Skip to content
Back to search results

Full Factorial Experiment — Complete Factorial Experimental Design

Introduced by
Ronald A. Fisher
Year
1926 (Fisher's foundational paper); codified by the 1950s–1960s
Type
Experimental design
Family
Experimental design
Font size

What does it do?

A full factorial experiment runs every possible combination of all chosen factor levels, making it the gold standard for simultaneously estimating main effects, two-way interactions, and higher-order interactions among multiple independent variables.

Introduced through Ronald Fisher's foundational work on factorial designs in the 1920s and systematised by Box, Hunter, and Montgomery, it provides complete information about how factors act individually and in combination on an outcome.

The intuition behind the method

Suppose you are testing two fertiliser types (A, B) and three watering schedules (daily, weekly, monthly) on crop yield. A full factorial design means you grow crops under all 2 × 3 = 6 combinations, not just each factor in isolation. This is powerful because the effect of fertiliser may depend on watering frequency — an interaction that one-factor-at-a-time studies completely miss. The full factorial captures every interaction by ensuring that each level of one factor appears with every level of every other factor, giving a balanced, unconfounded view of the entire factor space.

History

Ronald Fisher introduced the concept of factorial experimentation in his landmark 1926 paper and developed it fully in The Design of Experiments (1935), arguing that studying multiple factors simultaneously is more efficient than one-at-a-time approaches. Frank Yates formalised the notation and computation of factorial effects in the 1930s. George Box, J. Stuart Hunter, and William Hunter systematised the practical application of factorial designs for industrial and scientific settings in their seminal 1978 textbook (updated 2005). Douglas Montgomery's widely adopted textbook (first edition 1976) has since made full factorial design central to engineering and applied-science curricula worldwide.

When to use it

Use a full factorial experiment when you have a small number of factors (typically k ≤ 4 or 5 for two-level designs) and need to estimate all main effects and interactions without ambiguity. It is the right choice when interactions are expected or their presence is scientifically important, when resources allow completing every combination, and when the outcome is measurable under controlled conditions. Do NOT use a full factorial design when k is large (say k ≥ 6 at two levels), because the required 64+ runs become infeasible — prefer fractional factorial or response-surface designs instead. Also avoid it when run cost or material limits make N = 2^k (or 3^k) prohibitive, or when you only care about main effects and have strong theoretical reasons to assume no interactions.

Limitations and considerations

  • Run count grows exponentially with the number of factors: 2^k = 32 runs for 5 factors, 64 for 6 — quickly becoming impractical.
  • High-order interactions (three-way and above) are estimable but rarely interpretable or scientifically meaningful, making many runs informationally redundant.
  • Assumes factors can be set precisely at the specified levels; factor control is often easier in the lab than in field or organisational settings.
  • Replication is resource-intensive; without it, error estimation relies on pooling negligible high-order interactions, which requires judgment.

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

Strengths

  • Estimates every main effect and every interaction without aliasing or confounding — the most complete picture possible from a controlled experiment.
  • Orthogonal design: all effect estimates are statistically independent, making interpretation straightforward.
  • Provides a pure error estimate when runs are replicated, enabling rigorous hypothesis testing via ANOVA.
  • Flexible structure: works with continuous, categorical, or mixed factor types and can incorporate blocking for nuisance control.
  • Scales naturally: a 2^k design with center-point runs can be augmented into a response-surface design (central composite or Box-Behnken) for curvature detection.

Applications

  • Pharmaceutical formulation: simultaneous optimisation of drug concentration, excipient type, and temperature to maximise dissolution rate.
  • Agricultural science: testing combinations of fertiliser, irrigation level, and planting density on crop yield.
  • Manufacturing process improvement: identifying how machine speed, temperature, and material grade jointly affect product tensile strength.
  • Behavioural and educational experiments: examining how teaching method, class size, and assessment type interact to affect learning outcomes.
  • Clinical and health research: evaluating combinations of behavioural intervention components to identify which combinations drive the therapeutic effect.
  • Software engineering: testing the impact of algorithm type, dataset size, and hardware configuration on computational performance.

How it works

How it works

1

Identify k experimental factors and specify the levels for each. A two-level design (each factor at low/high) is denoted 2^k; a three-level design is 3^k. Factor levels should be chosen to span the range of practical or scientific interest. The total number of treatment combinations equals the product of all level counts across factors.

2

The total number of experimental runs is N = the product of all level counts (e.g., 2^3 = 8 for three two-level factors). This grows exponentially with the number of factors, so full factorial designs are most feasible when k is small (typically k ≤ 5 for two-level factors). A design matrix is constructed listing every combination; replication of each combination strengthens error estimation.

