What Is DOE (Design of Experiments) and How Does It Improve Quality?

Design of Experiments (DOE) is a structured statistical method for planning, conducting, and analyzing tests that change multiple input factors simultaneously to identify their effects on an output response. It is a core tool in quality engineering, process optimization, and product development.

What It Is

DOE, also known as factorial experimentation, is a systematic approach to determine the relationship between factors (independent variables) and a response (dependent variable). Unlike one-factor-at-a-time (OFAT) testing, DOE varies several factors at once, allowing you to detect interactions—situations where the effect of one factor depends on the level of another. The method is fully described in standard textbooks such as Montgomery’s Design and Analysis of Experiments (Wiley), which covers full factorial, fractional factorial, and orthogonal designs.

How It Works / Key Steps

The most common DOE design is the 2^k factorial design, where k factors are each tested at two levels (often coded as −1 and +1). This requires 2^k experimental runs for a full factorial. For example, with 3 factors, you run 2³ = 8 experiments.

The core analysis uses contrasts and analysis of variance (ANOVA):

1. Define the problem and response – Choose a measurable output (e.g., tensile strength, defect rate).
2. Select factors and levels – Identify the key inputs and set low/high values.
3. Choose a design – Full factorial (all combinations) or fractional factorial (a subset) to save resources.
4. Run experiments in random order – Randomization reduces bias from uncontrolled variables.
5. Compute effects – For each factor, the effect is the difference between the average response at the high level and the average response at the low level. For a two-factor interaction, the contrast is computed from the four corner combinations.
6. Perform ANOVA – Partition total variation into components due to each factor, interactions, and error. Compare mean squares against the error mean square using an F-test to determine statistical significance.

A Worked Illustrative Example

Example data (illustrative only): Suppose a chemical process has two factors: Temperature (A) at 150 °C (−1) and 170 °C (+1), and Catalyst concentration (B) at 1% (−1) and 2% (+1). A full 2² factorial with two replicates gives the following yield (in %):

Run | A | B | Yield 1 | Yield 2

• 1 | −1 | −1 | 72 | 74
• 2 | +1 | −1 | 80 | 82
• 3 | −1 | +1 | 75 | 77
• 4 | +1 | +1 | 91 | 93

Main effect of A: Average yield at A = +1 minus average at A = −1.
Mean at A+ = (80+82+91+93)/4 = 86.5; Mean at A− = (72+74+75+77)/4 = 74.5.
Effect of A = 86.5 − 74.5 = 12.0.

Main effect of B: Mean at B+ = (75+77+91+93)/4 = 84.0; Mean at B− = (72+74+80+82)/4 = 77.0.
Effect of B = 84.0 − 77.0 = 7.0.

Interaction AB: Contrast = (average at A+B+ − average at A+B− − average at A−B+ + average at A−B−) / 2.
= (92 − 81 − 76 + 73)/2 = 8/2 = 4.0.

ANOVA would then compare these effects against the error variance (from the two replicates). In this illustrative case, all three effects appear large relative to the run-to-run noise, suggesting Temperature, Catalyst, and their interaction are all significant.

Common Pitfalls

• Confounding in fractional designs: When you use a fractional factorial, some effects are aliased with others. Always verify the resolution and defining relation before interpreting results.
• Ignoring randomization: Running trials in a fixed order can introduce hidden time-related bias.
• Overlooking interactions: A significant interaction means the main effects alone do not tell the full story—always plot and interpret interaction effects.
• Insufficient replicates: Without replication, you have no estimate of pure error, making ANOVA impossible.

---

To plan your own factorial experiments and compute effects quickly, try the free DOE tool at https://www.6sq.com/tools/doe/ — it guides you through design selection and analysis in a few clicks.