A practical guide to Design of Experiments (DOE) — full and fractional factorial designs, 2^k experiments, main effects and interactions, response surface methodology, and Taguchi methods.
What Is Design of Experiments?
Design of Experiments (DOE) is a statistical methodology for planning experiments so that the influence of several input factors on one or more output responses can be learned efficiently and rigorously. The traditional alternative — changing one factor at a time (OFAT) — is inefficient and, worse, blind to interactions between factors. DOE varies multiple factors simultaneously in a designed pattern, extracting more information from fewer runs and revealing how factors combine.
DOE is a core tool in the Improve phase of Six Sigma and throughout process and product engineering, used to optimize settings, identify dominant factors, and build predictive models.
Key Terminology
- Factor — an input variable that is deliberately changed (e.g., temperature, pressure).
- Level — a specific value a factor is set to (e.g., low/high).
- Response — the measured output (e.g., yield, strength).
- Run — a single experimental trial at a specified combination of factor levels.
- Replication — repeating runs to estimate experimental error.
- Randomization — running trials in random order to guard against lurking variables.
Factorial Designs
A full factorial design tests every combination of factor levels. The most common form is the 2k design: k factors, each at two levels (low = −1, high = +1), requiring 2k runs. A 22 design has 4 runs, a 23 has 8, and a 24 has 16. From these runs you can estimate every main effect and every interaction.
| Run | A | B | C |
|---|
| 1 | − | − | − |
| 2 | + | − | − |
| 3 | − | + | − |
| 4 | + | + | − |
| 5 | − | − | + |
| 6 | + | − | + |
| 7 | − | + | + |
| 8 | + | + | + |
This is the design matrix for a 23 full factorial in factors A, B, and C.
Main Effects and Interactions
A main effect is the average change in the response as a factor moves from its low to high level, computed as (average response at high) − (average response at low). An interaction exists when the effect of one factor depends on the level of another — for instance, when temperature only improves yield at high pressure. Interactions are invisible to OFAT experiments but fall out naturally from a factorial design, which is the single biggest reason to use DOE.
Fractional Factorial Designs
When the number of factors is large, full factorials become expensive (27 = 128 runs). A fractional factorial runs a carefully chosen subset — for example a 27−3 design with only 16 runs — accepting that some higher-order effects become confounded (aliased) with each other. Because high-order interactions are usually negligible, fractional factorials are excellent for screening: separating the vital few factors from the trivial many before deeper study.
Response Surface Methodology
Two-level factorials assume effects are roughly linear, which is fine for screening but cannot locate a curved optimum. Response Surface Methodology (RSM) extends DOE to model curvature using designs such as the central composite design (CCD) and Box-Behnken design, which add center points and axial points so a quadratic model can be fit:
y = β₀ + Σβᵢxᵢ + Σβᵢᵢxᵢ² + ΣΣβᵢⱼxᵢxⱼ + ε
The fitted surface lets engineers find the settings that maximize, minimize, or target the response, and visualize the "hills and valleys" of the process. The model coefficients are estimated by least-squares regression, so RSM is essentially regression applied to a designed set of points.
Taguchi Methods
Genichi Taguchi developed a distinctive DOE philosophy focused on robustness — making a product or process insensitive to "noise" factors (environment, variation in materials, usage conditions) that cannot be controlled in the field. Taguchi designs use orthogonal arrays (a form of fractional factorial) to study many factors economically, and introduce the signal-to-noise (S/N) ratio as the performance measure to maximize. The guiding idea is the quality loss function: any deviation from target — not just falling outside specification — imposes a cost, so the objective is to minimize variation around the target.
Choosing a Design
| Goal | Recommended Design |
|---|
| Screen many factors quickly | Fractional factorial / Taguchi orthogonal array |
| Quantify main effects and interactions | Full 2k factorial |
| Optimize with curvature | Central composite or Box-Behnken (RSM) |
| Achieve robustness to noise | Taguchi parameter design |
Best Practices
- Randomize run order to neutralize time-related lurking variables.
- Replicate to estimate experimental error and test significance.
- Use center points to detect curvature before committing to a quadratic model.
- Screen first, optimize later — use cheap fractional designs to find the vital few factors, then characterize them in detail.
- Confirm — always run confirmation trials at the predicted optimum before locking in new settings.