Fractional Factorial Experiment — Fractional Factorial Experimental Design
- Introduced by
- D. J. Finney (formal development); foundations in Ronald Fisher's factorial design work
- Year
- 1945 (Finney); broader development 1950s–1970s by Box, Hunter
- Type
- Quantitative experimental design
- Family
- Experimental design
What does it do?
A fractional factorial experiment is a resource-efficient experimental design that tests only a carefully chosen fraction of all possible factor-level combinations.
By exploiting the principle that high-order interactions are usually negligible, it identifies the main effects and low-order interactions of k factors using far fewer runs than a full factorial design — making it the workhorse of industrial and engineering screening experiments.
The intuition behind the method
Suppose you want to study 7 factors, each at two levels. A full factorial would require 128 runs — impractical and expensive. A fractional factorial takes a strategically selected half (or quarter) of those combinations: 64 or 32 runs. The trick is that each omitted run corresponds to a high-order interaction that is almost certainly zero in practice. You trade the ability to estimate those unlikely interactions for a dramatic reduction in experimental cost, while still capturing the effects that matter most. The key is knowing exactly which effects get confounded (aliased) with each other, so you can interpret results correctly.
History
Fractional factorial designs grew out of Ronald Fisher's foundational work on factorial experiments at Rothamsted Experimental Station in the 1920s and 1930s. D. J. Finney formally introduced fractional replication in a 1945 paper in the Annals of Eugenics, showing that a carefully chosen subset of runs could estimate main effects while sacrificing only high-order interactions. George Box, J. Stuart Hunter, and William Hunter popularized and systematized the methodology for industrial experimentation through their influential textbook series starting in 1978, establishing resolution and aliasing as the key concepts practitioners use today. The designs remain a cornerstone of Six Sigma, Design of Experiments (DOE) training, and quality engineering worldwide.
When to use it
Use a fractional factorial design when you have many factors (typically five or more) to screen and cannot afford a full factorial, when resources (time, material, cost) are limited, and when high-order interactions are assumed negligible — a reasonable assumption in most physical, chemical, and engineering settings. It is the standard first step in a sequential experimentation strategy (screen → optimize). Do NOT use it when you need to estimate all high-order interactions (use a full factorial instead), when you have fewer than four or five factors (a full factorial is affordable), when aliasing of two-factor interactions is unacceptable without follow-up runs, or when the response surface must be characterized in detail (use response surface methodology instead).
Limitations and considerations
- Aliasing means some effects cannot be independently estimated within a single fraction; resolving aliases requires additional runs.
- The assumption that high-order interactions are negligible may not hold in biological, social, or complex systems.
- Requires statistical literacy to correctly specify generators, check resolution, and interpret the alias pattern.
- With very few runs (e.g., 8-run Resolution III design for 7 factors), there is little power to detect all but large effects.
Source conditions do not replace independent scientific review or validation by an analysis engine.
Strengths
- Dramatically reduces the number of experimental runs compared to a full factorial, saving time and resources.
- Systematic alias structure means the researcher always knows exactly which effects are confounded — there are no hidden ambiguities.
- Sequential and combinable: fractions can be folded over or combined with new fractions to de-alias important effects in a second stage.
- Widely supported in statistical software (R, SAS, JMP, Minitab) and extensively documented in the literature.
- Orthogonal design structure makes parameter estimation straightforward and statistically efficient.
Applications
- Screening dozens of process variables in chemical or pharmaceutical manufacturing to identify the few that most affect yield or purity.
- Optimizing agricultural treatments — fertilizer, irrigation, pesticide combinations — over a growing season with limited plots.
- Engineering product design studies where material, geometry, and process parameters must all be explored within a fixed prototype budget.
- Software and web experimentation (feature flags, UI elements) when running a full factorial across many features simultaneously is infeasible.
- Analytical chemistry method development — identifying which reagent concentrations, temperatures, and instrument settings drive assay performance.
- Quality improvement projects in manufacturing using DMAIC or Six Sigma frameworks.
How it works
How it works
Identify all k factors to investigate and set their levels (typically two levels — low and high — coded as -1 and +1). Clarify the response variable(s) to be measured. With more than five or six factors, a fractional design is typically warranted; with fewer than four, a full factorial is usually feasible.
