Practical Application of Regression Analysis in Six Sigma DMAIC: A Comprehensive Case Study from Variable Screening to Process Optimization

1. Introduction: When Experience and Intuition Fail in Complex Processes

A leading automotive electronics company, Dingxin Electronics (a pseudonym), faced a challenging issue in 2025. The final test pass rate of their BMS (Battery Management System) control module had been consistently hovering around 91%, falling short of the company's target of 98% by 7 percentage points. More critically, there were over 30 potential factors affecting the pass rate: peak reflow temperature, furnace airspeed, solder paste printing pressure, SMT machine placement speed, incoming PCB batch, environmental temperature and humidity, etc. The engineering team spent three months investigating each factor based on experience, but the pass rate only improved by less than 2 percentage points.

The Quality Director candidly stated at a Six Sigma project review meeting: "Our current approach is like feeling an elephant in the dark—each engineer claims their factor is critical, but no one can explain the relationships between these factors or the extent of their individual impacts. We need a method to clearly articulate the relationships between over 30 variables and the final quality."

This is precisely where regression analysis (Regression Analysis) plays a central role in Six Sigma projects. As one of the most powerful statistical tools in the Analyze phase of the DMAIC framework, regression analysis does more than just "look at correlations." It establishes mathematical models to quantify the functional relationships between multiple input variables (X) and output responses (Y), helping the team to precisely identify the key few factors from a large dataset and predict process performance under different factor combinations.

This article will present a complete Six Sigma project case from Dingxin Electronics, systematically demonstrating the practical application of regression analysis in the DMAIC framework—from data collection design, model building and diagnosis, to the final process optimization.

2. Case Background: Why Can't the BMS Module Welding Pass Rate Improve?

Dingxin Electronics' BMS control module uses a double-sided reflow soldering process, with key operations including solder paste printing, component placement, reflow soldering, and AOI optical inspection. At the project's inception, the first-pass rate of AOI inspection was approximately 91%, with the main defect types being: BGA voids (42%), tombstoning (28%), solder ball splashing (18%), and minor defects such as insufficient solder and short circuits.

The project team, led by Green Belt Engineer Chen, included one process engineer, one equipment engineer, and one quality engineer. In the Define phase, the team defined the project Y as "first-pass rate of AOI after reflow soldering," with a current baseline of 91% and a target of ≥98%. The project scope was limited to the BMS product family on a main SMT production line.

In the Measure phase, the team completed MSA analysis, confirming that the GR&R of the AOI inspection equipment was 8.5% (acceptable). They also collected process parameter data and corresponding pass rate data from 30 consecutive production batches. After brainstorming and using a C&E Matrix, the team initially identified 12 potential key factors (X1~X12).

3. The Three-Step Practical Approach to Regression Analysis

(1) Step One: Experimental Design and Data Collection—the Foundation of Regression Analysis

Many quality teams make a fundamental mistake when applying regression analysis: directly using historical production data for regression modeling. The biggest issue with historical data is multicollinearity among factors—such as the peak reflow temperature and furnace airspeed, which are often adjusted together in actual production, leading to high correlation between the two variables and making it difficult for the model to accurately separate their independent effects.

The Dingxin Electronics team used a fractional factorial design (Fractional Factorial Design) to generate modeling data. The specific plan was as follows:

• Screening Phase: For the 12 potential factors, a 2^(12-8) resolution IV fractional factorial design was used, running 16 experimental combinations, each repeated 3 times, resulting in 48 data sets. A resolution IV design ensures that main effects are not confounded by second-order interactions, sufficient for factor screening.
• Optimization Phase: After identifying 4 significant factors (peak reflow temperature X1, soak zone heating rate X2, solder paste printing pressure X3, and incoming PCB storage time X4), a central composite design (Central Composite Design, CCD) was used to generate 30 experimental data sets for fitting a second-order regression model.

The team emphasized two data quality principles: first, all experimental combinations were conducted under normal production conditions (using mass production batches), ensuring that the experimental results were practically reproducible; second, the AOI equipment was calibrated before and after each experiment to eliminate additional variation caused by measurement drift.

(2) Step Two: Model Fitting and Diagnosis—From Numbers to Insights

With reliable data, the team entered the core modeling phase. Using Minitab software, with the pass rate (Y) as the response variable and the four significant factors as independent variables, they first fitted a first-order linear regression model:

Y = β₀ + β₁X₁ + β₂X₂ + β₃X₃ + β₄X₄ + ε

The model output showed R²=0.762 and adjusted R²=0.731, indicating that the four factors could explain about 73% of the pass rate variation. However, residual analysis revealed a critical issue: the scatter plot of residuals versus fitted values showed a clear "funnel shape"—residuals were more dispersed at higher fitted values, suggesting heteroscedasticity and possibly missing important interaction terms or quadratic terms.

