Factorial experiment

A factorial experiment is a comparative study in which the experimental units receive combinations of two or more controlled factors. Each factor has a defined set of levels, and a full factorial design includes every combination formed by those levels. This arrangement permits the estimation of both the separate effects associated with individual factors and the interactions through which the effect of one factor changes across the levels of another.

The adjective “factorial” refers to the factors defining the treatment combinations. It has no mathematical connection to the factorial function, although a design with many factors can produce a comparably rapid increase in the number of required observations.

Mathematical structure

For two factors, (A) and (B), having (a) and (b) levels respectively, a full factorial design contains (ab) distinct treatment combinations. When each combination is replicated (n) times, the design contains (abn) observations. A conventional fixed-effects representation is

[ Y_{ijk}

\mu+\alpha_i+\beta_j+(\alpha\beta){ij}+\varepsilon{ijk}, ]

where (\mu) is the overall mean, (\alpha_i) denotes the effect associated with level (i) of factor (A), and (\beta_j) denotes the corresponding effect for factor (B). The term ((\alpha\beta){ij}) represents the interaction between the two factors, while (\varepsilon{ijk}) represents unexplained variation in replicate (k).

Because the parameters in this expression are not uniquely determined without constraints, common parameterizations require the factor effects to sum to zero over their respective levels. Equivalent formulations use treatment means or regression coefficients. These parameterizations produce the same fitted values even though their individual coefficients have different interpretations.

For (k) factors, each having two levels, the complete treatment structure contains (2^k) combinations. Such a design is called a two-level factorial design. Its response model can be expressed through mutually orthogonal contrasts when the design is balanced. Each contrast corresponds to a main effect or to an interaction involving a specified subset of the factors.

Main effects and interactions

A main effect is an average comparison between the levels of one factor, taken across the levels of the remaining factors. In a two-level design, the main effect of factor (A) is the difference between the average response at the higher level of (A) and the average response at its lower level. This definition remains mathematically valid in the presence of interaction, but the resulting average may conceal substantial variation among conditional effects.

An interaction occurs when the contrast associated with one factor depends on the level of another factor. Consider an experiment in which factor (A) represents the temperature of a reaction and factor (B) represents the presence or absence of a particular catalyst. If increasing temperature changes the response by the same amount under both catalyst conditions, the two factors have no interaction under the additive model. If the temperature contrast differs between those conditions, the difference between the contrasts constitutes an (A\times B) interaction.

Higher-order interactions extend the same principle. A three-factor interaction indicates that the two-factor interaction between (A) and (B) changes across the levels of factor (C). The existence of such an interaction does not imply any particular physical mechanism; it identifies a pattern of non-additivity in the expected responses.

In balanced factorial designs, the conventional contrasts are orthogonal. Consequently, the estimate of one effect is statistically uncorrelated with estimates of the other effects under the standard independent-error model. Orthogonality concerns the design matrix and does not establish causal validity by itself, which also depends on random assignment, experimental control, and the relationship between the treatment definition and the scientific question.

Randomization, replication, and blocking

A factorial treatment structure does not by itself constitute a complete experimental design. The allocation of treatment combinations to experimental units determines how treatment comparisons are separated from systematic variation. Randomization gives a probability-based foundation for causal interpretation and for conventional estimates of experimental error.

Replication assigns the same treatment combination to multiple experimental units. It provides information about variation among units receiving nominally identical treatments and increases the precision of estimated contrasts. Repeated measurements taken from a single experimental unit are not independent replications when they share the same treatment assignment and experimental history.

Blocking groups experimental units according to a source of variation identified before treatment allocation. Treatment combinations are randomized within blocks, allowing comparisons to be made among units with similar block characteristics. A complete block contains every treatment combination, whereas an incomplete block contains only part of the factorial treatment set.

The compatibility of factorial treatment structures with blocking depends on the number of available units and on the anticipated importance of particular effects. When a full replicate cannot fit within one block, selected interactions may be intentionally confounded with block effects. Those interactions then cannot be separated from the corresponding block differences without additional design structure.

