Must the outcome be ordinal?

No. The label concerns ordering of the specified group comparison.

Can an ordinal interaction be practically important?

Yes. A changing magnitude can alter where an intervention is worthwhile.

Does a crossing outside the observed range prove a reversal?

No. It is an extrapolated implication of the fitted model and may be highly uncertain.

AI Glossary / Statistics and probability / Intermediate

Ordinal interaction

Definition

An ordinal interaction is an interaction in which a specified group difference changes in magnitude across levels of another variable without reversing direction over the range considered. One group remains above the other, but the gap is not constant on the chosen outcome scale. This contrasts with a crossover, or disordinal, pattern in which the group's ordering reverses.

What it means

Suppose one interface performs better than another on both easy and difficult tasks, but its advantage is much larger on difficult tasks. The ordering is consistent, yet task difficulty modifies the size of the interface difference.
Ordinal in this term describes the ordering of effects or group means. It does not mean that the outcome must be measured on an ordinal scale. The direction being compared and the range of the other variable should be stated explicitly.

How it works

Define the outcome scale and the comparison of interest. For two groups observed at two levels of another factor, calculate the group difference at each level. An interaction on the additive scale occurs when these differences are unequal.
If the differences have the same nonzero sign, the specified comparison has an ordinal interaction pattern. If the signs reverse, it has a disordinal pattern. A zero difference at a boundary needs careful description rather than an automatic strong claim about a reversal.
Plot estimated outcomes with uncertainty, but assess the interaction directly in an appropriate model. Nonparallel lines in a sample plot suggest an interaction; sampling variation can also produce apparent nonparallelism. Statistical evidence and practical importance are separate questions.
For a continuous moderator, inspect the relevant range. Two fitted lines may cross outside the observed data, but extrapolated crossing does not establish a reversal within the studied population. Uncertainty in the crossing location matters as well.
Interaction depends on the response scale. A difference can vary on the raw scale while a ratio stays constant. A logarithmic transformation can therefore change whether an additive interaction is present, even when the ordering of positive fitted values is preserved.
Report conditional effects rather than only an average across conditions. The average may still answer a useful question, but it can hide where the advantage is large or small. Causal language additionally requires a suitable design and identification assumptions.

Three practical examples

Interface evaluation: an AI assistant improves task scores in both novice and experienced users, with a larger estimated benefit among novices.
Service planning: a faster process reduces waiting time in both quiet and busy periods, but the reduction is greater when demand is high.
Learning research: two teaching methods retain the same ordering across task difficulties while their score difference changes.

Think of it this way

Think of two runners on separate tracks. One stays ahead at each checkpoint, but the gap widens and narrows. The changing gap is the interaction pattern; staying ahead is the ordinal part. The analogy does not establish why the gap changes or whether noisy measurements support the pattern.

Worked example

Consider original mean scores for two interfaces, A and B. On easy tasks, A scores 80 and B scores 70. On difficult tasks, A scores 65 and B scores 40.
The A-minus-B difference is ten points on easy tasks and 25 points on difficult tasks. The difference of differences is 25 - 10 = 15 points. Because both conditional differences are positive, this is an ordinal interaction pattern for the interface comparison on the score scale.
No standard error or significance test follows from these four means alone. Sample sizes, variability, and dependence between observations would be needed.
For a separate scale illustration, suppose B has positive outcomes 10 and 20 across two conditions, while A has 20 and 40. The raw differences are ten and 20, but the A-to-B ratio is two in both conditions. On a logarithmic scale, both differences are log(2), so the additive interaction disappears.
The examples show that effect direction, effect magnitude, and measurement scale should be reported separately.

Illustrative case study

Fictional case: Maple Training compares an AI tutor with a standard tutorial for simple and complex exercises. The estimated tutor advantage is positive in both categories but larger for complex exercises.
The researchers report the two conditional differences and assess their difference directly. They also check uncertainty and whether the scoring scale has ceiling effects on simple exercises. The company avoids describing the tutor as equally beneficial everywhere or treating the visual pattern alone as a proven interaction.

Common misconceptions and mistakes

An ordinal interaction is still an interaction. A consistent direction does not imply that the effect has the same size in every condition.
Do not infer an interaction because one subgroup comparison is statistically significant while another is not. The difference between subgroup effects needs its own assessment.
A graph's lines failing to cross is insufficient unless the lines are also nonparallel on the relevant scale. Perfectly parallel lines represent a constant additive difference, not an ordinal interaction.
Avoid extrapolating beyond observed moderator values. Also be explicit about which factor's conditional contrast retains its sign, especially in designs where different comparisons lead to different descriptions.

Business use

Businesses use this distinction when deciding whether an intervention has a broadly consistent direction but should be tailored by context. A positive average improvement may conceal small gains in one setting and much larger gains in another. Examine absolute and relative effects, uncertainty, and costs before choosing a uniform rollout or differentiated service. A statistical pattern alone does not explain its cause.

Check your understanding

What makes the first example ordinal?

Answer: A exceeds B in both task categories while the advantage changes from ten to 25 points.

Why did the second example lose its additive interaction after logging?

Answer: Its multiplicative ratio was constant, making the log-scale differences equal.

Related terms

Simple effect
Disordinal interaction
interaction
effect modification
moderator
crossover interaction
outcome scale
difference of differences

Learn next

calculate conditional differences and ratios from a two-factor table, then compare which interaction statements each scale supports.

Sources

Sources: David Howell, University of Vermont, Ordinal interaction: https://www.uvm.edu/~statdhtx/methods8/Glossary/Glossary.html . NIST/SEMATECH, Exploring First Order Interactions: https://www.itl.nist.gov/div898/handbook/ppc/section4/ppc423.htm . Widaman and colleagues, Distinguishing Ordinal and Disordinal Interactions: https://pmc.ncbi.nlm.nih.gov/articles/PMC3553243/ . The first two sources were read; the paper's search excerpt was checked but full text encountered a browser challenge. Examples are original.

Previous: Ordering constraints Next: Ordinal probit regression