The team then fitted a complete model including all second-order terms (interaction and square terms):

Y = β₀ + ΣβᵢXᵢ + ΣβᵢⱼXᵢXⱼ + ΣβᵢᵢXᵢ² + ε

The variable selection results from stepwise regression were encouraging: the model's R² increased to 0.913, and the adjusted R² to 0.887. More importantly, the analysis revealed two highly valuable relationships:

First, there is a significant interaction between peak reflow temperature (X₁) and soak zone heating rate (X₂) (p=0.003). In simpler terms, the "combination" of peak temperature and heating rate is more important than their individual effects—when the heating rate is low, the impact of peak temperature on the pass rate is minimal; but when the heating rate is high, even a small change in peak temperature can cause significant fluctuations in the pass rate. This finding explains why previous adjustments to peak temperature alone were inconsistent—engineers did not simultaneously consider the setting of the heating rate.

Second, the effect of solder paste printing pressure (X₃) on the pass rate follows an inverted U-shaped curve (X₃² term p=0.008). Low printing pressure results in insufficient solder transfer, leading to insufficient solder (BGA voids); high printing pressure causes solder to be squeezed out, leading to solder balls and bridging. Engineers had previously set the pressure at 8.5N based on experience, but regression analysis showed that the optimal value should be around 10.2N.

(3) Step Three: Model Validation and Residual Diagnosis—Rejecting "Pseudo-Regression"

After confirming the model's goodness of fit, the team conducted rigorous model validation, a crucial but often overlooked step in many Six Sigma projects:

1. Residual Normality Test: Using the Anderson-Darling test, the p-value was 0.214 > 0.05, indicating that the residuals follow a normal distribution, which aligns with the basic assumptions of regression analysis.

2. Residual Independence Test: The Durbin-Watson statistic was 1.92, close to the ideal value of 2.0, indicating no significant autocorrelation in the residuals.

3. Homogeneity of Variance Test: The scatter plot of residuals versus fitted values showed a random distribution with no clear pattern, indicating that the heteroscedasticity issue had been resolved.

4. Multicollinearity Diagnosis: All independent variables had VIF (Variance Inflation Factor) values less than 3.5, far below the warning line of 10, indicating no severe multicollinearity issues.

5. Model Extrapolation Validation: The team predicted the pass rates for 5 new process parameter combinations that were not used in the model building. The average prediction error was only 1.2 percentage points. This step is crucial—a model that performs well on training data must also maintain accuracy in actual predictions to be considered a "reliable model."

4. From Model to Improvement: How Regression Analysis Drives Process Parameter Optimization

After establishing the regression model, the team used the Response Optimizer to find the process parameter combination that maximizes the pass rate:

The model predicted a pass rate of 98.6% (confidence interval 95%~99.8%) under this combination.

The team conducted a two-week validation production run according to the optimized parameters, producing 12 batches and approximately 960 products. The results were as follows:

• First Week (6 batches): Average pass rate 97.8%, highest batch 99.2%
• Second Week (6 batches): Average pass rate 98.4%, highest batch 99.5%
• Overall Average for Two Weeks: 98.1%, meeting the project target

More importantly, the key factors and interaction relationships identified by regression analysis provided the team with interpretable process knowledge rather than just a set of "parameter tuning results." The process engineers understood why the heating rate needed to be adjusted in conjunction with the peak temperature—because a faster heating rate ensures that the flux in the solder paste is fully volatilized before entering the reflow zone, reducing internal void defects; and a higher peak temperature ensures that BGA solder balls are fully melted and self-aligned. This understanding of the causal mechanisms transformed the process parameter settings from "trial and error based on experience" to "scientific design."

5. Case Two: Application of Regression Analysis in a Service Industry Six Sigma Project

In addition to manufacturing, regression analysis also plays an indispensable role in Six Sigma projects in the service industry. The following case comes from a commercial bank's credit card center.

The bank launched a Six Sigma project in 2024 to "reduce the credit card approval cycle." The current average approval cycle was 7.2 working days, with a target of compressing it to within 3 working days. The project Y was defined as "the total time from customer application submission to approval completion (hours)."

In the Analyze phase, the team identified 8 variables that might affect the approval cycle: application channel (online/offline), completeness of submitted documents (score 0～100), credit inquiry response time (seconds), number of manual reviews, application time (weekday/weekend), applicant age, application amount, and historical card ownership (yes/no).