Formula
N = \prod_{i=1}^{k} l_i
3

Experimental units are randomly assigned to each treatment combination, or the order of runs is fully randomised. Randomisation protects against systematic bias from time trends, fatigue, or equipment drift. If some nuisance variables are known (e.g., batch, operator, day), blocking is added to isolate their variance without confounding factor estimates.

4

The response variable is measured for each experimental run under its assigned combination of factor levels. Data are recorded systematically, preserving the link between run order, factor settings, and observed outcome. Multiple replications per cell are recommended whenever resources allow, as they provide a pure estimate of experimental error.

5

Main effects and all interaction effects are estimated using the method of least squares (or ANOVA). In a 2^k design each effect estimate equals a linear contrast of the cell means; the formula beta-hat = (X'X)^{-1} X'y applies where X is the model matrix of coded factor settings. Because every combination is tested, all effects are orthogonal and estimable without aliasing.

Formula
\hat{\beta} = (X^\top X)^{-1} X^\top y
6

ANOVA tables partition total variance into main effects, two-factor interactions, three-factor interactions, and so on. Significant interactions are interpreted using interaction plots: parallel lines indicate no interaction; crossing or diverging lines signal that the effect of one factor changes across levels of another. Higher-order interactions (three-way or above) are rare in practice but are estimable and testable in a full factorial design.

Common pitfalls

  • Ignoring interactions when choosing the design: if you expect no interactions, a full factorial wastes runs — but if interactions exist and you used a one-factor-at-a-time approach, you will miss them entirely.
  • Using too many levels without justification: a 3^k design has 3^5 = 243 runs for five factors; unless curvature is specifically hypothesised, two-level designs are usually sufficient and far more efficient.
  • Confusing a full factorial with a fractional factorial: a fractional factorial deliberately omits certain combinations and aliases some interactions — if you need all interactions estimable, only the full factorial provides this.
  • Failing to randomise run order: systematic trends (equipment warm-up, operator learning, reagent degradation) inflate or deflate effect estimates if runs are executed in a fixed, non-random order.
  • Analysing main effects in isolation when interactions are significant: a statistically significant interaction means main-effect conclusions depend on the level of the moderating factor and must be interpreted conditionally.

Frequently asked questions

What is the difference between a full factorial and a fractional factorial design?

A full factorial runs every possible combination of factor levels, estimating all main effects and interactions without aliasing. A fractional factorial runs only a carefully chosen fraction of those combinations, deliberately aliasing certain higher-order interactions to reduce run count. Use the full factorial when interactions matter and resources allow; use the fractional factorial when k is large and you are willing to assume certain high-order interactions are negligible.

How do I choose how many levels to use for each factor?

Two levels (low and high) are sufficient when you expect a roughly linear response across the factor range and are primarily interested in direction and magnitude of effects. Three levels are needed when you suspect curvature (a nonlinear response), but they cube the run count for each additional factor. A practical strategy is to start with 2^k designs augmented with center-point runs to screen for curvature before committing to a 3^k design.

Do I need to replicate each treatment combination?

Replication is strongly recommended. Without replication, the only source of error estimation is the pooled sum of squares from negligible high-order interactions (assumed zero), which is a risky assumption. With even one full replicate (doubling all runs) you obtain a genuine pure-error estimate, tighter confidence intervals, and more reliable significance tests.

What statistical test do I use to analyse a full factorial experiment?

The standard analysis is a factorial ANOVA (or its regression-model equivalent), which partitions the total variance into components attributable to each main effect and each interaction term. Effect significance is judged by F-tests. When the outcome is non-normal or ordinal, nonparametric alternatives or generalised linear models are used instead.

At what point should I switch to a response surface design instead?

If your 2^k full factorial (possibly with center points) reveals statistically significant curvature in the response, the linear factorial model is inadequate and a response surface design — such as a central composite or Box-Behnken design — should be used. Response surface designs add axial (star) points to estimate quadratic terms and map the full curvature of the response surface.

Related methods

Sources

  1. Box, G. E. P., Hunter, J. S., & Hunter, W. G. (2005). Statistics for Experimenters: Design, Innovation, and Discovery (2nd ed.). Wiley-Interscience.

    ISBN: 978-0471718130
    Copied.
  2. Montgomery, D. C. (2017). Design and Analysis of Experiments (9th ed.). Wiley.

    ISBN: 978-1119492443
    Copied.