03 · What You Need to Know
When the No-Interaction Assumption Matters in Factorial Research
What Does No Interaction Actually Mean?
No interaction means that the effect of one factor does not change across levels of another factor on the specified effect scale.
Consider two binary interventions, A and B, with a continuous outcome.
On an additive scale, no interaction means that the difference associated with A is the same whether B is present or absent.
For example, if A improves performance by 10 points when B is absent, it also improves performance by 10 points when B is present.
This is a statement about treatment effects, not about whether the interventions are biologically, psychologically, or operationally unrelated.
Two interventions may act through different mechanisms and still exhibit statistical interaction. Conversely, interventions may share mechanisms without showing interaction on a particular outcome scale.
Does Every Factorial Design Require This Assumption?
No. A full factorial design systematically includes combinations of factor levels. Its allocation structure can support estimation of both main effects and interactions.
Indeed, an interaction-focused factorial experiment is deliberately designed to investigate whether effects differ across combinations.
The no-interaction assumption becomes particularly relevant when researchers simplify the analysis or justify efficiency by treating the interventions as producing effects that can be summarized independently.
Researchers should therefore distinguish the factorial arrangement itself from assumptions introduced by the chosen model, sample size calculation, or interpretation.
Factorial design structure
Specifies which combinations of factor levels are observed or experimentally assigned.
No-interaction assumption
States that the effect of one factor does not vary across levels of another on the selected effect scale.
The first concerns experimental allocation. The second concerns the relationship among treatment effects.
Why Do Efficiency-Oriented Factorial Trials Often Discuss No Interaction?
An efficiency-oriented factorial trial aims to evaluate multiple interventions using a coordinated experiment rather than conducting separate trials.
In a balanced 2×2 trial, the researcher may estimate the marginal effect of A by comparing participants assigned A with participants not assigned A, averaging across the two levels of B.
The corresponding comparison for B uses the same trial population.
This can reduce duplicated recruitment and data collection.
When interactions are absent, the marginal effect of A corresponds to its effect at either level of B on the specified additive scale. Interpretation is therefore straightforward.
When interactions exist, the marginal effect instead represents an average of different conditional effects. It may remain valid and useful, but it no longer describes a constant effect across treatment combinations.
The distinction is important when researchers use factorial efficiency to justify a sample size or make recommendations about implementing interventions separately.
Can Marginal Effects Be Estimated When Interactions Exist?
Yes. An interaction does not automatically invalidate a marginal effect.
Suppose AI feedback improves writing scores by 4 points without peer discussion and by 12 points with peer discussion.
In a balanced 2×2 design, the equally weighted marginal effect of AI feedback is 8 points.
That average is mathematically meaningful. However, it does not mean that AI feedback improves performance by 8 points in both settings.
Its practical relevance depends on whether the equal weighting of peer-discussion conditions represents the intervention policy or population of interest.
If a university plans to introduce AI feedback only in courses without peer discussion, the 4-point conditional effect may be more relevant than the 8-point marginal effect.
Consequently, the existence of interaction changes the interpretation and possible transportability of marginal effects, not necessarily their estimability.
How Is No Additive Interaction Expressed Mathematically?
For a continuous outcome, no additive interaction means that the difference between the effect of A when B is present and the effect of A when B is absent equals zero.
Does No Interaction Mean the Combined Intervention Has No Additional Benefit?
No. This is a consequential misunderstanding.
If A and B each improve an outcome and their effects are additive, the combined intervention can outperform either intervention alone without an interaction.
For example, suppose A improves a score by 10 points and B improves it by 8 points.
Under an additive model with no interaction, their combination improves the score by 18 points relative to neither intervention.
The combined intervention produces a larger total benefit, but it does not exceed the sum of the individual effects.
Interaction concerns departure from the specified combination rule, not whether combining interventions produces a better outcome than using only one.
Does the No-Interaction Assumption Depend on the Outcome Scale?
Yes. Interaction is scale-dependent.
For continuous outcomes analyzed using mean differences, researchers may evaluate additive interaction.
For binary outcomes, effects may be expressed as risk differences, risk ratios, or odds ratios.
Two interventions can exhibit no interaction on one scale while showing interaction on another.
For example, a constant risk ratio across levels of another intervention does not generally imply a constant risk difference.
