Skip to content
Back to search results

Design of Experiments — DOE

Introduced by
Ronald A. Fisher
Year
1935
Type
Experimental planning framework
Family
Engineering methods
Font size

What does it do?

Design of Experiments (DOE) is a systematic framework for planning, conducting, and analyzing controlled experiments to determine how multiple input factors simultaneously affect one or more responses.

Introduced by Ronald A. Fisher in 1935, DOE allows researchers and engineers to identify causal relationships, quantify factor effects, and find optimal settings efficiently — using far fewer runs than one-factor-at-a-time approaches. It is foundational in engineering, manufacturing, agriculture, and applied sciences.

The intuition behind the method

Suppose you want to improve the yield of a chemical process and suspect that temperature, pressure, and reaction time all matter — but testing every combination individually would take months. DOE solves this by prescribing a structured set of experimental runs that, together, let you estimate each factor's effect and all their interactions simultaneously. The key insight is that by changing multiple factors at once in a deliberate pattern, you gather far more information per experiment than by varying only one thing at a time, and you can detect synergistic or antagonistic interactions that a one-factor approach would miss entirely.

History

Ronald A. Fisher developed the foundational principles of experimental design — randomization, replication, and blocking — while working at Rothamsted Experimental Station and published them in The Design of Experiments (1935). Frank Yates and Oscar Kempthorne extended the theoretical foundations through the 1940s and 1950s. George Box and colleagues pioneered response surface methodology in the 1950s, adapting DOE for process optimization. Genichi Taguchi introduced his orthogonal array approach for robust engineering design in the 1980s. Douglas Montgomery's textbook (first edition 1976, now in its ninth edition) became the standard industrial reference.

When to use it

DOE is the method of choice whenever you need to understand or optimize a process that has multiple controllable inputs and measurable outputs, and where the cost of unplanned experimentation is high. It is ideal for process development, product formulation, engineering optimization, and quality improvement. Use DOE when you can actively set factor levels and control experimental conditions. Do not use it when factors cannot be manipulated (observational data only — use regression or observational study designs instead), when you have only a single factor to vary (a simple controlled experiment suffices), or when physical experiments are too expensive and a computer simulation surrogate (metamodel) has not been built first.

Limitations and considerations

  • Requires the ability to actively control and set factor levels — purely observational settings are out of scope.
  • Full factorial designs grow exponentially with the number of factors, making them impractical beyond five or six factors without fractionation.
  • Model validity depends on the chosen design resolution; low-resolution fractional designs confound main effects with interactions, potentially misleading conclusions.
  • Assumes a stable process during the experiment; process drift or uncontrolled shifts invalidate the randomization assumption.

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

Strengths

  • Estimates all main effects and interactions simultaneously, revealing synergies that one-factor-at-a-time experiments miss.
  • Dramatically reduces the total number of runs needed compared to exhaustive one-factor-at-a-time testing.
  • Randomization and blocking provide a rigorous basis for causal inference and valid statistical tests.
  • Scales from simple two-level screening designs to complex response surface optimization within the same framework.
  • Results are directly actionable: optimal settings, prediction equations, and confidence intervals on effects.

Applications

  • Optimizing manufacturing process parameters (temperature, pressure, feed rate) to maximize product yield and quality.
  • Formulating consumer products — identifying ingredient levels and processing conditions that meet sensory and stability targets.
  • Accelerated reliability testing — determining which stress factors most affect product life.
  • Agricultural field trials — evaluating fertilizer types, irrigation levels, and crop varieties simultaneously.
  • Pharmaceutical process development and analytical method validation under ICH Q8/Q9 guidelines.
  • Software performance benchmarking — systematically varying configuration parameters to identify bottlenecks.

How it works

How it works

1

The experimenter defines the objective (maximize yield, minimize defects, understand relationships), identifies the response variable(s) to be measured, and selects the input factors and their levels. Nuisance variables that cannot be controlled are identified and handled by blocking or randomization. A clear problem statement prevents wasted runs on the wrong question.

