Skip to content
Back to search results

Factorial Experiment — Factorial Experimental Design

Introduced by
Ronald A. Fisher
Year
1926–1935
Type
Quantitative experimental design
Family
Experimental design
Font size

What does it do?

A factorial experiment is an experimental design in which two or more independent variables (factors) are manipulated simultaneously, and every combination of their levels is tested.

Introduced by Ronald Fisher in the 1920s–1930s, it is the standard approach whenever a researcher needs to detect not only the main effect of each factor but also whether the effect of one factor depends on the level of another — the interaction effect.

The intuition behind the method

Imagine testing a new fertilizer (yes/no) and two irrigation levels (low/high) on crop yield. A one-factor-at-a-time approach would need two separate experiments. A factorial experiment tests all four combinations — fertilizer-yes/irrigation-low, fertilizer-yes/irrigation-high, fertilizer-no/irrigation-low, fertilizer-no/irrigation-high — in one study. This not only saves resources but is the only way to discover whether the fertilizer works better under high irrigation than low irrigation. That crossover pattern is an interaction, and factorial designs are specifically built to detect it.

History

Ronald Fisher formalized the logic of factorial experiments in his 1926 paper on agricultural experiments and elaborated the theory in his landmark 1935 book The Design of Experiments, where he introduced concepts of randomization, replication, and factorial crossing as the basis of sound empirical inference. Fisher's work was motivated by the practical need to test multiple agronomic interventions simultaneously rather than one at a time. By the mid-twentieth century, factorial designs had been adopted across psychology, engineering, medicine, and the social sciences. Douglas Montgomery's widely used textbook Design and Analysis of Experiments (first edition 1976) brought factorial and fractional factorial methods to engineering and industrial audiences, cementing the approach as a cornerstone of experimental methodology.

When to use it

Use a factorial experiment when you have two or more factors you wish to study in one study and when detecting interactions is scientifically important. It is preferable to one-factor-at-a-time designs whenever resources allow, because it provides the same information about each factor at lower total sample cost and additionally yields interaction estimates. Avoid it when the number of factors is large (four or more factors with multiple levels each creates an unmanageably large number of cells); in those cases prefer a fractional factorial design or a response-surface method. Do not use a full factorial design when you have only one factor — a simple RCT or one-way ANOVA is sufficient. Also avoid it when it is operationally impossible to deliver all combinations of factor levels.

Limitations and considerations

  • Sample size requirement grows multiplicatively with the number of factors and levels, making large designs expensive.
  • Full factorial designs with many factors are often infeasible; researchers must resort to fractional designs at some loss of information.
  • Interpretation becomes complex when multiple higher-order interactions are significant, requiring careful graphing and simple-effects follow-ups.
  • All standard factorial ANOVA assumptions (normality, homogeneity of variance, independence) must be checked and, if violated, addressed.

Source conditions do not replace independent scientific review or validation by an analysis engine.

Strengths

  • The only design that can estimate interaction effects between factors — the most informative feature of multi-factor research.
  • More statistically efficient than separate one-factor studies: each observation contributes to estimating every main effect.
  • Supports causal inference when factors are fully randomized, meeting the same standard as a randomized controlled trial.
  • Highly flexible: factors can be between-subjects, within-subjects (repeated measures), or mixed.
  • Generalizable findings: conclusions about factor A apply across all tested levels of factor B rather than under a single arbitrary condition.

Applications

  • Psychology: studying how the effects of a cognitive intervention differ by participant age group or gender (2 × 2 between-subjects design).
  • Agriculture and biology: testing combinations of fertilizer type and irrigation level on crop yield.
  • Engineering and manufacturing: optimizing product quality by simultaneously varying process temperature, pressure, and material grade.
  • Pharmacology: evaluating drug dose and administration route combinations on bioavailability.
  • Education: assessing whether a new teaching method is equally effective across different class sizes and subject areas.
  • Marketing: testing combinations of message framing and media channel on consumer purchase intention.

How it works

How it works

1

Identify all independent variables (factors) of theoretical or practical interest and fix the set of levels for each. A 2 x 3 design, for example, has one factor at 2 levels and a second at 3 levels, yielding 6 treatment conditions. The choice of levels should be grounded in prior research or substantive theory, not arbitrary convenience.