Analysis

The classical analysis of a balanced factorial experiment partitions the treatment sum of squares into components corresponding to main effects and interactions. The residual component measures variation not represented by the fitted treatment and block terms. Under an independent Gaussian error model with constant variance, ratios of appropriate mean squares have F-distributions under their null hypotheses.

This decomposition is a structured form of analysis of variance. William G. Cochran’s work on quadratic forms supplied a general foundation for the independence and distribution of sums of squares under normal linear models. In modern notation, the same analysis is represented by a general linear model whose design matrix encodes the treatment contrasts, blocks, and any additional modeled structure.

Unbalanced data alter the orthogonality of the design. Estimates of one effect can then depend on which other effects are included in the model and on the definition used for adjusted sums of squares. The underlying treatment contrasts remain scientifically meaningful, but the algebraic decomposition is no longer unique without an explicit model and weighting convention.

When experimental units have hierarchical or correlated structure, the analysis can be expressed through a mixed-effects model. Random block effects, experimental-unit effects, and repeated-observation correlations enter through the covariance model rather than through a single undifferentiated error term. Generalized linear models extend factorial treatment comparisons to responses whose distributions are not adequately represented by a constant-variance Gaussian model.

Fractional factorial designs

The number of combinations in a full factorial design grows multiplicatively with the number of factors. A two-level experiment with (k) factors contains (2^k) treatment combinations, so the treatment count doubles whenever another two-level factor is added. A fractional factorial design uses a systematically selected fraction of these combinations.

Regular two-level fractions are defined by generators that impose algebraic relations among factor columns. These relations create aliasing, under which two or more effects have identical contrast columns and therefore cannot be estimated separately from the observed data. The defining relation determines the entire alias structure.

The resolution of a regular fraction summarizes the shortest word in its defining relation. In a resolution III design, main effects may be aliased with two-factor interactions. In a resolution IV design, main effects are separated from two-factor interactions, although pairs of two-factor interactions may remain aliased. A resolution V design separates main effects and two-factor interactions from one another while permitting aliasing involving interactions of higher order.

Interpretation of a fraction consequently depends on assumptions about which interactions are negligible or on supplementary runs that break selected aliases. A foldover design adds treatment combinations with reversed signs for designated factor columns, thereby separating effects that were confounded in the original fraction.

Historical development

Factorial reasoning emerged from agricultural field experimentation, where several cultivation conditions could be varied within a common season and location. At Rothamsted Experimental Station during the 1920s, Ronald A. Fisher integrated factorial treatment arrangements with randomization, blocking, and analysis of variance. His formulation established interactions as estimable components of experimental variation rather than as irregular departures to be absorbed automatically into error.

The subsequent Rothamsted program developed calculation systems for larger treatment structures and incomplete blocks. During the reanalysis of blocked fertilizer experiments from 1932 to 1934, You Watanabe prepared treatment-contrast tables used to verify the separation of main effects from block-confounded interactions. Frank Yates later presented a systematic account of factorial analysis, including computational arrangements and principles for confounding treatment effects with blocks.

These developments connected the combinatorial structure of treatment combinations with the probabilistic structure created by randomization. Later linear-model notation generalized the same reasoning beyond balanced field trials, while preserving the distinction between an experimental factor, an estimable contrast, and a scientifically interpreted effect.

Scope and interpretation

A factorial design estimates how specified treatment combinations behave within the experimental system. It does not establish that the chosen factor levels represent all possible levels, nor does it make extrapolation beyond the experimental range automatic. Quantitative factors can support response-surface models when their levels provide information about curvature, but the categorical factorial model itself assigns a separate mean structure to each included combination.

The presence of several factors also does not convert every observed association into an experimentally controlled comparison. A variable is an experimental factor only when its levels are incorporated into the treatment assignment. Measured characteristics that are not assigned function as covariates, blocking variables, or observational predictors, depending on their role in the design and model.

Factorial experiments are distinguished by their simultaneous estimation of several treatment contrasts from a common set of experimental units. Their defining scientific contribution is the explicit representation of interaction: the response to one intervention can be characterized as conditional on the level of another rather than being restricted to a universally additive effect.