Interaction (statistics)

A model relation in which the association or effect of one predictor on an outcome changes with the level of another predictor.

Version

v2 · 2026-09-19 · History

Domain-specific #

5062

Domain group

Formal Sciences

Origin domain

Experimental Design & Statistics

Subdomain

Statistical Modeling → Experimental Design & Statistics

Aliases

Moderation

Core Idea

Statistical interaction is scale- and parameterization-dependent, it need not imply biological or causal interaction, main-effect coefficients become conditional and absence on an additive scale does not imply absence on multiplicative scales. A product term or stratified response surface lets the slope for one variable vary with another; contrasting predicted outcomes tests departure from the chosen additive or multiplicative no-interaction model. The abstraction is therefore identified by a declared carrier, a transformation or constraint over that carrier, and an invariant that tells an analyst whether the named structure is genuinely present.

How would you explain it like I'm…

 

No faithful explanation at this level. A five-year-old version collapses into 'two things team up to make something happen,' teaching that interaction is real causal teamwork, when it is only a departure from a no-interaction model on a chosen scale.

It Depends on the Other

An interaction in statistics means the effect of one factor on an outcome changes depending on the level of another factor. Without an interaction, each factor just adds its own fixed amount. With one, you can't describe one factor's effect without saying what the other factor is. Whether you find an interaction depends on how you measure: something that looks like no interaction when you add amounts may show one when you compare by multiplying, like doubling. And a statistical interaction doesn't automatically mean the two things truly work together in real life.

Scale-Dependent Effect Modification

In statistics, an interaction means the effect of one predictor on an outcome depends on the level of another predictor in your model. It's usually modeled with a product term, such as x × z, which lets the slope for x change with z, or with a fitted surface for different groups. Testing for interaction checks whether the data depart from a 'no interaction' model, meaning the effects simply add, or simply multiply, depending on the scale you chose. Interaction depends on scale: effects that add perfectly might show interaction on a multiplicative scale, and the reverse is also true. Once an interaction term is included, the 'main effect' of x describes its effect only at a reference value of z. Interaction is also different from confounding, mediation, and plain correlation, and it doesn't prove a real causal or biological interaction.

 

Interaction in statistics is a departure from a declared no-interaction model—additive or multiplicative—for how two or more predictors jointly relate to an outcome. It is typically represented by a product term or a stratified or nonparametric response surface that lets the conditional slope of one variable vary with another, and it is assessed by contrasting predicted outcomes against the additive null on the chosen scale. The concept is scale- and parameterization-dependent: interaction can be absent on an additive (risk-difference) scale yet present on a multiplicative (ratio, log-link) scale, and vice versa. Including an interaction term makes the main-effect coefficients conditional on the reference levels and coding of the other variables, which is why hierarchy or marginality principles matter. Making a claim properly requires stating the outcome, predictors, model family and link, coding, conditional marginal effects and contrasts, and estimation uncertainty, usually with visualization. Statistical interaction need not imply biological or causal interaction; claiming effect modification in a causal sense requires explicit causal assumptions. It must also be distinguished from confounding, mediation, and mere correlation between predictors.

Scope of Application

Interaction (statistics) belongs to statistical modeling and is useful where the analyst can specify the typed statistical modeling carrier, including objects, relations, parameters, conventions, evidence, boundaries, and comparison targets, then evaluate the outcome and two or more predictors, model family and link scale, main effects and interaction term or nonparametric surface, reference levels and coding, conditional marginal effect and contrast, null of additivity on the chosen scale, estimation uncertainty and visualization, causal assumptions when effect modification is claimed, hierarchy or marginality and distinction from confounding mediation and simple correlation are explicit.

Clarity

The abstraction clarifies a crowded vocabulary by making the outcome and two or more predictors, model family and link scale, main effects and interaction term or nonparametric surface, reference levels and coding, conditional marginal effect and contrast, null of additivity on the chosen scale, estimation uncertainty and visualization, causal assumptions when effect modification is claimed, hierarchy or marginality and distinction from confounding mediation and simple correlation are explicit the center of the account.

Manages Complexity

Without the abstraction, an analyst must reason directly over many local details: the carrier roles, admissibility assumptions, competing conventions, derived invariants, boundary cases, and proof or validation obligations specific to Interaction (statistics). Interaction (statistics) compresses them into the roles in the structural signature. That compression permits comparison across instances without erasing the variables that determine validity. It also exposes which details may be varied safely and which are constitutive.

Abstract Reasoning

1. Identify the carrier. State what the elements, states, objects, or observations are: the typed statistical modeling carrier, including objects, relations, parameters, conventions, evidence, boundaries, and comparison targets. Reject examples whose alleged carrier belongs to a different problem.
2. Lock the constitutive rule. Express the outcome and two or more predictors, model family and link scale, main effects and interaction term or nonparametric surface, reference levels and coding, conditional marginal effect and contrast, null of additivity on the chosen scale, estimation uncertainty and visualization, causal assumptions when effect modification is claimed, hierarchy or marginality and distinction from confounding mediation and simple correlation are explicit independently of one notation or implementation.

Knowledge Transfer

Knowledge transfers strongly among subfields of statistical modeling because they reuse the typed statistical modeling carrier, including objects, relations, parameters, conventions, evidence, boundaries, and comparison targets, A product term or stratified response surface lets the slope for one variable vary with another; contrasting predicted outcomes tests departure from the chosen additive or multiplicative no-interaction model., and type the carrier, state every parameter and convention in the definition, test that the outcome and two or more predictors, model family and link scale, main effects and interaction term or nonparametric surface, reference levels and coding, conditional marginal effect and contrast, null of additivity on the chosen scale, estimation uncertainty and visualization, causal assumptions when effect modification is claimed, hierarchy or marginality and distinction from confounding mediation and simple correlation are explicit, compare the nearest accepted identity, and report counterexamples, uncertainty, and limiting cases.

Relationships to Other Abstractions

Current abstraction Interaction (statistics) Domain-specific

Parents (1) — outgoing typed relations

• Interaction (statistics) is a kind of Relation Prime

  It remains its own entry because its identity is fixed by the outcome and two or more predictors, model family and link scale, main effects and interaction term or nonparametric surface, reference levels and coding, conditional marginal effect and contrast, null of additivity on the chosen scale, estimation uncertainty and visualization, causal assumptions when effect modification is claimed, hierarchy or marginality and distinction from confounding mediation and simple correlation.

Hierarchy path (1) — routes to 1 parentless root

• Interaction (statistics) → is a kind of Relation

Neighborhood in Abstraction Space

Interaction (statistics) sits in a crowded region of the domain-specific corpus (11^th percentile for distinctiveness): several abstractions share nearly its structure, so a description that fits it tends to fit its neighbors too.

Family — Dispersion Measures & Multilevel Regression (11 abstractions)

Nearest neighbors

• Deviance (statistics) — 0.92
• Principle of marginality — 0.92
• Standard error — 0.91
• Hierarchical generalized linear model — 0.91
• Studentization — 0.91

Computed from structural-signature embeddings · 2026-10-25