Researchers should specify the effect measure before evaluating whether the no-interaction assumption is plausible.
It is not methodologically defensible to describe interventions as universally independent simply because an interaction term is absent from one statistical model.
What Happens When a Regression Model Omits the Interaction Term?
For two binary factors and a continuous outcome, an additive linear model without an interaction term can be written as:
In a balanced randomized 2×2 experiment with equal cell weights, an additive linear model may still recover marginal factor contrasts even when the true cell means exhibit interaction. This follows from the orthogonality of the balanced factorial allocation.
However, the fitted model may misrepresent combination-specific outcomes, and its conventional standard errors or interpretations require appropriate assumptions.
With unequal allocation, missing data, nonlinear models, or covariate adjustment, the relationship between model coefficients and desired marginal effects can become more complicated.
Researchers should not assume that omitting an interaction term is harmless in every analysis.
Does a Nonsignificant Interaction Test Justify Assuming No Interaction?
No. A nonsignificant test indicates that the data do not provide sufficient evidence to reject the specified null hypothesis at the chosen significance level.
It does not establish that the interaction is zero or substantively negligible.
Interaction estimates can be imprecise, particularly when a factorial trial was powered primarily for marginal intervention effects.
A confidence interval may include both zero and interactions large enough to change practical decisions.
In that situation, the findings remain compatible with meaningful effect modification despite the nonsignificant test.
Watch Out
Do not use a nonsignificant interaction test as proof that interventions operate independently. If the validity of a practical conclusion depends on negligible interaction, evaluate the estimate and its uncertainty against a substantively justified margin.
What Does Sufficiently Small Interaction Mean?
In some research settings, the assumption need not be that interaction is exactly zero. The relevant question may be whether any interaction is small enough that it would not materially alter the intended interpretation or decision.
For example, a university may consider a difference of one point in an interaction contrast negligible but regard a difference of eight points as educationally important.
Such thresholds should be justified before examining the results whenever possible.
A formal equivalence approach may be appropriate when the objective is to establish that interaction lies within prespecified bounds, but this requires suitable sample size planning and statistical analysis.
Simply observing a small point estimate is insufficient if the confidence interval remains wide.
How Does Interaction Affect Sample Size Planning?
Sample size requirements depend on whether the trial is intended to estimate marginal effects or interactions.
An efficiency-oriented factorial trial may be powered for the marginal comparison of A versus no A and B versus no B.
An interaction-focused trial instead needs sufficient information to estimate the difference between conditional treatment effects.
Interaction estimation often requires a larger sample because the contrast combines information from multiple treatment combinations.
Researchers should therefore distinguish the statistical requirements for testing intervention interactions from those for marginal treatment comparisons.
A trial with adequate power for two main effects may still have inadequate precision to rule out a practically important interaction.
What If Interactions Are Plausible Before the Trial?
Researchers should use substantive theory and prior evidence to identify whether one intervention could modify another's effects.
For example, AI feedback and peer discussion may influence overlapping learning processes. Their combined effect could be enhanced or diminished depending on how students use both forms of support.
If the interaction is central to the scientific question, the study should be designed and powered to estimate it.
If the primary objective remains marginal intervention evaluation, researchers should explain why those averages are meaningful and how important interaction uncertainty will be addressed.
The decision should not be postponed until the interaction test produces a p-value.
When Should Researchers Report Conditional Effects?
Conditional effects describe the effect of one intervention at a specified level of another.
They are particularly informative when interaction is substantial or when decisions concern specific intervention combinations.
For example, the effect of AI feedback in courses without peer discussion may be more relevant to an institution that does not provide structured peer activities.
Reporting conditional effects alongside marginal effects can prevent an average from concealing important differences between combinations.
However, exploratory subgroup comparisons should be interpreted cautiously, especially when the study was not designed to estimate them precisely.
What Should Be Reported About the No-Interaction Assumption?
Researchers should identify whether the primary analysis assumes no interaction, estimates marginal effects regardless of interaction, or explicitly models interaction.
They should report the treatment-combination outcomes, relevant interaction estimates, and uncertainty when these are important to the research question.
If the no-interaction assumption supports a sample size calculation or analytical simplification, its rationale should be documented.
The CONSORT and SPIRIT factorial trial extensions emphasize alignment among the trial objective, interaction assumptions, sample size, analysis, and reporting.
