InsightFab
Knowledge Base/2^k-p Fractional Factorial Design: Managing Confounding
DOE6 min read

2^k-p Fractional Factorial Design: Managing Confounding

Have you ever faced persistently low product yield despite extensive Design of Experiments (DOE), only to have management react negatively to the results? This article addresses such a predicament, where parameters seem fine but yield remains significantly poor. The issue might not be the absence of DOE, but a lack of understanding of 'confounding,' a critical aspect of fractional factorial designs that can inadvertently blend factor effects and lead to misidentification of root causes.

That day, When the CPK 1.08 Report Came Out, The Air in the Room Froze for Three Seconds

I recall an instance where we conducted a large-scale DOE to improve the yield of a new product. The production line exhibited peculiar behavior during that period; machine A's parameters appeared normal, yet its product Cpk was only 1.08, significantly lower than machine B's 1.35. Upon seeing the report, the boss's face turned green, and he immediately summoned several senior engineers into the conference room, demanding, "What exactly is going on? How can the DOE results be like this?" Honestly, the atmosphere in the conference room at that moment was more oppressive than nitrogen in a cleanroom.

Where Was the Problem? It Wasn't That You Didn't Do DOE; It Was That You Didn't Understand 'Confounding'

We did conduct DOE, but the problem was that we actually used a 2^k-p fractional factorial design. This, in essence, is a 'resource-saving' approach. When you have too many factors, a full factorial design would require hundreds of runs, making it prohibitive in terms of time and cost. Therefore, we opt to run only a subset of combinations. But what is the cost of saving resources? It is that your factor effects might be 'confounded' with other factors or interaction effects.

In other words, the A factor effect you observe might actually include the interaction effect of B and C. For instance, when you see a Cpk of 1.08, you might assume it's a problem with machine A, but it could actually be a specific parameter of machine A performing particularly poorly under certain environmental temperatures. At this point, if you directly adjust machine A's parameters, you might not solve the problem and could even trigger new disasters.

How to Do It in Practice? Clearly Understand Your 'Generators'

In a 2^k-p design, the most critical aspect for managing confounding is your 'Generators'. Generators determine which factors will be confounded with which other factors.

For example, let's say we are conducting a 2^(5-2) design, meaning there are 5 factors, but we only run 2^3 = 8 experiments. The '-2' here indicates that we used two generators.

  1. Assume our generators are D = AB and E = AC.
  2. This means factor D will be confounded with the AB interaction, and factor E will be confounded with the AC interaction.
  3. Furthermore, if you multiply D = AB and E = AC, you get DE = A^2BC = BC. This indicates that the DE interaction will be confounded with the BC interaction.

So, the key is to know what your generators are, and then check their 'Resolution'. A higher resolution means a lower degree of confounding for your main effects or lower-order interaction effects.

  • Resolution III: Main effects are confounded with two-factor interactions.
  • Resolution IV: Main effects are not confounded, but two-factor interactions are confounded with two-factor interactions.
  • Resolution V: Neither main effects nor two-factor interactions are confounded.

Frankly, Resolution V is the best, but it also entails the least reduction in experimental runs compared to a full factorial. If the budget allows, it is usually recommended to aim for at least Resolution IV.

The Most Common Pitfall: Wanting to Save and Be Accurate, Ending Up With Neither

The most common pitfall I've observed is that people, when selecting a design, focus solely on minimizing experimental runs and thus choose a design with only Resolution III. Then, when the experimental results come out, they find a significant main effect and immediately rush to adjust its parameter. What's the result? Yield doesn't improve; instead, more strange problems arise.

It was only later discovered that the significant 'main effect' was actually confounded with a critical two-factor interaction. You adjusted the main effect but completely missed the truly influential interaction. This is like treating a headache when the cause is foot pain; naturally, it's ineffective.

In our Cpk 1.08 case, it was eventually found that a low-resolution design had confounded a parameter of machine A with the interaction of environmental humidity. On that particular day, the humidity was exceptionally high, which amplified the effect of that parameter, making it appear as a problem with machine A.

One Thing You Can Do Today

Next time you conduct a DOE, please first confirm the 'Resolution' of your fractional factorial design.

Want to try it yourself?

Every tool mentioned in this article is available on InsightFab — just upload a CSV to analyze.

Go to Tools

More in this category

DOE: Not Just Running Experiments, But Hitting the Most Targets with the Fewest Bullets

A full factorial experiment with four factors and three levels each requires 81 runs. Using a Taguchi L9 orthogonal array, only 9 runs are needed, saving 89%, while simultaneously estimating the main effects of all four factors without mutual contamination.

DOE Taguchi Method: Finding Optimal Parameters in 9 Experiments, Not by Luck

A full factorial design for 4 factors at 3 levels requires 81 runs, while a one-factor-at-a-time approach, though shorter at 36 runs, misses crucial factor interactions. The Taguchi L9 method offers a highly efficient alternative, completing the analysis in just 9 experiments and achieving an 89% reduction in experimental effort.

Response Surface Method: Finding the Optimal Process Sweet Spot

After DOE identifies the significant factors, RSM (Response Surface Methodology) helps pinpoint the precise parameter combination that optimizes the outcome. This marks a crucial transition from understanding 'which factors are important' to determining 'what settings are best'.