A factorial trial randomizes participants with respect to more than one factor. Each factor contains an intervention and its comparator or, in more complex designs, multiple levels. In a full 2×2 trial, two factors with two levels each produce four allocated treatment cells.[1][2]
The four cells are not automatically the four main comparisons. Many factorial trials combine cells to estimate the main effect of each factor efficiently. That analysis depends on a clearly defined treatment effect and assumptions about how the factors relate. Understanding the cells, comparisons, and interaction is therefore more informative than seeing “2×2” in the methods section.
Map the four cells before reading any result
Suppose factor A compares A with not-A, and factor B compares B with not-B. A full 2×2 design contains:
| B | Not-B | |
|---|---|---|
| A | A plus B | A without B |
| Not-A | B without A | Neither A nor B |
Each row-column combination is a randomized group or cell. The trial may have two main factor comparisons:
- A main comparison: everyone allocated to A versus everyone allocated to not-A, combining across B assignment.
- B main comparison: everyone allocated to B versus everyone allocated to not-B, combining across A assignment.
This is often called a factorial or “at-the-margins” analysis. A cell-by-cell or multi-arm analysis instead compares specific combinations inside the table. The methods paper accompanying the factorial extensions emphasizes that the target effect and main comparisons must be stated because the analysis groups may not be identical to the conditions in which the effect is intended to be interpreted.[1]
Factor, level, cell, and comparison are different
Confusing these terms leads to exaggerated claims.
| Term | Meaning in a factorial trial |
|---|---|
| Factor | One randomized intervention question together with its comparator levels |
| Level | A specific option within a factor, such as active versus control |
| Cell or treatment group | A unique combination of levels across all factors |
| Main comparison | The groups combined or contrasted to answer a prespecified objective |
| Interaction | Dependence of one factor’s effect on the level of another factor, on the chosen analysis scale |
A four-cell trial does not necessarily provide four equally powered pairwise answers. Its sample size may have been chosen for two marginal comparisons, not for each cell against every other cell. Extract the prespecified objective, comparison, effect measure, and analysis population before interpreting a cell estimate.
Full, partial, and fractional factorial designs are not interchangeable
In a full factorial design, every combination of factor levels is represented and every enrolled participant is eligible for randomization on every factor. In a partial factorial design, some participants are not randomized for particular factors. This can occur because eligibility differs across factors or because a factor opens or closes at a different time. In a fractional factorial design, some combinations of levels are omitted.[1]
Partial designs can improve feasibility. A participant ineligible for factor B may still contribute to the randomized comparison for factor A. That person must not be placed into the factor-B comparison, because there was no factor-B randomization for that participant. Factor-specific denominators and flow are essential.
Different eligibility or enrollment periods can also change which population contributes to each factor. The A comparison may include more people, different calendar time, or different baseline risk than the B comparison. The shared protocol does not make those evidence sets identical.
A 2-in-1 design usually targets efficient main effects
A common reason to use a factorial design is to evaluate two interventions within one trial without materially increasing the sample size that separate two-arm trials would require. The 2025 explanation-and-elaboration paper calls this a “2-in-1” objective.[1][2] It commonly uses the marginal comparisons because every eligible randomized participant contributes information to both factor questions.
The efficiency is not free. It generally relies on an assumption that the interventions do not interact in a way that makes the marginal estimates misleading for the intended treatment effects. Trialists may choose factors thought unlikely to interact, but an interaction usually cannot be ruled out in advance.
A trial designed primarily to detect interaction has a different objective. Detecting interaction generally requires a larger sample than estimating main effects. The factorial guidance explains that an efficiency-focused trial and an interaction-focused trial usually require different design and sample-size choices.[1] Readers should not expect an efficient main-effect trial to give a precise interaction answer it was not sized to provide.
Interaction belongs to an analysis scale
An interaction means that the effect of one factor depends on the other factor’s level, relative to the relationship assumed on the chosen analysis scale. A difference-scale interaction, ratio-scale interaction, and model-term interaction need not have the same appearance.
This is why interaction cannot be diagnosed by checking whether a result is statistically significant in one subgroup and nonsignificant in another. Separate p values do not test the difference between the effects. A direct interaction estimate, its direction, and its uncertainty are needed.
For example, an estimated effect for A among people assigned B and an estimated effect for A among people assigned not-B may differ numerically. The question is how large that difference is on the prespecified scale and how uncertain it remains. A wide confidence interval may allow interactions that are negligible, helpful, or harmful.
A large interaction p value is not proof of no interaction
Many 2-in-1 trials are powered for their factor main effects, not for interaction. In that setting, an interaction test can be imprecise. Failure to reject a no-interaction hypothesis means the data did not provide strong evidence against it at the selected threshold; it does not establish that the factors do not interact.
