
Suppliers, resins, alloys, machine types, operators, test rigs. These factors shape your process results, but they do not sit on a number line. Supplier A is not half of Supplier B, and Machine 3 is not three times Machine 1.
This article shows you how to include qualitative factors in a designed experiment without inventing a numeric scale. You will see how level selection drives the scope of your conclusions, how indicator coding works, when effects coding makes more sense, and how supplier-by-process interactions reveal settings that shrink performance gaps. A labeled supplier example runs through the piece so the reasoning stays concrete.
Key Takeaways
- Treat supplier and machine labels as categories, not measurements.
- Choose levels that match the decision and limit claims to those tested.
- Use balanced runs where feasible and plan replication.
- Check supplier-by-process interactions before recommending settings.
- Exclude infeasible combinations during design, then validate the final setting.
Before the design mechanics, it helps to settle the coding question that trips up most teams.
How Do You Include Categorical Factors in DOE?

In a regression model with an intercept, represent a factor with k levels using k − 1 indicator columns. Each coefficient compares a level with the chosen reference when other model terms are held fixed, as shown in Penn State's dummy variable explanation. No ordering, no spacing, no pretend measurement scale.
The temptation is obvious. Someone types 1, 2, 3 into a spreadsheet for three suppliers because the software wants numbers.
That single choice forces a straight line through three labels and manufactures a trend that does not exist. The fitted slope would claim the step from Supplier A to B equals the step from B to C, which is meaningless.
With indicator coding, the omitted category provides a reference. Effects coding instead uses a sum-to-zero constraint; in a balanced design, its intercept corresponds to the grand mean. See Penn State's categorical coding notes.
- Indicator (dummy) coding: coefficients read as differences from the reference level.
- Effects coding: coefficients read as departures from the unweighted level mean; with balanced data, that matches the observed grand mean.
- Two-level categorical factors: a single column coded minus one and plus one works and fits standard two-level analysis.
You might be wondering which one to pick. If one supplier is your incumbent, indicator coding with that supplier as reference answers the business question directly.
Coding is only half the job. Level selection decides what your experiment can even speak to.
Choosing Factor Levels for Categorical Factor DOE
Every categorical level you include becomes part of the answer, and every level you leave out stays invisible. A test of three suppliers tells you nothing about a fourth. That boundary sounds obvious written down, yet reports routinely overreach it.
NIST's experimental design overview describes DOE as a way to plan studies for valid, objective conclusions. Level selection is where that objective becomes operational.
1. Pick Levels That Reflect Real Decisions
Include the suppliers you would actually buy from next quarter. Testing a vendor you cannot qualify burns runs for no payoff.
2. Keep the Level Count Honest
Each added supplier multiplies the combinations in a full factorial design. Choose the number of suppliers using the decision, available runs, and precision you need.
3. Hold Levels Stable Across the Study
Keep lots consistent where possible, and record any changes. If lots must change, consider blocking so lot variation does not obscure supplier comparisons.
4. Document What Each Label Contains
"Supplier B" might mean a different grade, a different plant, or a different shipping route. Write it down before the first run.
5. Match Levels to the Inference You Want
If the goal is a qualification decision, test incumbent plus candidates. If the goal is robustness, include the worst plausible material.
6. Avoid Convenience Levels
Material grades chosen because samples sat on the shelf produce conclusions nobody trusts. Scope the study to the decision instead.
With levels fixed, the next concern is whether comparisons stay fair across them.
Balanced Comparisons and Dummy Coding in Supplier Comparison Studies

Balance means each categorical level receives the same number of runs at each combination of the other factors. Balanced designs place the same number of observations in each treatment combination, as NIST's DOE glossary defines balance. Unequal counts can still be analyzed, but comparisons may have different precision.
Picture a study with three suppliers crossed with two process factors: temperature at 180 and 200 degrees C, and dwell time at 20 and 40 seconds. A full crossing gives 12 distinct conditions. Two replicates per condition yields 24 runs.
| Supplier group | Supplier | Temperature (C) | Dwell (s) | Runs |
|---|---|---|---|---|
| 1 | A (incumbent) | 180 / 200 | 20 / 40 | 8 |
| 2 | B | 180 / 200 | 20 / 40 | 8 |
| 3 | C | 180 / 200 | 20 / 40 | 8 |
Supplier A serves as the reference. Two dummy columns capture B and C, so the model reports how B and C shift the response relative to A at a given setting.
Randomize run order where practical. If day or operator effects require blocking, plan the blocks and randomize within them; see Air Academy's guide to blocking and randomization.
Historical data may omit combinations or mix supplier effects with other changes. A planned study can make the comparisons clearer. Main effects alone rarely close the supplier question, which brings interactions into focus.
Factor Interactions in DOE With Categorical Factors
NIST's definition of an interaction describes an effect that changes with another factor's level; a full factorial design can estimate included interactions. For categorical factors, an interaction means each supplier responds differently to the same process change. That is the finding worth chasing.
In the example above, suppose Supplier C produces high scrap at 180 C but matches Supplier A at 200 C. An overall supplier average could hide that difference across temperatures. Compare suppliers at each relevant setting.
- Supplier by temperature: different slopes per supplier reveal settings where material differences flatten out.
- Material by machine type: one alloy may run cleanly on a newer press and poorly on an older one.
- Operator by method: a new procedure may help less experienced staff more than veterans.
A supplier-by-temperature interaction can show whether the performance gap narrows at a different setting. Confirm that setting with follow-up runs before changing the process.
When an interaction matters, report the relevant supplier comparisons at specified settings alongside any overall main effect.
Plot the interaction with supplier labels on separate lines to make the conditional comparison clear.
Run counts grow quickly once categorical levels multiply, so feasibility deserves attention.
Feasible Run Combinations in Mixed Categorical Factor DOE

