AI Glossary / Statistics and probability / Advanced
Simple interaction
Definition
A simple interaction is the interaction between two factors evaluated at one specified level of a third factor. It asks whether the effect of one factor depends on another within a fixed context. In a three-factor design, for example, it can examine whether an AI tool's effect depends on task type among novice users. This entry uses that conditional factorial meaning; some terminology uses simple interaction more loosely for a two-way interaction without emphasizing a third factor.
What it means
An interaction concerns a difference in effects, rather than one effect considered alone. Holding a third factor fixed creates a simple interaction within that stratum. Comparing simple interactions across the third factor addresses a three-way interaction. These are distinct questions. Evidence of a two-way interaction in one stratum but not another does not automatically establish that the interactions themselves differ. Their direct comparison requires an appropriate contrast and uncertainty calculation.
How it works
Specify the two factors being examined and the level of the conditioning factor. Express the interaction as a contrast of cell means or predictions on the chosen outcome scale. For two binary factors, it is a difference of two differences. Estimate that contrast using a model that includes the relevant factorial structure and compute its uncertainty from the covariance of the estimated means. Preserve clustering or repeated measurements where present. If several strata or contrasts are examined, define the comparison family and address multiplicity or exploratory selection.
Three practical examples
Example 1: A study examines whether AI assistance has different effects on easy and difficult tasks among beginners. Example 2: Engineers test whether material type changes a temperature effect at one selected pressure level. Example 3: A service experiment examines the interaction of message style and delivery channel among customers in one prespecified group. Each fixes one context while investigating how two other factors combine. The result depends on the measurement scale and the validity of the study design.
Think of it this way
Think of opening one layer of a stack of comparison tables. Within that layer, compare the column differences across rows. That is the simple interaction. Open another layer and repeat the calculation; comparing the two layer-specific interactions asks a further question. This layered view prevents a conditional pattern, an overall average and a three-way difference from being described as though they were the same finding.
Worked example
Suppose factors A and B each have levels zero and one, and context C has two levels. At C = 0, cell means for (A,B) = (0,0),(1,0),(0,1),(1,1) are 10,14,12 and 20. The A effect is four when B = 0 and eight when B = 1, so the simple A-by-B interaction is 8 - 4 = 4. At C = 1, the corresponding means are 20,23,25 and 28. Both A effects are three, so that simple interaction is zero. The three-way contrast, C = 1 minus C = 0, is 0 - 4 = -4. An equal-weight average of the simple interactions is two, which answers a different marginal question. If four independent cell means each have estimated variance 9/16, the estimated standard error of one four-cell interaction contrast is square root of four times 9/16 = 1.5; other designs require their actual covariance structure.
Illustrative case study
In this fictional case study, Horizon Tasks studies AI assistance across task difficulty and user experience. Analysts prespecify the assistance-by-difficulty interaction among beginners and among experienced users. They report both contrasts and directly evaluate their difference, rather than comparing whether separate p-values cross a threshold. A plot shows cell means alongside uncertainty so stakeholders can see the underlying pattern. The team avoids presenting a complex interaction as a universal recommendation for all users or tasks.
Common misconceptions and mistakes
Common mistakes include calling a simple effect an interaction, interpreting significance in one stratum and non-significance in another as a proven three-way interaction, and overlooking the outcome scale. A difference-of-differences on a raw score scale need not imply the same interaction on a logarithmic or probability scale. Another error is fitting a model that omits necessary terms and then interpreting a coefficient as the desired conditional interaction. Sparse cells and many comparisons can make apparent patterns unstable.
Business use
Businesses use simple interactions when a product or intervention combines with another factor differently within a relevant operating context. AI studies may examine task and interface combinations separately by experience, device or language group. These analyses can guide follow-up experiments, but they require enough information in each cell and careful interpretation. Reports should show the contrast definition, conditioning level, scale and uncertainty. A clear graphical display of the underlying means often helps explain more than a table of isolated significance labels.
People also ask
Is a simple interaction the same as a three-way interaction?
No; the former fixes a third factor, while the latter compares how a two-way interaction changes across it.
Does an interaction imply that both main effects are large?
No; an interaction describes a pattern of conditional effects and can occur even when marginal effects are small.
Can a simple interaction be causal?
Only when the design and assumptions support the relevant causal comparisons; the mathematical label alone is insufficient.
Check your understanding
What is the simple interaction at C = 0 in the worked example?
Answer: Four units.
Why is it zero at C = 1?
Answer: The A effect is three at both B levels.
What is the three-way contrast with the stated direction?
Answer: Minus four.
Can separate p-values substitute for directly testing that three-way contrast?
Answer: No; their difference in significance does not measure the difference between interactions.
Related terms
- Simple effect
- Contrast
- Related terms include factorial design
- two-way interaction
- three-way interaction
- effect modification and marginal effect
Learn next
Learn next how factorial contrasts can be written as weighted sums of cell means. Explore how the covariance matrix determines their standard errors, then compare conditional and marginal interactions. Practice plotting cell means before interpreting higher-order coefficients, and plan scientifically motivated contrasts before searching many possible subgroups.
More in Statistics and probability
- Signal magnitude area - Signal magnitude area, commonly abbreviated SMA, is a feature that summarizes the absolute magnitude of a signal over a time window.
- Significance level - A significance level is a prespecified probability threshold used to calibrate a statistical hypothesis test, commonly denoted alpha.
- Simalto - SIMALTO is a market-research approach that uses a grid of product or service attributes and levels to explore customer priorities under trade-offs.
- Simple linear regression - Simple linear regression models the mean of one response variable as a straight-line function of one predictor.
- Simple regression - Simple regression is regression with a single explanatory variable.
- Simpson's paradox - Simpson's paradox is a statistical phenomenon in which an association observed within separate groups reverses when the groups are combined.
Sources
Sources: University of Vermont, Statistical Methods for Psychology glossary, Simple interaction, https://www.uvm.edu/~statdhtx/methods8/Glossary/Glossary.html; B. C. Dudek, Implementation of Three-factor Between-group Analysis of Variance, https://bcdudek.net/3wayanova/3way_factorial_Rmarkdown.html. The eight cell means, interaction calculations and fictional AI-task study are original illustrations with contrast directions and variance assumptions specified.
Previous: Simple function Next: Simple linear regression