2

An experimental design matrix is chosen based on the number of factors, desired resolution, and available budget of runs. Options range from full factorial designs (all combinations of factor levels) to fractional factorials, response surface designs (Central Composite, Box-Behnken), and Taguchi arrays. The matrix specifies exactly which factor-level combinations to test and in what randomized order.

3

Experiments are conducted in the order specified by the randomized design. Randomization guards against systematic bias from time trends or operator learning effects. Replication — running the same treatment more than once — provides an estimate of pure experimental error, enabling valid significance tests.

4

Responses are modeled as a linear (or quadratic) function of the factors and their interactions: y = beta_0 + sum(beta_i * x_i) + sum(beta_ij * x_i * x_j) + epsilon. Analysis of variance (ANOVA) partitions total variability into factor effects and error, identifying which factors and interactions are statistically significant. Effect plots, interaction plots, and residual diagnostics support interpretation.

Formula
y = \beta_0 + \sum \beta_i x_i + \sum \beta_{ij} x_i x_j + \varepsilon
5

Significant effects are translated into actionable decisions: optimal factor settings are identified (possibly using response surface optimization), confirmation runs are performed to validate predictions, and findings are documented. If key assumptions (normality, constant variance) are violated, transformations or alternative models are applied before drawing conclusions.

Common pitfalls

  • Omitting important factors — if a key factor is excluded, its variation inflates the error term and can mask real effects; thorough cause-and-effect analysis before designing the experiment is essential.
  • Skipping randomization — running experiments in a systematic (non-random) order allows time-related confounders to masquerade as factor effects, invalidating significance tests.
  • Ignoring interactions — analyzing only main effects when strong interactions exist leads to incorrect optimal settings; always plot and test interaction terms.
  • Insufficient replication — without replication there is no estimate of pure error, making it impossible to distinguish real effects from noise; at least some center points or replicate runs should be included.
  • Overfitting the response model — fitting a high-order polynomial to a small dataset produces an equation that fits the data but predicts poorly; use adjusted R-squared and residual plots to guard against this.

Frequently asked questions

How is DOE different from simply trying different combinations?

Ad hoc trials change one factor at a time, which cannot detect interactions and requires many more runs to cover the same experimental space. DOE uses a statistically balanced design matrix that ensures every factor is tested at every level in combination with every other factor (or a carefully chosen fraction thereof), enabling simultaneous estimation of all effects with minimum runs and valid error estimates.

How many factors can I include in a DOE?

Practically, screening designs (Plackett-Burman, Resolution III fractional factorials) can handle 7–30+ factors efficiently to identify the vital few. Once the active factors are identified, a smaller follow-up design (response surface or full factorial on 2–5 factors) is used for optimization. Starting with too many factors in a detailed design wastes resources; start broad and screen first.

What is resolution in a fractional factorial design?

Resolution describes the degree to which effects are confounded. Resolution III designs confound main effects with two-factor interactions — acceptable for initial screening. Resolution IV designs confound two-factor interactions with each other but not with main effects — better for understanding interactions. Resolution V designs keep all main effects and two-factor interactions estimable — preferred for optimization. Always choose the minimum resolution appropriate for your goals.

Do I need special software to run DOE?

Software significantly eases design generation, randomization, and analysis. JMP, Minitab, Design-Expert, and R packages (FrF2, rsm) are widely used. However, understanding the underlying principles — balance, randomization, confounding — is necessary to choose the right design and interpret results correctly; software alone cannot substitute for that knowledge.

When should I use a response surface design instead of a factorial design?

Use a factorial design when you want to screen or compare factors at discrete levels. Move to a response surface design (Central Composite or Box-Behnken) when you need to model curvature in the response and find an optimum operating point, because these designs include the center and intermediate levels required to fit a second-order polynomial.

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: 978-1119492443
    Copied.