements from the same unit.

Estimation methods for a mixed model. REML estimates variance components using a restricted likelihood; ML uses the ordinary likelihood. Compare models with different fixed effects using ML because their REML likelihoods are not comparable.