Using historical data (approximately 12,000 records over the past 6 months) for multiple linear regression analysis, the results showed:

• Significant Factors (p＜0.05): Completeness of submitted documents (β=-0.48, p＜0.001), number of manual reviews (β=1.62, p＜0.001), credit inquiry response time (β=0.31, p=0.002), and application channel (β=0.22, p=0.015).
• Model R²=0.784, indicating that the four factors can explain 78% of the approval cycle variation.
• Among these, the number of manual reviews had the largest standardized regression coefficient (β=1.62), meaning that each additional manual review extends the approval cycle by about 1.6 days on average—this finding drove the digital transformation of "pre-rule automatic review."
• Completeness of submitted documents increased by 10 points, reducing the approval cycle by about 0.48 days—this directly led to the launch of the "AI pre-review + document completion reminder" feature.

In the Improve phase, the project implemented four improvements:

1. Launch an AI pre-review system to automatically assess document completeness and guide customers to complete them in real-time.
2. Optimize the credit inquiry interface to reduce the average response time from 4.2 seconds to 0.8 seconds.
3. Establish an automatic approval rule engine to reduce the proportion of applications requiring manual review from 42% to 15%.
4. Digitize the offline application process to reduce the paper flow.

Three months after implementation, the average approval cycle was reduced from 7.2 working days to 2.8 working days, and the customer complaint rate decreased by 63%. The value of regression analysis here lies not only in "identifying which factors are important" but also in quantifying the business impact of each factor—this helped the project team and business departments reach a consensus on resource allocation priorities, avoiding the situation where "every department thinks their process is the most critical."

6. Common Traps and Countermeasures in Regression Analysis in Six Sigma Projects

The cases from Dingxin Electronics and the bank credit card center demonstrate the correct way to apply regression analysis. However, in actual Six Sigma projects, many teams fall into the following five typical traps:

Trap One: Over-reliance on P-values, Ignoring Practical Significance. In large samples, even minor effects can show p＜0.05. It is essential to always consider the size of the regression coefficients (practical effect size) and the change in R², not just the p-values. For example, in the bank case, although the application channel was significant (p=0.015), its standardized regression coefficient was only 0.22, indicating a very limited actual business impact and should not be a priority for improvement.

Trap Two: Directly Modeling with Historical Data, Ignoring Data Quality. Historical data often has issues such as multicollinearity, uncalibrated measurement systems, and unhandled outliers. The Dingxin Electronics team chose to use DOE to generate modeling data rather than directly using historical production data, which was the primary guarantee for the success of the regression analysis.

Trap Three: Overfitting Due to Excessive Model Complexity. Especially when there are many factors and limited data, models with numerous interaction and quadratic terms may perform well on training data but poorly on new data. It is always necessary to use adjusted R², predicted R², Cp statistics, or cross-validation to evaluate the model's generalization ability.

Trap Four: Ignoring the Basic Assumptions of Regression Analysis. Linearity, independence, normality, and homoscedasticity—these four assumptions are essential. The Dingxin team systematically tested each assumption during the model diagnosis phase and confirmed it with both residual plots and statistical tests. This rigorous attitude is the cornerstone of model reliability.

Trap Five: Ending the Project After Model Building, Without Converting to Operational Process Windows. The ultimate value of regression analysis is not just obtaining a set of coefficients but converting the model into process specifications that engineers and operators can directly use. In the Control phase, the Dingxin team wrote the optimized parameters into the standard work instructions (SOS) and set parameter range alarm limits in the MES system, achieving a closed loop from model to the production floor.

7. Conclusion: Regression Analysis—From "Guessing" to "Data-Driven"

Reflecting on the Six Sigma project for Dingxin Electronics' BMS module, the most impressive transformation was not the increase in the pass rate from 91% to 98%, but the fundamental change in the engineering team's mindset. Before the project, the logic for setting process parameters was "senior engineer's experience + asking peers in WeChat groups"; after the project, the team learned to speak with data—every parameter adjustment is supported by a regression model, and every improvement measure has a quantitative basis.

This is the true value of regression analysis in the Six Sigma DMAIC framework: it is not a complex statistical black box but a mathematical language that converts experience and intuition into verifiable, reproducible, and transferable knowledge. For any quality team facing a multi-variable complex system, mastering regression analysis is not an elective but a mandatory course—because when intuition fails in complex processes, only data can illuminate the path.

Regression analysis is not the exclusive domain of statisticians but the sharpest data tool in the hands of every Six Sigma practitioner—learn to use it, let the data speak, and make improvements predictable.

Knowledge code: 6.1.1

Version: v20260729

Author: Quality Think Tank The 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.