An interaction term is a variable in a statistical model that captures how two factors work together to influence an outcome, beyond what either factor does on its own. It’s created by multiplying two predictor variables together and adding that product to a regression equation or similar model. If a pain medication works differently in men than in women, for example, the interaction term between drug dose and sex is what detects and quantifies that difference.
The Core Idea: Effects That Depend on Context
Most statistical models start with an assumption that each variable has its own independent effect on the outcome. Drug dose reduces pain by a certain amount. Being male or female shifts the baseline pain level. These individual contributions are called main effects, and they add up in a straightforward way.
An interaction term asks a different question: does the effect of one variable change depending on the value of another? If a higher dose reduces pain more dramatically in males than in females, the two variables aren’t just adding their effects together. They’re combining in a way that produces something beyond the sum of their parts. That “something beyond” is the interaction. As one useful framing puts it, an interaction evaluates whether and how the relationship between an exposure and an outcome is modified by the value of a third variable.
A simple way to think about it: if the differences aren’t different, there’s no interaction. If taking the drug lowers pain by 10 points regardless of sex, drug dose and sex don’t interact. If the drug lowers pain by 10 points in males but only 3 points in females, they do.
How It Works in a Model
In a regression equation, you start with something like this: outcome = baseline + (effect of X) + (effect of Z). To test for an interaction, you add a third term that’s the product of X and Z: outcome = baseline + (effect of X) + (effect of Z) + (effect of X × Z). That product term is the interaction term. Its coefficient tells you how much the effect of X changes for each unit increase in Z, or equivalently, how much the effect of Z changes for each unit increase in X.
Here’s a subtlety that trips people up: once you add an interaction term, the individual coefficients for X and Z no longer mean what they meant before. In a model without the interaction, the coefficient for X represents its average effect across all values of Z. In a model with the interaction, the coefficient for X represents its effect only when Z equals zero. If Z never actually equals zero in your data, or if zero is an extreme or meaningless value, that coefficient can look strange or even misleading. Researchers sometimes misinterpret these coefficients as if they still represent average effects, which they don’t.
A Concrete Example With Categories
Imagine you’re studying whether a treatment affects men and women differently. You have two categorical variables: treatment (yes or no) and gender (male or female). Without an interaction term, your model assumes the treatment effect is the same for both genders.
Add the interaction term, and each combination gets its own predicted outcome. Say the baseline (male, no treatment) is 50. The coefficient for “female” might be +3, meaning untreated women average 53. The coefficient for “treatment” might be +5, meaning treated men average 55. The interaction coefficient might be -10, meaning when you combine female and treatment, you add all the coefficients together: 50 + 3 + 5 + (-10) = 48. The treatment increases the outcome by 5 points for males but decreases it by 5 points for females. The effect of treatment completely flips depending on gender.
When you plot this, the lines for males and females cross each other, which is sometimes called an antagonistic or interference interaction. The steeper the crossing angle, the stronger the interaction.
Ordinal vs. Disordinal Interactions
Interactions come in two visual patterns. In an ordinal interaction, one group always scores higher than the other across all levels of the second variable, but the gap between them widens or narrows. The lines on a plot fan apart or converge but never cross within the range of your data. In a disordinal (or crossover) interaction, the lines actually cross: one group scores higher at low values of the second variable but lower at high values. Disordinal interactions tend to be more dramatic and easier to detect because the relationship literally reverses direction.
Synergistic and Antagonistic Interactions
In biology and medicine, interactions are often described as synergistic or antagonistic. These terms are anchored to an additive model, which predicts what should happen if two factors simply stack their individual effects. If factor A reduces a symptom by 10 points and factor B reduces it by 15, an additive model predicts a combined reduction of 25.
A synergistic interaction means the combined effect is greater than that additive prediction. Maybe the two factors together reduce the symptom by 40 points. An antagonistic interaction means the combined effect is less than predicted. Maybe together they only reduce it by 12 points, as if one partially cancels the other out. In statistical terms, a significant interaction term in the model is what tells you the combined effect deviates from simple addition.
When Main Effects and Interactions Coexist
It’s entirely possible for main effects and interactions to exist simultaneously. Each factor can independently account for variability in the outcome while also combining with the other factor to produce an additional effect. A common example: a higher drug dose is more effective than a lower dose for both sexes (main effect of dose), but the difference between doses is much more pronounced in males than in females (interaction between dose and sex). Both the main effect and the interaction are real and meaningful.
When a significant interaction is present, interpreting the main effects in isolation can be misleading. Saying “the higher dose is more effective” is technically true on average, but it hides the fact that the dose effect is five times larger in one group than the other. The interaction is where the interesting story lives.
The Hierarchy Principle
A widely followed rule in model building is that if you include an interaction term, you should also include the main effects for both variables that make up that interaction. You wouldn’t include the X × Z term without also including X and Z on their own. Dropping a main effect while keeping the interaction can distort the model’s estimates and make the interaction coefficient uninterpretable.
Mean Centering: Helpful but Not a Universal Fix
Because an interaction term is the product of two variables, it tends to be highly correlated with those variables on their own. This correlation (called multicollinearity) can inflate the uncertainty around your estimates. A common recommendation is to mean-center your variables before creating the interaction term: subtract each variable’s average so the values are centered around zero. This can make the individual coefficients easier to interpret, since “zero” now means “average” rather than some extreme or meaningless value.
However, centering doesn’t always reduce multicollinearity the way textbooks suggest. Research in Educational and Psychological Measurement found that when data points cluster in groups rather than spreading symmetrically, centering can actually increase the multicollinearity problem. The practical advice: plot your data first. If the two predictor variables are spread roughly symmetrically, centering helps. If the data appear in distinct clusters, centering may not improve anything.
Why Interaction Terms Matter
Without interaction terms, statistical models assume every factor operates independently, like ingredients that contribute to a recipe without affecting each other. In reality, factors modify each other constantly. Exercise might lower blood pressure more in older adults than younger ones. A fertilizer might boost crop yield in wet soil but harm it in dry soil. A teaching method might work well in small classes but backfire in large ones. Interaction terms are how models capture these conditional relationships instead of forcing everything into a simple additive framework.
If you’re reading a study that reports only main effects without testing for interactions, it may be telling an incomplete story. The averages might look unremarkable while dramatic differences hide within subgroups. Interaction terms are the tool that pulls those differences into view.