Not every combination of a categorical level and a continuous setting can be run. Some materials may not cure at low temperature, and some machines may not reach the target pressure.
Omitting impossible combinations from a standard layout can weaken the intended design. NIST discusses constrained design spaces and alternative designs.
Check feasibility with the process owner before the design is final. That check can prevent unusable planned runs.
Steps to Keep the Design Runnable
- List hard constraints first. Identify combinations that are physically or contractually impossible, and remove them before sizing the study.
- Size the design against available material. Confirm each supplier can deliver enough qualified stock for its full share of runs.
- Consider a fractional or screening approach. NIST notes that studies with many factors may require fractional factorial or screening designs to control run count.
- Consider optimal designs for constraints. Constrained or mixed-level settings often call for computer-generated designs rather than a standard catalog layout.
- Protect replication where variation matters. Plan enough true replication to estimate experimental error and support the comparisons that matter.
- Lock the measurement system before running. Measurement variation can obscure a genuine supplier difference; review the measurement system before the experiment. Air Academy's measurement system analysis guide explains the check.
When constraints remove planned combinations, revisit which effects the remaining design can estimate before reducing the run count.
AirAcad Training for Experiments With Categorical Factors
Handling qualitative factors well takes more than a coding rule, since level choice, balance, and interaction reading all affect the decision you hand to leadership. These four resources from Air Academy Associates target exactly those skills.
Mixed Factor/Mixed Level Designs Short Course
The Mixed Factor/Mixed Level Designs Short Course introduces experiments that combine different factor types and level counts. It is relevant when a sourcing comparison must also account for process settings. Explore the mixed-factor short course.
2-Level Designs: Full Factorial Short Course
The 2-Level Designs: Full Factorial Short Course covers main effects and interactions in complete two-level designs. That foundation helps with a two-supplier comparison across process settings. View the full factorial course.
Operational Design of Experiments Course
The Operational Design of Experiments Course covers planning, analysis, and interpretation, including mixed-factor/mixed-level designs and measurement system analysis. It supports teams that need to choose levels, run a study, and interpret results. See the operational course details.
Understanding Industrial Designed Experiments
Understanding Industrial Designed Experiments is a reference for planning and discussing industrial studies. Use it alongside your study objective when comparing full factorial, fractional, and screening layouts. Browse the designed experiments book.
A few interpretation habits separate credible categorical results from overreach.
Reporting Results From DOE With Categorical Factors
State the reference level in plain language. "Supplier B averaged 1.8 units lower than Supplier A at 200 C" beats a bare coefficient every time. Name the setting, because the difference may not hold elsewhere.
Avoid any phrasing that implies distance between labels. There is no halfway point between two machine types.
- Limit claims to the levels tested, and say so explicitly.
- Show interaction plots with labeled lines rather than numeric axes for the categorical factor.
- Report precision per level so unbalanced data does not mislead. Air Academy's DOE Pro XL software overview describes tools for analysis and plots.
- Run confirmation trials at the recommended condition before changing specifications.
An ordered categorical factor, such as three grades of a filler, may have a meaningful order. Even then, treat it as categorical first and check whether a trend shows up afterward.
Final Thoughts
Labels deserve labels, not invented numbers. If your team wants help designing a supplier or material study, request a consultation or view our training schedule, and we will point you toward the right format.
Air Academy Associates can help you plan a study that fits your process. Explore our Design of Experiments training options to learn more.
FAQs
How Do You Include Categorical Factors in a Designed Experiment?
Treat each category (e.g., Supplier A/B/C, Machine Type 1/2) as a distinct level in your DOE, then select a design that supports those levels (often factorial, fractional factorial, or mixed-level designs). In analysis, model them as categorical terms (indicator/dummy coding) rather than continuous variables, and plan runs to balance levels and avoid confounding. Air Academy Associates teaches practical setup and analysis so teams can design realistic experiments without forcing "fake" numeric inputs.
Why Should Supplier Codes Not Be Treated as a Numeric Scale?
Supplier codes are labels, not quantities—so assigning numbers (e.g., 1, 2, 3) incorrectly implies order and equal spacing, which can create misleading effects, trends, and predictions. Proper DOE treats suppliers as categorical levels so the model compares each supplier appropriately and supports valid conclusions. This is a common pitfall we address in our DOE training because it directly impacts decision quality.
How Are Categorical-Factor Interactions Interpreted?
An interaction means the effect of one factor depends on the category of another factor—for example, a material that performs best on one machine type but not another, or a supplier whose results change under different process settings. You interpret it by comparing differences between levels across the other factor (often via interaction plots and estimated effects), then confirming practical significance and follow-up runs if needed. Our instructors emphasize interpreting these interactions in operational terms so you can act confidently on the results.