Choose the fraction 2^(k-p), where p defines how many factors' columns are generated by multiplying existing columns. A half-fraction (p=1) uses 2^(k-1) runs; a quarter-fraction (p=2) uses 2^(k-2) runs. The resolution of the design (III, IV, V) determines which effects are aliased: Resolution III confounds main effects with two-factor interactions; Resolution IV confounds two-factor interactions with each other; Resolution V or higher keeps main effects and two-factor interactions clear of each other.
Formula
2^{k-p}Build the design matrix by first writing out a full factorial in the base k-p factors, then creating the additional p columns by multiplying existing columns according to the chosen generator(s). Verify the aliasing structure (alias pattern) to confirm it is acceptable for the research question.
Perform the experimental runs in a randomized order to guard against systematic time-related biases. Apply blocking if there is a nuisance factor such as batch, day, or operator. Record the measured response for each run, including any replicated or center-point runs included to estimate pure error.
Estimate main effects and interaction effects using ordinary least squares or Yates's algorithm. Calculate effect magnitudes and their standard errors. Use a normal probability plot of effects, half-normal plot, or Lenth's method to distinguish active (real) effects from the noise, especially when there are no replicated runs to provide a direct error estimate.
Interpret every estimated effect in light of the alias structure: a significant estimate may reflect the named effect, its alias, or both. Use subject-matter knowledge, follow-up confirmation runs, or a fold-over design to de-alias effects that cannot be resolved from the current fraction. Report the resolution and alias pattern transparently so readers can assess what was and was not estimable.
Common pitfalls
- Choosing too low a resolution (e.g., Resolution III) when two-factor interactions are plausible — main effect estimates are then biased by those interactions, and results can be seriously misleading.
- Ignoring randomization: running the design in standard order introduces confounding with time trends that the aliasing structure was not designed to handle.
- Failing to report the alias pattern in the write-up, making it impossible for readers (or reviewers) to assess which effects are estimable.
- Treating a screening result as final without a confirmation run or follow-up optimization phase — fractional factorial designs are inherently a first step, not the last word.
Frequently asked questions
What does 'resolution' mean in a fractional factorial design?
Resolution describes which effects are aliased (confounded) with which. Resolution III: main effects are aliased with two-factor interactions — use only when two-factor interactions are known to be negligible. Resolution IV: main effects are clear of two-factor interactions, but two-factor interactions are aliased with each other. Resolution V or higher: both main effects and two-factor interactions are estimable independently — preferred when two-factor interactions are plausible.
How do I choose between a half-fraction and a quarter-fraction?
A half-fraction (2^(k-1)) retains higher resolution but only halves the run count. A quarter-fraction (2^(k-2)) is smaller but typically has lower resolution, increasing the aliasing problem. If you have seven or more factors and resources are very tight, start with a Resolution IV quarter-fraction; if you have four to six factors, a half-fraction at Resolution IV or V is usually achievable and preferable.
How is a fractional factorial different from a Plackett-Burman design?
Both are screening designs for many factors in few runs, but they use different construction principles. Fractional factorial designs (2^(k-p)) are based on two-level full factorials and have a clear, interpretable alias structure. Plackett-Burman designs use Hadamard matrices; they are more flexible in run size (any multiple of 4) but have a more complex, non-hierarchical alias structure that makes interpretation of two-factor interactions harder.
What should I do after a fractional factorial identifies active factors?
A fractional factorial is a screening design, not an optimization design. Once active factors are identified, move to a follow-up phase: fold-over the original fraction to resolve aliased effects, then use a response surface method (central composite design or Box-Behnken) or a new full factorial on only the active factors to characterize and optimize the response.
Can I use a fractional factorial with more than two levels per factor?
Yes. Three-level fractional factorials (e.g., 3^(k-p) designs) exist and are used when curvature is suspected. However, they require substantially more runs than two-level designs. For initial screening, two-level fractional factorials are almost always used; three-level or mixed-level designs are more common in the optimization phase.
Related methods
Sources
Box, G. E. P., Hunter, J. S., & Hunter, W. G. (2005). Statistics for Experimenters: Design, Innovation, and Discovery (2nd ed.). Wiley-Interscience.
ISBN: 978-0471718130Copied.Finney, D. J. (1945). The fractional replication of factorial arrangements. Annals of Eugenics, 12(1), 291–301.
Open DOI referenceCopied.