Data Collected, but Root Cause Still Elusive? —— Case Study Analysis of a Breakthrough in the A Phase of a Six Sigma Project at an Automotive Parts Company

Many Six Sigma projects fail not in the measurement phase but in the analysis phase. The M phase involves meticulous data collection planning and a validated measurement system, but once the project enters the A phase, the team is often at a loss with a screen full of data: fishbone diagrams are drawn, brainstorming sessions are held, parameters are tested, yet the defect rate remains unchanged. This article analyzes a real case: a die-casting porosity project at an automotive parts company that was stuck in the A phase for three months. The project finally broke through by reconstructing the analysis path, revealing four "analysis traps" that almost every DMAIC project encounters.

1. Case: Stuck in the A Phase for Three Months, Project on the Brink of Failure

An automotive parts company specializes in aluminum die-cast components, supplying transmission housings to vehicle manufacturers. Recently, the porosity defect rate of a key product increased from 1.5% to 4.8%, prompting the client to issue an ultimatum. The company assembled a core team to form a Six Sigma Black Belt project, aiming to reduce the defect rate to below 1% within three months.

The first two phases of the project went smoothly: the Define phase clearly defined the problem as "porosity defects in the die-casting process," with a comprehensive goal, scope, and benefit assessment. The M phase was even more rigorous—over 2,000 data points were collected according to the plan, the measurement system GR&R was validated, and a preliminary process capability study was conducted. The project team entered the A phase with high confidence, expecting to identify the root cause easily.

However, the project got stuck for three months. The team used a fishbone diagram to list over 30 potential factors, then began "testing" them: adjusting the injection speed today, changing the mold release agent tomorrow, and modifying the mold temperature the day after. Each adjustment was followed by a small batch trial. Despite over a dozen rounds of testing in three months, the defect rate fluctuated between 4% and 5%, showing no significant improvement. The atmosphere at project meetings became increasingly tense, and management began to question why the analysis was not yielding results despite the extensive data collection efforts.

2. Symptoms: Not Lack of Effort, but Incorrect Analysis Method

Reviewing the past three months, the project team was actually very diligent: they worked overtime to test parameters and filled several notebooks with records. However, breaking down their actions reveals the problem—they were not "analyzing" but "guessing."

The first symptom is that the data was left in tables, and analysis relied on "visual data pulling." Over 2,000 data points were only summarized once in the M phase, and no one systematically dissected the data in the A phase: no stratification by shift, no grouping by mold number, no comparison by material batch. The data was collected, but it was never "interrogated."

The second symptom is that experience replaced evidence. Among the 30 factors listed in the fishbone diagram, the experienced master insisted that "injection speed is the key." The project team focused on adjusting the injection speed, treating other factors as mere formalities. Intuition became a filter, prematurely narrowing the analysis scope.

The third symptom is single-factor trial and error. Each parameter was adjusted individually while others remained unchanged, and if there was no effect, the next parameter was tried. On the surface, this approach seemed rigorous, but in practice, it was both slow and blind—it could never detect the interaction effects where two parameters together produce a significant impact.

The fourth symptom is unverified conclusions. Each adjustment that "seemed to improve" was taken as evidence, but when the defect rate fluctuated, the team would start over. The entire analysis process lacked hypothesis testing and relied entirely on intuition. The three months of trial and error essentially turned the A phase into a game of chance.

3. Diagnosis: Four Traps Leading to a Dead End in Analysis

After the quality director intervened, the project team's analysis path was reviewed, and four typical traps were identified.

Trap one: Mean obscures stratification. The project team consistently used the average defect rate of all data for decision-making. However, when the data was stratified by shift, the truth emerged: the day shift had a defect rate of 2.9%, while the night shift had a defect rate of 6.1%, nearly double that of the day shift. The mean value conflated two vastly different subpopulations into a single number, much like averaging an elephant and an ant to get a "medium-sized animal." Without stratification, even more data would be a mess.

Trap two: Correlation mistaken for causation. The experienced master claimed that "faster injection speed leads to more porosity," and scatter plots seemed to support this. However, after stratification, the truth became clear: high injection speeds coincided with the night shift, during which mold temperatures were generally lower. The actual influencing factor might be mold temperature, with injection speed merely being a correlated variable. Mistaking correlation for causation can lead to misidentifying the true culprit and wrongly blaming the innocent.

Trap three: Ignoring interaction effects. Adjusting the concentration of the mold release agent alone showed no change in the defect rate, and adjusting the mold temperature alone also had minimal impact. However, when both factors were adjusted simultaneously, the defect rate fluctuated significantly—indicating an interaction effect. The single-factor trial and error method inherently cannot detect interaction terms, which is the fundamental reason for the project's lack of progress over three months.

