Factor Selection and Level Determination in DOE: A Five-Step Method

A quality engineer at a plastic injection molding company spent three weeks conducting a round of DOE, involving three factors: melt temperature, holding pressure, and cooling time. Each group had three repetitions, the test sequence was randomized, and a center point was added. The report concluded, "Melt temperature has no significant impact on shrinkage, and cooling time is a minor factor."

However, on the production line, as soon as the mold temperature fluctuates by more than 5°C, the shrinkage defect rate doubles immediately.

During the post-mortem analysis, the issue quickly surfaced: in the experiment, cooling time was only set at three levels—10, 12, and 14 seconds—while the actual range allowed by the production line was 8 to 20 seconds. The three levels of melt temperature were all set in the middle of the process window, without utilizing the high and low ends.

The experiment itself was conducted correctly—randomization, repetition, and the center point were all in place. However, if the factors and levels are determined incorrectly, the data will not yield the right answers: you are not "validating" the process; you are merely confirming that "this small, flat area is indeed flat" within the most stable part of the process window.

For a company that can only schedule two rounds of experiments per week, this kind of waste is particularly costly.

1. Distinguish the Three Roles: Factors, Levels, and Responses

The inputs in a DOE are limited to a few elements. Distinguishing them clearly will make half of the subsequent choices straightforward.

There is a rule that almost resolves all debates: If it can be set, maintained, and recorded, it is a factor; if it can only be passively recorded, it is a covariate or a block; if it cannot even be recorded, it is noise—use randomization and repetition to average it out. Treating variables like shifts and environmental temperature and humidity as "factors" is the most common pitfall for beginners: the experiment may appear more comprehensive, but the conclusions will not be reproducible by different people or in different seasons.

2. Three Hard Rules for Level Determination

The specific values of the levels are more critical than the number of levels in determining the success or failure of the experiment. Follow these three rules in sequence.

Rule 1: The level range must cover the actual variation range in the field. Plot the actual variation range of the factors in the field, the specification range, and the adjustable range of the equipment on the same number line. The experimental levels should fall within their intersection, not in the center of the process window. If the range does not cover the actual variation, the conclusions cannot be extrapolated to the field—no matter how significant the results appear.

Rule 2: The range must be wide enough to make the effect "stand out above the noise." A common rule of thumb is that the expected effect size should be at least three times the measurement system variation (GR&R or standard deviation of repeatability). If the range is too narrow, the real differences between levels will be drowned out by measurement variation and material variation, leading to a series of conclusions with high p-values, making you mistakenly believe "these factors are not important."

Rule 3: Continuous factors should have at least three levels. Two levels can only show a straight line, not a curve—yet processes often exhibit non-linearity at the extremes of the window (material decomposition at high temperatures, flash at high pressures). Three levels (low—medium—high) plus a center point can help determine if there is curvature and whether the optimal value lies within the interval.

3. Five-Step Method: From Problem to an Executable Factor Table

Step 1: Determine the response first, not the factors. The response must meet three criteria: measurable, repeatable, and aligned with the problem to be solved. Using composite indicators like "appearance pass rate" as responses will always lead to unclear conclusions. After determining the response, confirm the measurement system—indicators with resolution lower than the expected effect will not reveal real differences in the data.

Step 2: List candidate factors, starting broad and narrowing down. The sources are not brainstorming sessions at the conference table but verifiable evidence: failure causes in the PFMEA, suspicious branches in the fishbone diagram, historical improvement records, and process window documents. List 10 to 15 candidates initially, allowing for "many listed but not adopted."

Step 3: Narrow down to 3 to 6 factors. Use engineering judgment, historical data, or a screening experiment to reduce the candidates to 3 to 6. Simultaneously handle noise: variables like shifts and batches should not be set as factors but should be blocked to prevent their interference from contaminating the factor effects.

Step 4: Determine the levels according to the three rules and document the rationale. This section must include a one-sentence explanation of "why these two endpoints" to ensure that the experiment can be reused in the future.

Step 5: Pre-enact and confirm. Before the formal experiment, conduct a trial run with an "extreme combination" (a combination where all factors are set at high levels) to confirm that the response has indeed changed and that the change can be detected by the measurement system. If the extreme combination shows no difference, go back and adjust the level range, rather than increasing the number of trials.

4. A Directly Applicable Factor and Level Determination Table

Fill out the table below, and the experimental plan will be largely set.

Three details are worth noting: the low level of holding pressure, 30MPa, is below the product specification requirement of 40MPa—this is not incorrect. The goal of the screening phase is to identify effects, and the level range should span critical points. Whether the results can be extrapolated to the specification boundaries is a matter for the next step; the upper limit of cooling time is constrained by the cycle time, indicating that the level range is often determined by production capacity rather than the process, which must be noted in the "Level Rationale" column; each row must have a rationale, and any blank cells will be reasons for future rework.

5. Four Common Pitfalls

• Treating noise as factors. Setting shifts, environmental temperature and humidity, and incoming material batches as factors will double the number of experiments and make the conclusions non-reproducible. They should be blocked.
• Determining level ranges arbitrarily. If the range is too narrow, no factors will be significant; if the range is too wide, the conclusions will fall into the unusable zone (material decomposition at high temperatures, flash at high pressures).
• Listing too many candidate factors at once. Eight two-level full factors result in 256 experimental groups, which most companies cannot schedule. Screen first, then optimize.
• Ignoring the resolution of the measurement system. If the expected effect is only 1.5 units and the measurement variation is 2 units, the data will not reveal real differences—conducting an MSA is more useful than running ten additional groups of experiments.

The quality of experimental design often does not depend on the number of experiments conducted or the complexity of the model used, but rather on the two hours before the experiment begins: whether the response was correctly chosen, whether the factors were correctly selected, and where the levels were set.

The quality of a DOE is largely determined before the experiment begins: selecting the right factors and setting the levels accurately ensures that the data will speak.

Knowledge code: 6.4.1

Version: v20261011

Author: QTank QTank is dedicated to providing systematic professional knowledge, methodologies, and practical tools to quality management practitioners, helping companies continuously improve their quality capabilities.