Does first order mean only one variable is involved?

Not in this glossary; it refers to an interaction of two variables.

Must the lines cross?

No. Different slopes or differences can indicate interaction without a reversal.

Is the interaction coefficient always the difference-in-differences?

Only under the appropriate coding and model scale, such as the worked zero-one linear model.

AI Glossary / Statistics and probability / Intermediate

First order interaction

Definition

A first order interaction is an interaction between two variables: the relationship between one variable and an outcome depends on the level of the other. In the source glossary's terminology, it is also called a simple interaction. The more explicit phrase two-way interaction helps avoid ambiguity because conventions for numbering interaction order vary. An interaction is a feature of the specified statistical model and outcome scale. It does not by itself establish a causal mechanism or require that the direction of an effect reverses.

What it means

An additive model assumes that the effect associated with one input stays the same across levels of another. An interaction relaxes that assumption. A training method, for example, might have a larger association with performance under one task condition than another. The difference in those associations is the interaction. Both simple effects can be positive while differing in magnitude, so crossing lines are not necessary. Conversely, main effects averaged across conditions can conceal important variation in the conditional relationships.

How it works

In a linear model with two numerical predictors, an interaction can be represented by y = b0 + b1 x + b2 z + b3 xz + error. The slope for x is b1 + b3 z, making its dependence on z explicit. Categorical predictors require suitable coding and may involve several interaction coefficients when they have more than two levels. Estimate and interpret relevant contrasts with uncertainty, retaining appropriate lower-order terms. Plot predicted outcomes or simple effects over meaningful ranges. The interpretation changes with coding, centering and link function, so the product-term coefficient should not be read without the full model specification.

Three practical examples

Process experiments: examine whether changing temperature has different effects at different material settings.
Service evaluation: test whether a new workflow's effect varies between two prespecified task conditions.
Predictive modeling: allow the association between one feature and an outcome to depend on another feature while validating the added complexity.

Think of it this way

Think of a volume control whose effect depends on another setting. Moving the first control by the same amount may produce different changes in output depending on the second. The interaction describes that dependency, not merely the fact that both controls influence the output.

Worked example

Consider hypothetical mean outcomes for two binary factors x and z. The four means are 40 at (0,0), 50 at (1,0), 44 at (0,1) and 70 at (1,1). The x difference is 50 - 40 = ten when z is zero, but 70 - 44 = 26 when z is one. The difference-in-differences is 26 - 10 = 16. With zero-one coding, the saturated mean model is 40 + 10x + 4z + 16xz, reproducing all four values. The x effect is positive in both conditions, so no reversal is required. With minus-one/plus-one coding, the conventional factorial interaction effect is eight and its regression coefficient is four; these different numbers represent the same cell means under different scaling. Replication and an error model are still needed for statistical uncertainty.

Illustrative case study

Fictional case: Spruce Training compares two instruction formats across simple and complex tasks in a planned factorial study. Analysts find that the estimated format difference varies by task condition, so they report conditional contrasts rather than one universal benefit. They show uncertainty and keep the coding convention explicit. A manager's question about a third task characteristic becomes a separate higher-order interaction question. This invented case illustrates interpretation of a two-variable interaction and does not claim that observed differences are statistically established or causally explained without the design evidence.

Common misconceptions and mistakes

Do not identify interaction by saying one subgroup result is significant and another is not; directly assess the difference between their effects. Do not assume an interaction requires crossing lines or opposite signs. A product term's coefficient depends on units and coding. In nonlinear models, interaction on the link scale need not equal interaction on the outcome-probability scale. Omitting lower-order terms can impose unintended restrictions. A statistical interaction may reflect model misspecification, selection or confounding, so causal language requires an appropriate design and assumptions.

Business use

Businesses can use interactions to understand when an intervention or operating choice works differently across conditions. Prespecify important comparisons and ensure the study has enough information to estimate them, since interaction estimates can be imprecise. Present predicted outcomes and conditional effects in familiar units, with clear uncertainty. Avoid searching many combinations and reporting only the most striking pattern. An interaction that improves prediction may still be difficult to interpret causally, and a small statistical interaction may have little operational importance.

Check your understanding

In the worked example, why is the interaction 16 under zero-one coding?

Answer: The x difference changes from ten to 26 as z changes from zero to one, giving 26 - 10 = 16.

Related terms

Main effect
Simple effect
Factorial design
Disordinal interaction
two-way interaction
product term
effect modification

Learn next

recode a two-by-two example without changing its fitted means, then compare interaction interpretations on additive, logarithmic and logistic model scales.

Sources

Sources: University of Vermont, Statistical Methods for Psychology glossary, First order interaction: https://www.uvm.edu/~statdhtx/methods8/Glossary/Glossary.html ; NIST, Estimate Main and Interaction Effects, coding and coefficient scale: https://itl.nist.gov/div898/handbook/pri/section6/pri615.htm ; NIST, Effects plot, factorial effect conventions: https://www.itl.nist.gov/div898/handbook/pri/section5/pri597.htm ; NIST, One variable at a time, interaction interpretation: https://www.itl.nist.gov/div898/handbook/pri/section2/pri212.htm .

Previous: First normal form (1NF) Next: First-order logic