Does three-way mean three outcome variables?

No; it usually refers to three interacting predictors for an outcome.

Does the sign have a universal meaning?

No. It depends on coding, comparison direction, and outcome scale.

AI Glossary / Statistics and probability / Intermediate

Higher order interaction

Definition

A higher order interaction is a statistical interaction involving three or more predictors or factors. A three-way interaction means that the interaction between two factors changes across levels of a third. The term describes a feature of a specified model and outcome scale; it does not by itself establish a causal mechanism or imply that every factor has a large average effect.

What it means

A two-way interaction says that the relationship between A and an outcome depends on B. A three-way interaction goes further: how B changes that relationship depends on C.
For example, the difference between two AI interfaces might depend on task difficulty, and that interface-by-difficulty pattern might itself differ between novice and experienced users. Reporting only one average interface difference could hide the relevant pattern.

How it works

Define the outcome scale and the factors clearly. Interactions are departures from a specified additive structure on that scale. An additive interaction on a raw-score scale is not the same claim as an interaction on a log or log-odds scale.
For three binary factors, compare the A difference at each B level within C = 0. Their difference is an A-by-B interaction contrast for that C level. Repeat within C = 1, then compare the two interaction contrasts. A nonzero difference is the three-way contrast.
A linear model can represent this through main effects, all pairwise products, and an A*B*C product term. A hierarchical specification retains the lower-order components whenever the higher-order term is included. This makes the model structure and conditional interpretations explicit.
Coding matters for numerical coefficients. With 0/1 coding, the three-way coefficient in a saturated binary-factor model equals the corresponding difference of differences of differences. With -1/+1 coding, that coefficient equals one eighth of that contrast. Neither convention changes the fitted cell means.
With more than two levels, an interaction can require several parameters. A three-factor interaction for factors with a, b, and c levels has (a - 1)(b - 1)(c - 1) degrees of freedom in a full factorial model with estimable terms. A single plotted difference may therefore represent only part of the interaction.
Use separate panels to display a two-way pattern across levels of the third factor. Show estimated cell means or contrasts with uncertainty, rather than expecting readers to interpret a product coefficient alone. For continuous predictors, choose meaningful values within the observed range.
Assess whether the design contains enough information. Sparse combinations, strong correlations, and fractional-factorial aliasing can make high-order terms hard to estimate or separate. Replication or an appropriate error model is needed for uncertainty; cell means alone do not establish statistical significance.
Keep exploration controlled. Searching many interactions and highlighting only the largest can overstate evidence. Prespecified questions, suitable multiplicity handling, and independent confirmation help distinguish a useful pattern from an unstable finding.

Three practical examples

An AI usability study examines interface-by-task interaction separately for novice and experienced participants.
A manufacturing experiment studies whether the interaction between temperature and material changes with machine speed.
A forecasting study asks whether the relationship between promotion and season differs across store formats.

Think of it this way

Think of a recipe where the benefit of extra mixing depends on oven temperature, and that dependency changes with the type of flour. A three-way interaction concerns the changing two-factor pattern, not merely three ingredients each having separate effects.

Worked example

Consider illustrative mean scores for binary factors A, B, and C. At C = 0 and B = 0, the means for A = 0 and A = 1 are 10 and 12, so the A difference is 2. At C = 0 and B = 1, they are 20 and 26, so the A difference is 6.
The A-by-B difference of differences at C = 0 is 6 - 2 = 4 points.
At C = 1 and B = 0, the means are 15 and 19, giving an A difference of 4. At C = 1 and B = 1, they are 25 and 37, giving an A difference of 12. The A-by-B contrast at C = 1 is 12 - 4 = 8 points.
The three-way contrast is therefore 8 - 4 = 4 points. The two-way interaction is stronger at C = 1 on this score scale. In a saturated model with all factors coded 0/1, the ABC coefficient is 4; with -1/+1 coding it is 4/8 = 0.5.
These stipulated means demonstrate interpretation and coding. Without sample sizes, variability, and design information, they do not supply a standard error, p-value, or causal conclusion.

Illustrative case study

Fictional case: Cedar Interface tests two assistant layouts on simple and complex tasks with novice and experienced users. The team plans the three-factor comparison before collecting results and retains all eight combinations.
Plots suggest that the layout-by-complexity pattern differs by experience. The analyst estimates the three-way contrast with uncertainty, checks whether repeated tasks from each user need a repeated-measures model, and reports the underlying cell means.
The product team uses the result to form a targeted design hypothesis. It does not conclude that a layout works for everyone or remove lower-order model terms solely because their individual p-values are large. A fresh study checks whether the pattern persists.

Common misconceptions and mistakes

Do not call any model containing three predictors a three-way interaction. Separate main effects are insufficient. Also avoid claiming that significance in one subgroup and nonsignificance in another proves their interactions differ; test the relevant contrast directly.
A nonsignificant higher-order term does not prove exact absence. Its uncertainty may be large. Strong observed patterns can also be scale dependent or confounded by the study design.

Business use

Higher order interactions can identify operating combinations that average results conceal. They are useful when a decision depends on context, but their complexity increases data requirements and communication effort. Reports should focus on the combinations that matter operationally and the uncertainty in those comparisons.

Check your understanding

What is the A-by-B contrast at C = 0?

Answer: Four points.

What is it at C = 1?

Answer: Eight points.

What additional information is needed for inference?

Answer: Sampling variability, sample sizes, and the study's dependence and assignment structure.

Related terms

Factorial design
Simple effect
Contrast
Aliasing (factorial experiments)
Interaction
effect modification
model hierarchy
aliasing
three-way interaction
outcome scale

Learn next

Calculate two-way contrasts within separate strata before forming a three-way contrast. Then compare coding conventions and practice presenting estimated cell means with uncertainty across small, clearly labeled panels.

Contrast

Sources

David Howell, glossary: https://www.uvm.edu/~statdhtx/methods8/Glossary/Glossary.html UC Davis, Treatment Structures and Comparisons: https://psfaculty.plantsciences.ucdavis.edu/PLS120/PLS120Book/docs/chTrtstr.html NIST, effects plots: https://www.itl.nist.gov/div898/handbook/pri/section5/pri597.htm

Previous: High-throughput computing Next: Higher-order logic