Interaction Effect

An interaction effect happens when the effect of one variable on an outcome depends on the level of another variable. Put plainly, it depends. A discount might lift sales strongly for new customers and barely move loyal ones, so the discount's effect interacts with customer type. When there is no interaction, a variable's effect is the same regardless of the others, and a discount adds the same lift to everyone. When there is an interaction, you cannot state one variable's effect without naming the setting of the other. Statisticians also call this moderation, because the second variable moderates, meaning strengthens, weakens, or reverses, the first variable's influence. Interactions are how you capture the reality that the same lever produces different results for different segments, channels, or conditions, rather than one tidy average for all.

You model an interaction by adding a product term to a regression, the two variables multiplied together. If that term's coefficient is meaningfully different from zero, the effect of each variable shifts with the other. Interactions come in flavors. A synergistic interaction means two levers together beat the sum of their parts, as when a great creative and a well-targeted audience each help, and together help more than expected. An antagonistic one means they undercut each other. The sharpest kind is a crossover, where an effect that is positive in one group is negative in another, so a bold, edgy ad that delights younger buyers may repel older ones, and the average effect hides two opposite truths. Interactions are where averages stop being enough to guide a decision.

A main effect is a variable's average influence on the outcome, ignoring the others, such as the overall lift from a discount pooled across all customers. An interaction effect is how that influence changes across levels of another variable. The two describe different things, and reading one for the other is a common, costly mistake. Suppose a promotion has a main effect of plus five percent on average. If it also interacts with customer type, plus fifteen for new buyers and near zero for loyal ones, the flat average is technically correct and practically misleading. Acting on the main effect alone, you would roll the promotion out to everyone. Acting on the interaction, you would aim it at new buyers, where the effect actually lives. Main effects summarize, while interactions reveal for whom and when a lever really works.

The danger is that a strong interaction can mask or distort main effects entirely. In a crossover interaction, a lever helps one group and hurts another by similar amounts, so the main effect nets to roughly zero, and a report that shows only the main effect concludes no effect when the truth is two large, opposite effects. This is why segmenting and testing for interactions matters, because an average that pools opposite responses is worse than useless, since it looks like knowledge. Multiple regression estimates main effects for each predictor, and adding interaction terms lets it estimate how those effects bend across conditions. The main effect answers what does this do on average. The interaction answers for whom, and where, does it do something different, which is often the more useful question.

To find interactions, you have to look for them on purpose. Add interaction terms to a regression, or run experiments that cross the variables you suspect combine, different creatives against different audiences, say, so you can read each cell, not just the margins. Plot the results, because interaction shows up as lines that are not parallel, and a crossover shows up as lines that cross. Interpret an interaction by describing the effect of one variable at specific levels of the other, the discount lifts new customers by fifteen points and loyal ones by two, never as a single number. And guard against false positives, because hunting through many possible interactions will turn up some by chance, so pre-specify the ones you care about or hold them to a higher bar of evidence and replication before you believe them.

The failures cluster around two errors. First, ignoring interactions and governing by averages, which buries the fact that a tactic helps one segment and harms another, the crossover that a main effect erases. Second, over-hunting interactions, because if you slice the data enough ways, spurious ones appear, tempting teams to build strategy on noise that will not repeat. Between these sits the discipline. Suspect interactions where experience says response differs, new versus loyal customers, high versus low intent, one channel versus another, test those deliberately, describe the effect at each level rather than collapsing it, and confirm the important ones before betting on them. Interactions are often where the real strategy hides, but only the ones that survive honest testing are worth acting on.

The term comes from analysis of variance in statistics, where an interaction names the part of an outcome that the combination of two factors explains beyond their separate main effects.

Etymology: source.

Search interest for this term over the last five years:

View interest-over-time on Google Trends →

Areas of marketing where interaction effect is a core concern: