Too Many Factors, Can't Afford the Tests? —— Practical Partial Factorial Design: 16 Tests to Lock Down Key Factors

In deep quality improvement efforts, almost everyone hits the same wall: there are seven or eight influencing factors, each one a suspect, but the tests are too costly to conduct. A full factorial design with two levels and seven factors requires 128 tests, which the production line simply cannot accommodate. Falling back on the old "one factor at a time" method means missing out on interaction effects, leading to endless parameter adjustments with no improvement in defect rates. This article uses a real SMT welding improvement case to explain the partial factorial design, a strategy that "uses the fewest tests to screen out key factors": how it saves tests, what the trade-offs are, and how to avoid pitfalls.

1. Why Full Factorial Design Fails When Factors Increase

Let's do the math. A full factorial design tests all combinations of factor levels. With two levels, each additional factor doubles the number of tests: 3 factors require 8 tests, 4 factors require 16 tests, 5 factors require 32 tests, and 7 factors require 128 tests. For processes like welding, which involve temperature changes and material adjustments, each test can take over half an hour. 128 tests mean the production line would need to be dedicated to testing for a week, impacting output, delivery, and costs, which the boss would not approve.

However, reducing the number of factors is not a viable solution. You can't be sure which factors are critical. Cutting some based on experience might eliminate the real culprits. Using the single-factor rotation method to investigate each factor individually results in a linear increase in the number of tests, but it fails to detect interactions between factors. For example, adjusting A alone has no effect, adjusting B alone has no effect, but adjusting A and B together can double the effect. This "1+1 greater than 2" phenomenon is often the key to process breakthroughs, and the single-factor rotation method will never reveal it.

Partial factorial design is the solution to this dilemma: instead of 128 tests, it carefully selects 16 or 32 tests to estimate main effects and screen out key factors. The savings come at a cost: "information discount" — some effects may be confounded, a statistical term for effects being mixed together. Whether this confounding will cause issues depends on the design choice, which is the focus of the following sections.

2. Case Study: SMT Welding Defects in an Electronics Manufacturing Company, 7 Factors, 16 Tests

An electronics manufacturing company produces power control boards. For two consecutive months, customers have complained about weak and cold solder joints, with the defect rate in outgoing sampling inspections rising to 1.1%. The production line has to repair over 200 solder joints daily, and the quality department is under immense pressure. The process, equipment, and quality teams each have their own theories: the process team blames uneven solder paste printing thickness, the equipment team points to incorrect reflow oven temperature settings, and the quality department suspects clogged stencil holes. After a week of arguments, no one has any data.

The quality engineer takes the lead, narrowing down the suspect factors to seven: peak temperature, preheat time, chain speed, solder paste printing thickness, squeegee pressure, squeegee speed, and stencil cleaning frequency. Each factor has two levels, with ranges based on historical parameters. A full factorial design with 7 factors would require 128 tests, but the production line can only spare two days. The engineer selects a partial factorial design with 2^(7-3) = 16 tests, with a resolution of IV — main effects are not confounded with any other main effects, only with third-order and higher interactions, and second-order interactions are confounded with each other. For the screening phase, this resolution is sufficient and the risk is manageable.

The execution phase is where the real skill comes into play. The 16 tests are arranged in a random order to avoid time-related noise such as temperature and humidity drift, operator fatigue, etc. Three center point tests (all factors at their mid-levels) are inserted between every two tests to check for linearity and curvature in the process. Each test involves 5 boards, with 6 solder joints measured on each board, using the number of defects as the response variable to ensure statistical robustness. Over two days, a total of 480 solder joints are evaluated.

The data analysis uses an effect Pareto chart: the main effects of peak temperature, solder paste printing thickness, and squeegee pressure are significantly prominent, and the interaction effect between peak temperature and solder paste thickness also exceeds the significance threshold. The other factors' effects are drowned out by the noise. The conclusion does not align with any of the three teams' hypotheses: chain speed and preheat time are not significant, and the real solution is the combination of "peak temperature × solder paste thickness." The engineer optimizes production using the best levels of the significant factors: increasing the peak temperature by 8°C, increasing the solder paste thickness from 0.12mm to 0.14mm, and making minor adjustments to the squeegee pressure, while maintaining the non-significant factors at their low-cost levels. Five batches of validation tests show that the solder joint defect rate drops from 1.1% to 0.12%, and weak solder complaints are eliminated for the month. The repair station is reduced from 3 to 1 person. The entire improvement effort only required 16 tests and two days of production line time, at a cost of less than one-eighth of a full factorial design.

