Design of Experiments is a method for changing several process inputs at once, in a planned pattern, so that you can measure how each input and each combination of inputs affects an output. It answers questions that trial and error cannot: which factors matter, how much, and whether they work together.
The payoff is efficiency and insight. A well-chosen design extracts the main effects and interactions from a small number of runs, and points to operating settings that are both better and more robust. This guide covers the vocabulary, the common designs, and a worked example you can reproduce by hand.
Open the DOE Quick Planner Read the Hypothesis Testing Guide
Why Design of Experiments Matters
Learns More From Fewer Runs
A planned design changes several factors together in a balanced pattern, so every run contributes to the estimate of every effect.
Finds Interactions
Real processes are full of factors whose effect depends on another factor. One-factor-at-a-time testing cannot see these; a factorial design can.
Replaces Trial and Error With Evidence
Instead of tweaking settings until something works, DOE gives a defensible map of which factors drive the output and by how much.
Supports Robust Settings
The result is a recommended operating window, not a single lucky setting, which is what makes a process stable in production.
Core Terms
| Term | Meaning |
|---|---|
| Factor | An input you deliberately vary, such as temperature, pressure or supplier. |
| Level | A specific setting of a factor. A two-level design uses a low (−) and a high (+) level. |
| Response | The output you measure, such as strength, defect rate or cycle time. |
| Main effect | The average change in the response when a factor moves from its low to its high level. |
| Interaction | When the effect of one factor depends on the level of another. |
| Replication | Repeating the same run to estimate random variation. |
| Randomization | Running the trials in random order so uncontrolled changes do not line up with a factor. |
| Blocking | Grouping runs by a known nuisance source, such as day or material lot, so it does not blur the factor effects. |
| Resolution | For fractional designs, how much main effects are confused (aliased) with interactions. |
Why Not Change One Factor at a Time?
Changing one factor while holding the others fixed feels rigorous, but it is inefficient and can mislead. It estimates each effect at only one setting of the other factors, so it cannot reveal interactions, and it uses many runs to estimate each effect once. A two-level factorial design changes all factors together in a balanced pattern. Every run then helps estimate every effect, and the interaction can be measured directly.
Common Designs
| Design | Runs | Use it when |
|---|---|---|
| Full factorial, 2 factors (2²) | 4 | Screening two factors and their interaction. |
| Full factorial, 3 factors (2³) | 8 | Three factors and all interactions are worth estimating. |
| Full factorial, 5 factors (2⁵) | 32 | Few factors and cheap runs; otherwise consider a fraction. |
| Half fraction, 5 factors (25−1, Resolution V) | 16 | Main effects and two-factor interactions are clear of each other. |
| Half fraction, 4 factors (24−1, Resolution IV) | 8 | Screening with main effects clear of two-factor interactions. |
| Fraction, 7 factors (27−4, Resolution III) | 8 | Early screening of many factors, accepting that main effects are aliased with interactions. |
| Response surface (central composite, Box-Behnken) | varies | Optimizing a few important factors after screening. |
Run counts are for a single replicate and exclude any center points. The DOE Quick Planner helps you build a run sheet and see how run count grows with factors.
A Practical DOE Sequence
- Define the response and how it is measured. Confirm the measurement system is capable first (see Measurement System Analysis).
- List candidate factors from process knowledge, a fishbone diagram or an FMEA, then pick the few worth testing.
- Choose levels far enough apart to produce a measurable effect but still safe and realistic to run.
- Select the design: full factorial for a handful of factors, a fraction for many.
- Randomize the run order and replicate where you can. Block on known nuisance factors.
- Run the experiment exactly as planned and record any anomalies.
- Analyze effects, interactions and residuals; confirm with ANOVA where appropriate.
- Confirm the predicted best settings with a few extra runs before changing the process, then update the control plan.