Interpretation should consider:
- the prespecified interaction scale and rationale;
- the interaction estimate and confidence interval;
- the plausible clinical importance of effects allowed by that interval;
- cell-specific outcomes and denominators;
- whether enrollment was concurrent for both factors; and
- sensitivity analyses using a different comparison structure.
The factorial methods paper notes that a cell-based multi-arm analysis avoids the no-interaction assumption but usually loses precision compared with the factorial analysis. It can therefore serve as a sensitivity analysis, not an automatic replacement for the prespecified main analysis.[1]
Sample size and multiplicity follow the objectives
Sample-size planning should be shown for each main comparison, including assumptions about effect size, variability or event rate, attrition, allocation, and interaction. In a 2-in-1 trial, investigators may calculate a required size for each main factor comparison and choose the larger. That does not imply adequate power for all pairwise cell comparisons or for interaction.[1]
More factors, outcomes, time points, and comparisons can also increase false-positive risk. ICH E9 asks protocols to identify primary variables and hypotheses and explain the approach to statistical inference, including effects on Type I error when several primary questions are involved.[4] Whether adjustment is needed depends on the confirmatory claims and comparison family; a report should state the strategy rather than imply that the factorial label solves multiplicity.
Flow and analysis must be factor-specific
A clear factorial report should identify:
- the rationale for the factorial design and whether interaction is a primary objective;
- every factor, level, and possible cell;
- full, partial, or fractional structure;
- eligibility and randomization timing for each factor;
- prespecified main comparisons and target effects;
- sample-size assumptions for each main comparison and for interaction when relevant;
- participant flow and denominators by factor and cell;
- interim decisions or early stopping for each factor;
- the estimated main effects and interaction, with uncertainty; and
- sensitivity analyses and any departures from the planned analysis.
The February 2025 primary paper is an explanation and elaboration for the extensions to CONSORT 2010 and SPIRIT 2013 for factorial trials.[1][2] The SPIRIT–CONSORT extensions overview currently lists factorial statements under those earlier baseline checklists.[3] These are reporting recommendations, not proof of sound design or regulatory compliance.
Joint assignment does not validate an outside mixture
Randomizing two factors within one protocol does not establish that a separately sourced combination is identical to the trial interventions. The trial evidence depends on the identities, controls, delivery, eligibility, monitoring, and analysis used in that study. It also does not show that combining the factors is better unless the prespecified comparisons directly support that conclusion.
A marginal effect can average across the other factor’s levels. It should not be rewritten as evidence for every combination, formulation, ratio, or commercial blend. Product identity and comparability require their own evidence.
A factorial-trial reading checklist
- Draw every allocated cell.
- Name each factor and level.
- Distinguish cells from prespecified main comparisons.
- Identify full, partial, or fractional structure and factor-specific eligibility.
- Record the treatment effect and analysis scale for each objective.
- Check sample-size assumptions for main effects and interaction separately.
- Extract factor- and cell-specific flow, denominators, and enrollment periods.
- Review the interaction estimate and confidence interval, not separate subgroup p values.
- Examine multiplicity, stopping, and sensitivity-analysis plans.
- Limit conclusions to the tested factors, combinations, and materials.
Important limitations
- Four randomized cells do not mean four equally supported pairwise claims.
- A marginal effect may not describe the effect within every cell.
- Partial factorial participants contribute only to factors for which they were randomized.
- A nonsignificant interaction test does not prove absence of interaction.
- An interaction depends on the chosen effect scale and model.
- A trial sized for main effects may be underpowered for interaction.
- A reporting checklist does not certify design quality.
- Factorial randomization cannot authenticate an outside blend or formulation.
Frequently asked questions
What are the four groups in a 2×2 factorial trial?
They are the four combinations of two factors: both active levels, A without B, B without A, and neither active level. Exact labels depend on the comparators used in the trial.
Is each cell a main comparison?
Not necessarily. Many trials combine cells to compare everyone assigned A with everyone assigned not-A, and separately everyone assigned B with everyone assigned not-B. The protocol should identify the intended comparisons.
What is a partial factorial design?
It is a design in which some participants are not randomized for certain factors, often because eligibility or enrollment timing differs. Those participants may contribute to another factor but not to the factor for which they were not randomized.
Does no significant interaction prove the factors act independently?
No. The interaction estimate may be imprecise, especially when the trial was sized for main effects. The estimate, uncertainty, analysis scale, and plausible importance all matter.
Can separate subgroup p values test interaction?
No. Statistical significance in one subgroup and not another does not test whether the effects differ. A direct interaction comparison is required.
Why can a factorial design be efficient?
Eligible participants can contribute information to more than one randomized factor comparison. That efficiency depends on the objectives, allocation, assumptions, and analysis; it does not guarantee adequate power for interaction or every cell contrast.
Does a factorial trial prove an untested mixture works?
No. It supports only the randomized interventions, combinations, conditions, and comparisons actually studied. It cannot establish the identity, composition, or comparability of another material.