2

Form every possible combination of factor levels. This crossing is what defines a factorial design: if factor A has a levels and factor B has b levels, the design contains a × b treatment cells. Full crossing ensures that main effects and all interactions are estimable without confounding.

3

Randomly assign experimental units (participants, plots, batches) to treatment cells or, in a within-subjects variant, randomize the order in which each unit receives all conditions. Proper randomization is essential for valid causal inference and for satisfying the independence assumption underlying factorial ANOVA.

4

Administer each treatment condition according to the protocol and record the outcome variable for every unit. Maintaining consistent procedures across all cells — except for the intentional manipulation of the factors — is critical for ruling out confounds.

5

Fit a factorial ANOVA (or regression with interaction terms) model: Y = mu + alpha_i + beta_j + (alpha*beta)_ij + epsilon_ijk. Significance tests and effect-size estimates (eta-squared, omega-squared) are computed for each main effect and each interaction term. The F-ratio for each source compares its mean square to the residual mean square.

Formula
Y = mu + alpha_i + beta_j + (alpha*beta)_ij + epsilon_ijk
6

Examine interaction plots (cell means plotted with one factor on the x-axis and lines for the other factor). Non-parallel lines signal an interaction: the effect of one factor is not uniform across levels of the other. When a significant interaction is found, main effects must be interpreted cautiously — simple effects analyses (one factor at each level of the other) clarify the pattern.

Common pitfalls

  • Ignoring significant interactions and interpreting main effects in isolation — when an interaction is present, a main effect is a misleading average across heterogeneous sub-groups; always plot cell means first.
  • Underpowering individual cells — power in a factorial design must be calculated for the smallest expected effect (often the interaction), not for the main effects alone; using main-effect power formulas leads to underpowered interaction tests.
  • Failing to randomize or balance cell sizes — unequal cell sizes introduce non-orthogonality, complicating the partitioning of variance; use Type III sums of squares and report the imbalance explicitly.
  • Treating a factorial design as if it were separate one-way ANOVAs — running individual ANOVAs for each factor ignores the crossed structure and inflates Type I error; always analyse the full factorial model.
  • Including too many factors — every added factor multiplies the number of cells and the complexity of interaction plots; four or more factors usually calls for a fractional or response-surface approach.

Frequently asked questions

What is the difference between a factorial experiment and a full factorial experiment?

Every full factorial experiment is a factorial experiment — the terms are often used interchangeably. When researchers distinguish them, 'factorial experiment' is the general category (includes both full and fractional variants), and 'full factorial' specifies that every possible combination of factor levels is included. A fractional factorial omits a systematic subset of cells to reduce cost, at the price of confounding higher-order interactions.

How many participants do I need per cell?

At minimum, two or three observations per cell are needed to estimate within-cell variance, but this is rarely sufficient for adequate power. A formal power analysis using the expected effect size for the interaction (typically smaller than main effects) and the desired alpha and power levels should determine cell size. Five to twenty participants per cell is a common working range in behavioural research, but engineering studies may use three or fewer with high measurement precision.

What if my interaction is significant — do I still report main effects?

Yes, but with caution. When a significant interaction is present, main effects describe averages across heterogeneous conditions, which can be misleading. Report them, but centre the interpretation on the interaction: conduct simple-effects analyses (testing one factor separately at each level of the other) and present interaction plots so readers can see the pattern directly.

Can factorial designs be used with within-subjects (repeated measures) factors?

Yes. Mixed factorial designs combine at least one between-subjects factor with at least one within-subjects factor. Fully within-subjects factorial designs are also possible. These designs increase statistical power because individual differences are controlled, but they introduce carryover and order effects that must be addressed through counterbalancing or washout periods.

When should I use a fractional factorial instead?

When the number of factors is large (typically four or more with multiple levels each), running all treatment combinations becomes prohibitively expensive. A fractional factorial design tests a carefully chosen fraction of the cells — sacrificing estimates of high-order interactions (which are often negligible in practice) in exchange for a dramatic reduction in run count. If interactions beyond two-way are theoretically implausible, a half-fraction or quarter-fraction design is often the better choice.

Related methods

Sources

  1. Fisher, R. A. (1935). The Design of Experiments. Oliver and Boyd.

    Open publication reference
    Copied.
  2. Montgomery, D. C. (2017). Design and Analysis of Experiments (9th ed.). Wiley.

    ISBN: 978-1119320937
    Copied.