Worked Example: Adhesive Bond Strength
An assembly cell bonds a bracket with structural adhesive. The team studies two factors, cure temperature (60 °C or 80 °C) and clamp pressure (20 or 40 psi), with two replicates per combination and randomized run order. The response is peel strength in lbf. The data are illustrative.
| Run | Temp | Pressure | Replicate 1 | Replicate 2 | Mean |
|---|---|---|---|---|---|
| 1 | − (60 °C) | − (20 psi) | 60 | 64 | 62 |
| 2 | + (80 °C) | − (20 psi) | 69 | 71 | 70 |
| 3 | − (60 °C) | + (40 psi) | 67 | 69 | 68 |
| 4 | + (80 °C) | + (40 psi) | 87 | 89 | 88 |
Effects are the average response at the high level minus the average at the low level:
- Temperature effect = (70 + 88)/2 − (62 + 68)/2 = 79 − 65 = +14 lbf.
- Pressure effect = (68 + 88)/2 − (62 + 70)/2 = 78 − 66 = +12 lbf.
- Interaction (T × P) = (62 + 88)/2 − (70 + 68)/2 = 75 − 69 = +6 lbf.
What the team learns. Both factors matter, and they reinforce each other. Running hot and at high pressure gives the best strength (88 lbf), well above what either change alone would suggest. A one-factor-at-a-time test at low pressure would have reported a temperature effect of only 8 lbf and understated the payoff of combining the changes. Before adopting 80 °C and 40 psi, the team would run confirmation batches and check that the settings do not create other problems, such as bracket distortion.
Self-Assessment Questions
- Is the response measured with a capable gauge, and is the metric tied to a customer requirement?
- Are the factor levels wide enough to show an effect but safe to run?
- Was the run order randomized, and are nuisance factors blocked?
- Did we include replicates or center points to estimate noise?
- Did we check residuals and confirm the predicted optimum with extra runs?
- Are the winning settings written into the control plan and standard work?
Common Mistakes
Levels Too Close Together
If the low and high settings barely differ, the effect will be buried in noise and the experiment will find nothing.
Skipping Randomization
Running all the low settings first lets drift, warm-up or operator changes masquerade as a factor effect.
Testing Too Many Factors at Once
A large screening design can be efficient, but a poorly chosen one confuses effects. Pick the design deliberately and note what is aliased.
Not Confirming the Result
A model is a prediction. Run confirmation trials at the recommended settings before changing production.
Quick Reference
Before the Experiment
- Validate the measurement system.
- Pick factors from process knowledge, not guesses.
- Choose levels with a real spread.
- Pick the design and randomize the order.
After the Experiment
- Look at interaction plots before main effects alone.
- Check residuals for problems.
- Confirm the best settings with extra runs.
- Update the control plan and standard work.
Design of Experiments (DOE): Frequently Asked Questions
What is Design of Experiments in simple terms?
Design of Experiments is a structured way to test several process factors at the same time by running a planned set of combinations of their settings. By analyzing the results, you can tell which factors affect the output, how large each effect is, and whether factors interact, using far fewer runs than testing one factor at a time.
What is an interaction in DOE?
An interaction occurs when the effect of one factor depends on the level of another factor. For example, raising cure temperature might add a small gain at low clamp pressure but a large gain at high pressure. Interactions appear as non-parallel lines on an interaction plot and can only be estimated when factors are varied together.
When should I use a fractional factorial design?
Use a fractional factorial when you have many factors and running every combination is too costly. It runs a carefully chosen fraction of the full set, so some effects are aliased, meaning confused, with others. The design's resolution tells you which effects are clear of each other, and screening designs typically accept lower resolution before focusing on the few important factors.
Sources and Further Reading
- Douglas C. Montgomery, Design and Analysis of Experiments.
- George E. P. Box, J. Stuart Hunter and William G. Hunter, Statistics for Experimenters.
- NIST/SEMATECH e-Handbook of Statistical Methods, Process Improvement chapter.
- ASQ Certified Six Sigma Black Belt Body of Knowledge.