Trap four: Experience is a clue, not a verdict. The experienced master's insights are valuable clues, but they must be statistically verified to become conclusions. The project team treated "what the master said" as "conclusive evidence," skipping the verification step, which is equivalent to letting the suspect judge their own case.

4. Breakthrough: Reconstructing the A Phase with "Stratification—Screening—Verification—Closure"

After identifying the root cause, the project team started over, reconstructing the analysis path in five steps.

Step one: Stratify before analyzing. All data was stratified by shift, machine, mold, and material batch, and stratification charts were drawn for comparison. Immediate clues emerged: the night shift defect rate was 6.1% compared to 2.9% for the day shift; the defect rate for two old molds was more than three times that of new molds; 80% of porosity defects were concentrated in the "night shift + old mold" combination.

Step two: Screen factors using data. The 30 factors listed in the fishbone diagram were quickly filtered based on "measurability, data availability, and variability." Only five factors could be verified with data: injection speed, mold temperature, mold release agent concentration, mold usage count, and shift. The remaining factors either lacked data support or could not be quantified in the short term, so they were put on hold.

Step three: Hypothesis testing to determine validity. For the five candidate factors, the team designed comparisons: a two-sample t-test to compare actual mold temperatures between the night and day shifts, and a chi-square test to confirm whether the defect rate was independent of the mold's age. Surprisingly, the injection speed was statistically insignificant—the master's intuition was refuted by the data. The significant factors were mold temperature and mold release agent concentration, as well as their interaction effect.

Step four: Small-scale DOE to confirm interaction. A 2² full factorial experiment was conducted for mold temperature and mold release agent concentration, with each combination repeated three times. The ANOVA results were clear: the main effect of mold temperature was significant, the main effect of mold release agent concentration was not, but the interaction term "mold temperature × mold release agent concentration" was highly significant—low temperatures amplified the effect of the mold release agent concentration, while high temperatures had almost no impact.

Step five: Mechanism closure. Statistical conclusions must be explainable in the field. The project team observed the night shift and found that the workshop's nighttime ambient temperature was low, and the mold temperature control system's actual heating capacity was insufficient. The mold temperature gauge showed 200°C, but the actual mold cavity temperature was around 185°C. At low temperatures, the mold release agent did not fully evaporate, and residual gases were entrapped in the aluminum melt during filling, causing porosity. The mechanism, data, and field observations aligned perfectly, locking down the root cause.

5. Effect: Defect Rate Reduced to 0.7%, A Phase Cycle Time Compressed from Three Months to Three Weeks

Measures were promptly implemented: insulation and compensation controls were added to the mold temperature control system, and preheating confirmation was mandatory before the night shift began. The mold release agent concentration was adjusted seasonally and included in the first article inspection. Three months later, the defect rate stabilized at 0.7%, and the project was closed on schedule, with benefits confirmed by the finance department.

More valuable than the numbers is the methodological sedimentation. The project team documented the "stratification—screening—verification—closure" process as the standard action list for the A phase, and subsequent projects followed this list. The average cycle time for the A phase was reduced from three months to three weeks. Within six months, the company used the same approach to solve two similar problems, and the Black Belts finally understood that the A phase is not about "not knowing how to use analysis tools" but about "not having an established analysis strategy."

6. Insights: When Stuck in the A Phase, Check the Strategy Before the Tools

This case offers three key insights for anyone working on DMAIC projects.

First, having data does not mean analysis has begun. Data must be stratified, compared, and verified to become evidence. The first step in the A phase should always be stratification—breaking down the data by shift, equipment, mold, material, and personnel. Most clues are hidden in the gaps of stratification.

Second, experience is a clue, but statistical verification is the verdict. The master's intuition can help narrow the scope, but only hypothesis testing can distinguish the true culprit from the scapegoat. Skipping the verification step is equivalent to leaving the project to chance. Correlation does not equal causation, and interaction effects can never be detected with single-factor trial and error—these lessons were learned over three months.

Third, the output of the A phase is not a pile of charts but a "statistically significant + mechanistically coherent" root cause conclusion. If the statistical significance does not align with the field observations, the data should be questioned. If the mechanism is reasonable but not statistically significant, the sample should be questioned. When both align, the improvement measures are worth the investment. A slower analysis phase can lead to a faster improvement phase.

Collecting data is just the beginning of analysis. Completing the four steps of stratification, screening, verification, and closure will bring the root cause to the surface.

Knowledge code: 6.1.1

Version: v20260820

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