Education statistics guide
How to Interpret an Interaction in Two-Way ANOVA
An interaction means the effect of one factor changes depending on the level of another factor.
Updated:
Problem
When an interaction is significant, saying only "the teaching method worked" may be misleading because it may work only under certain conditions.
Who should use this
Use this when reading two-way ANOVA output, SPSS tables, or research result sections.
Formula and concept
Common interpretation order: interaction -> simple main effects -> main effects.
A main effect asks whether one factor differs overall. An interaction asks whether that difference depends on another factor. Significant interactions can hide or change main-effect interpretations. If you only have one categorical factor, this two-factor logic does not apply at all — use the one-way ANOVA guide instead. And if the interaction test itself is not significant, do not chase an interaction interpretation: report the two main effects on their own, because forcing an interaction narrative onto a non-significant result overstates the finding.
Simple main effects examine differences within a specific level of another factor, such as comparing two teaching methods separately for lower and upper grades. In the worked example below, the gap between methods is only 1 point in lower grades (78 vs. 77) but 8 points in upper grades (88 vs. 80) — that change in the size of the gap across grades, not just whether each individual gap happens to be significant, is what a significant interaction is describing.
This page helps with calculation and reporting, but users should still confirm their research design, assumptions, and statistical interpretation.
Step by step
- Check the interaction row first, such as method × grade.
- If significant, inspect a means table or plot.
- Run simple main effects or follow-up comparisons based on the research question.
- Then describe main effects carefully as supporting context.
Worked example
Method A and B are compared across lower and upper grades. In lower grades, A = 78 and B = 77. In upper grades, A = 88 and B = 80. If the interaction is F(1, 76) = 5.90, p = .018, the method effect changes by grade, so grade-specific comparisons are needed. As a quick internal check, the interaction is really about the change in the gap: the lower-grade gap is 78 − 77 = 1, the upper-grade gap is 88 − 80 = 8, so the gap widens by 7 points across grades. If you recompute this from a data set and the gaps turn out nearly identical across grades instead, that pattern argues against an interaction regardless of what a p value from a different run might suggest.
Common mistakes
- Interpreting only main effects after a significant interaction.
- Drawing conclusions without cell means or a plot.
- Calling two significant main effects an interaction — these are three separate questions (is factor A significant overall, is factor B significant overall, and does factor A’s effect change across levels of factor B), and a study can land on any combination of significant and non-significant results among them.
- Running many simple effects without a hypothesis or comparison control.
Editorial and reproducibility review:
Turn the guide into an auditable decision
This section reworks the guide’s own two-way example (a 1-point gap in lower grades widening to an 8-point gap in upper grades) and builds a counter-example where the gaps stay flat, to check that "interaction" means a changing gap, not just any gap.
Reproduce the example’s gap calculation, not just the p value
The example has Method A=78, B=77 in lower grades (gap = 1) and A=88, B=80 in upper grades (gap = 8). Compute the change in the gap directly: 8 − 1 = 7 points. The interaction test, F(1, 76) = 5.90, p = .018, is describing that 7-point widening — not the fact that a gap exists in either grade individually. Checking only whether each grade’s gap was significant on its own would answer a different question than the interaction test answers.
Build a case with flat gaps to see what "no interaction" looks like
Set up a counter-example: lower grades A=80, B=79 (gap = 1), upper grades A=90, B=89 (gap = 1). The gap is identical across grades — nothing widens or narrows — so this pattern would typically not produce a significant interaction, even though a method effect might still exist within each grade separately. Comparing this to the worked example (gap growing from 1 to 8) is the check for whether you are looking for "a gap" or for "a gap that changes," since only the second is what an interaction describes.
Self-check the guide’s point that three questions can combine in any pattern
The guide states that a significant method main effect, a significant grade main effect, and a significant interaction are three separate questions, and a study can land on any combination of significant and non-significant results among them. Self-test: knowing that both main effects are significant tells you nothing about whether the interaction is significant — you would need to check the interaction row itself, the same way the worked example’s F(1, 76) = 5.90 had to be read on its own rather than inferred from the main effects.
Completion criteria
- Computed the gap change in the worked example directly (1-point gap widening to 8 points, a 7-point change) rather than reading only the p value.
- Built a flat-gap counter-example (equal 1-point gaps in both grades) and can explain why that pattern would not typically produce a significant interaction.
- Can state that a significant main effect for one factor says nothing about whether the interaction is significant.
- Would check the interaction row’s own F and p value rather than inferring significance from either main effect.
Recommended tools
Related guides
FAQ
- Do I run simple main effects when interaction is not significant?
- Usually no, unless there is a clear planned reason.
- What if a main effect and interaction are both significant?
- Interpret the interaction first, then discuss the main effect cautiously.
- Do I need an interaction plot?
- Not always, but it helps reveal the direction and pattern. A quick alternative without plotting software is comparing the cell-mean gaps by hand, the way the lower- and upper-grade gaps are compared in the worked example above.
- Does SPSS automatically provide simple main effects?
- Not always. You may need EM Means or additional comparison settings.
Next step
Use the SPSS interpreter to organize interaction and main-effect results, then draft the APA wording.