3. The Key to Partial Factorial Design: Choosing the Right Resolution

The most easily overlooked and crucial aspect of this case is the term "resolution IV." The essence of partial factorial design is to test only a subset of the full factorial combinations. Different subsets result in different confounding structures. Statistics uses resolution to classify designs: a resolution III design confounds main effects with second-order interactions, potentially misattributing interaction effects to main effects, leading to skewed conclusions. A resolution IV design confounds main effects only with third-order and higher interactions (which are rarely significant in engineering), and second-order interactions are confounded with each other, ensuring that main effects are sufficiently clear. A resolution V design clearly separates main effects and second-order interactions but requires more tests.

Here are three practical tips for frontline engineers:

1. Start with at least resolution IV in the screening phase; don't opt for resolution III to save tests — saving 8 tests and getting a potentially misleading effect chart is not worth it.
2. If you suspect a strong interaction between two factors, check the confounding relationship table during design to ensure that the interaction is not confounded with other effects. If necessary, upgrade the design to resolution V.
3. The "key factor list" from partial factorial screening is just a hypothesis, not a conclusion. It must be confirmed with a full factorial or validation test. This is the red line discussed in the next section.

4. Four Common Pitfalls, Any One Can Ruin the Effort

Reviewing numerous DOE projects, failures rarely occur due to statistical calculations but almost always due to the following four pitfalls.

The first pitfall: choosing too low a resolution. To reduce the number of tests to 8, selecting a resolution III design results in true interaction effects being confounded with main effects, leading to a completely wrong optimization direction that cannot be corrected later. Remember, the goal of screening design is to "not miss the real culprits," not to "minimize the number of tests."

The second pitfall: ignoring confounding and treating screening results as precise conclusions. The effects given by partial factorial design are "discounted." A significant factor could be due to its own effect or a combined effect with a confounded interaction. Without checking the confounding structure and conducting validation tests, directly locking in parameters for mass production based on screening results is like using discounted information as full value.

The third pitfall: not conducting validation tests. This is the most common cause of failure. After 16 tests screen out 3 factors, the improvement is declared complete, but the defect rate bounces back during mass production — because the screening tests have a small sample size and some factors are confounded, the conclusions are inherently uncertain. The correct approach is to conduct at least 3 to 5 independent validation batches using the optimal combination, confirm the effect with actual defect data, and then solidify the changes.

The fourth pitfall: using screening design for optimization. The purpose of partial factorial design is to "screen out key factors," not to "find the optimal parameters." For fine-tuning after identifying key factors, full factorial or response surface design should be used to find the optimal window, rather than repeatedly fine-tuning based on the results of 16 tests. Screening, full factorial, and response surface designs are sequential steps; the mindset of trying to achieve everything in one step is the biggest enemy of DOE.

5. When to Use Partial Factorial Design

Finally, let's clarify the applicable boundaries. Consider partial factorial design if all three conditions are met:

1. There are many candidate factors, typically more than 5, making full factorial tests impractical.
2. You are in the problem diagnosis screening phase, aiming to "identify key factors" rather than "find optimal values."
3. There is a suspicion of interaction effects, and the single-factor rotation method is insufficient.

Conversely, if there are only 2 to 4 factors and the test cost is manageable, go directly for a full factorial design to get the most complete information and the cleanest conclusions. After screening, use response surface design for fine optimization. Partial factorial design is not a panacea; it is the best solution for the specific scenario of "many factors, limited budget, and the need to quickly lock down the direction."

Returning to the electronics company's case: 16 tests and two days of production line time resulted in a nearly 90% reduction in defect rates. This success is not due to luck but to "saving where it counts and holding the line where it matters" — using resolution IV to control confounding risks, randomization and center points to ensure data quality, and validation tests to secure the conclusions. The value of DOE is not in "doing more tests," but in "using the right number of tests to get actionable information." The next time your production line is dealing with seven or eight factors, consider whether 16 tests are enough.

When there are many factors and a limited budget, use partial factorial design to screen directions, maintain resolution IV, and use validation tests to ensure the conclusions. Screening results are always hypotheses, not conclusions.

Knowledge code: 6.4.1

Version: v20260822

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