Design of Experiments (DOE) Quick Reference & Sample Layouts

Purpose and audience

This quick reference helps improvement practitioners, team leads, and technicians run small, well-formed experiments that produce usable causal insight. It emphasizes practical planning, simple layouts, basic analysis you can do in a spreadsheet, and common pitfalls that waste time or produce misleading conclusions.

When to use DOE

• When you have several controllable factors and want to discover which matter and how they interact.
• To optimize a process step or product feature with a small number of runs.
• To trade off cost, quality, speed, or other measurable responses systematically rather than by trial-and-error.
• Not ideal when the process is highly unstable, when effects are dominated by noise that cannot be controlled, or when only one factor is of interest and a simple one-factor-at-a-time check suffices.

Key terms (plain language)

• Factor: a variable you choose to change (temperature, speed, tool type).
• Level: the setting for a factor (low/high, or numeric values).
• Response: the measured outcome (defects, cycle time, yield).
• Run: a single trial with a specific combination of factor levels.
• Replication: repeating runs to measure variation.
• Randomization: running trials in random order to avoid systematic bias (time, operator, shift).
• Blocking: grouping runs that share a nuisance condition (machine, shift) and analyzing blocks to reduce error.
• Alias/Confounding: when two effects cannot be separated in the chosen design (common in fractional designs).
• Resolution: a simple way to describe which effects are aliased (higher resolution = fewer confounded important effects).

Quick planning checklist

1. State the practical question and pick a measurable response tied to business value.
2. List candidate factors (2–6 recommended for small designs) and choose 2 levels each (low/high) that are safe and meaningful.
3. Decide on the design size (full factorial vs fractional) based on runs you can afford and whether interactions matter.
4. Plan randomization and blocking (who, when, which machine) to avoid bias.
5. Decide on replication and any center points (for detecting curvature).
6. Prepare a simple run sheet and data-capture method (spreadsheet or interactive form) before starting.

Simple designs and sample layouts

2^2 full factorial (2 factors, 4 runs)

Use when you have two factors and can run four trials. This shows main effects and their interaction.

2^3 full factorial (3 factors, 8 runs)

Shows main effects and all two-way and three-way interactions. Useful when eight runs are practical.

2^(3-1) half-fraction (4 runs) — when runs are limited

A half-fraction reduces runs but creates aliasing. A common generator is I = A·B·C, which means ABC is aliased with the intercept and some effects are confounded. Use fractional designs only when you accept the aliasing or prior knowledge says certain interactions are negligible.

Blocking and randomization (practical rules)

• Randomize the run order within the constraints of safety and setup time to avoid trends or shift effects.
• If a known nuisance (machine, operator, shift) will affect results, include it as a block and assign runs so each block has a balanced mix of factor combinations.
• When blocking adds a column to your layout, analyze it to remove block-to-block variation before estimating factor effects.

Simple analysis steps you can do in a spreadsheet

1. Compute the average response at the high (+) and low (−) level for each factor.
2. Estimate the main effect: effect = (average at high) − (average at low). This gives the practical change in response when the factor moves from low to high.
3. For interactions, compare averages of combinations in the same way (or use linear regression with coded levels +1/−1 to estimate coefficients).
4. Create a Pareto chart of absolute effects (largest to smallest) to focus on the biggest contributors.
5. Check assumptions: include residual plots or simple checks for outliers and trends. If you added center points, test for curvature before fitting only linear effects.
6. Confirm practical significance: consider measurement noise, cost, and feasibility before changing the process.

Regression and ANOVA (basics)

When you want a single method that handles main effects, interactions, blocks, and continuous factors, fit a linear regression model using coded factor levels (−1, +1). ANOVA partitions variance and provides F-tests if you have replication. For small practical experiments, effects and Pareto ranking are often sufficient to decide next steps.

Templates — run sheet

Use this simple table to record runs and observations.

Common pitfalls and how to avoid them

• Running without randomization — risk: time or operator trends masquerade as effects. Fix: randomize order or include blocking.
• Insufficient replication — risk: you cannot distinguish effect from noise. Fix: add repeats or use tighter measurement methods.
• Misusing fractional designs when key interactions are unknown — risk: confounded conclusions. Fix: use a full factorial or a higher-resolution fractional design if interactions might be important.
• Poorly chosen factor ranges — risk: no observable effect or unsafe conditions. Fix: select realistic, safe, and work-relevant levels.
• Focusing on statistical significance alone — risk: small statistically significant effects that are not practically useful. Fix: always interpret effect size in business or process terms.

Practical next steps after analysis

1. Implement the most promising changes in a pilot or controlled rollout and measure the impact over time.
2. If results show curvature or you need an optimum inside the tested region, run a response-surface experiment (e.g., central composite design) or follow-up factorial with narrower ranges.
3. Document findings, decisions, and updated standard work so improvements stick.

Where this fits in continuous improvement

DOE provides a disciplined way to learn how multiple factors influence outcomes. Use it when your team wants to move from anecdote to causal insight and to make changes that are repeatable and measured.

References and tools

• Use a spreadsheet with coded levels (−1, +1) and simple formulas for averages and differences to start.
• Consider readily available DOE tools or statistical packages when you need ANOVA, regression diagnostics, or response-surface designs.

Keep this sheet handy when planning experiments. If you want, we can turn the run sheet into an interactive form that captures runs, stores data, and produces basic effect calculations automatically.