7. DMAIC: Analyze and Root-Cause Validation

Introduction and Context

The Analyze phase of DMAIC converts an observed performance problem into an evidence-based explanation. Define and Measure establish what the problem is, where it occurs, how large it is, and whether the measurement system is reliable. Analyze asks the more demanding question: Why does the problem occur, and which causes are sufficiently supported by evidence to justify action?

This distinction is central to Six Sigma. Teams frequently move from a visible symptom to an attractive explanation without testing the explanation. A high defect rate may be attributed to operator carelessness, a delay to inadequate staffing, or customer dissatisfaction to poor communication. Such explanations may be plausible, but plausibility is not validation. Root-cause analysis requires a disciplined movement from data exploration to causal reasoning and, where possible, statistical confirmation.

The Analyze phase therefore combines qualitative and quantitative methods. Pareto analysis, stratification, cause-and-effect diagrams, and the 5 Whys help teams organize knowledge and generate hypotheses. Graphical analysis, contingency tables, regression, and analysis of variance help determine whether suspected factors are associated with or predict variation in the critical-to-quality characteristic. The strongest conclusions arise when process knowledge, data patterns, and statistical evidence converge.

Analyze is not merely a search for a single cause. Complex processes often contain several contributors, interactions, measurement effects, and special conditions. The practical purpose is to identify the vital few causes that can be controlled and that materially influence the response of interest.

Learning Objectives

After studying this chapter, the learner should be able to:

1. Explain the purpose and logic of the Analyze phase within DMAIC.

2. Distinguish symptoms, contributing causes, root causes, and control factors.

3. Apply Pareto analysis, stratification, cause-and-effect diagrams, and 5 Whys to generate and prioritize hypotheses.

4. Select suitable graphical and statistical methods for continuous, categorical, and count data.

5. Interpret regression, ANOVA, contingency analysis, and confidence intervals in a Six Sigma context.

6. Distinguish association from causation and recognize confounding, interaction, and model limitations.

7. Validate root causes using practical, statistical, and process evidence.

8. Translate analytical findings into improvement priorities while maintaining ethical and professional standards.

The Logic of Root-Cause Validation

A useful analytical chain is:

Problem statement → response measure → suspected factor → mechanism → evidence → validated cause → improvement action.

The response is the outcome being improved, such as cycle time, defect probability, tensile strength, or customer waiting time. A factor is a process input or condition that may influence the response. A mechanism explains how the factor produces the observed effect. For example, excessive injection temperature may reduce viscosity, increase flash formation, and consequently raise the defect rate. The mechanism makes the hypothesis technically meaningful rather than merely correlational.

A suspected cause becomes a validated root cause when credible evidence shows that changing or controlling the factor is expected to change the response, and when alternative explanations have been reasonably examined. Statistical significance is useful but insufficient. A very small effect can be statistically significant with a large sample, while an operationally important effect may fail to reach significance in a small sample. Six Sigma decisions should therefore consider effect size, confidence intervals, practical thresholds, process knowledge, and feasibility of control.

Several distinctions are important:

• A symptom is the observable problem, such as late delivery.
• A proximate cause is an immediate contributor, such as an incomplete order record.
• A root cause is an underlying condition whose removal or control prevents recurrence.
• A common cause is a source of routine process variation built into the system.
• A special cause is an unusual, identifiable source of variation.
• A controllable factor can be adjusted by the organization; an uncontrollable or noise factor cannot be eliminated but may be managed.

The term root cause should not be used mechanically. A chain may contain several legitimate levels of causation. Stopping at “the operator made an error” is usually weak analysis because it ignores training, interface design, workload, standard work, and error-proofing. A mature analysis examines why the error was possible and why the system did not detect it.

Beginning with Data: Pareto Analysis and Stratification

Pareto analysis

Pareto analysis ranks categories according to their contribution to a total effect. A defect Pareto may rank defect types, while a complaint Pareto may rank complaint categories. The bars are arranged from greatest to least frequency, often with a cumulative-percentage line. The purpose is prioritization: it helps the team focus on categories that offer the greatest opportunity for improvement.

A Pareto chart must be constructed carefully. Categories should be mutually exclusive and collectively meaningful. The denominator should be clear, and the period, product family, location, and unit of analysis should be stated. Frequency alone can mislead. A rare defect with severe safety consequences may deserve priority over a frequent cosmetic defect. Teams should therefore consider frequency, severity, cost, and detectability rather than treating the 80/20 rule as a law.

A Pareto chart identifies where to look; it does not prove why the problem occurs. If Line A accounts for most defects, this may reflect greater production volume rather than poorer performance. Rates per opportunity, per unit, or per hour are often more informative than raw counts.

Stratification

Stratification separates data into meaningful subgroups so that hidden patterns become visible. Common strata include machine, shift, operator group, supplier, material lot, product model, geographic region, customer segment, and time period. A process average can conceal large differences between strata. For example, an overall first-pass yield of 96% may comprise 99% on one machine and 90% on another.

Stratification is especially useful before formal modeling because it reveals possible confounders and suggests interactions. However, excessive subgrouping can produce unstable small samples and false patterns. Each stratum should have a process-based rationale, and comparisons should account for exposure or volume.

A practical sequence is to examine the problem overall, stratify by major process dimensions, compare rates or distributions, and then investigate whether the observed differences persist after adjusting for other factors. Time-based stratification is also important: a gradual deterioration suggests wear, drift, or changing material, whereas an abrupt shift suggests a special cause.

Generating Causal Hypotheses

Cause-and-effect diagrams

The cause-and-effect diagram, often called a fishbone or Ishikawa diagram, organizes possible causes under broad categories. In manufacturing, the familiar categories are often Manpower, Machine, Method, Material, Measurement, and Environment. In service processes, categories may be people, policy, technology, information, workflow, and customer demand. The categories are not rules; they are prompts for systematic thinking.

A sound diagram is built from a clearly defined effect and supported by process knowledge. Causes should be written specifically. “Poor quality” is not a cause; “calibration interval exceeds the supplier recommendation” is testable. The team should distinguish possible causes from confirmed causes by using separate labels or evidence markers. Brainstorming creates hypotheses, not conclusions.

The 5 Whys

The 5 Whys method repeatedly asks why an observed condition occurred. Its value lies in moving from an event to conditions that allowed the event. The number five is not mandatory; some problems require three questions and others require seven. The questioning should stop when the team reaches a cause that is specific, controllable, supported by evidence, and connected to a preventive action.

For example:

• Why was the shipment late? The order was released after the promised date.
• Why was it released late? Credit approval remained pending.
• Why did approval remain pending? The system did not flag incomplete financial data.
• Why did the system not flag it? The validation rule was not included for this customer class.
• Why was the rule absent? Ownership of exception requirements was undefined during system design.

This chain is useful only if each link is verified. Teams should avoid blaming individuals, assuming that one chain explains all incidents, or forcing a linear explanation onto a multifactor problem. A fault-tree or causal tree may be preferable when several conditions must combine to produce the failure.

Graphical Methods for Exploring Relationships

Graphs often reveal structure before formal tests are applied. The correct graph depends on the type of response and factor.

A histogram displays the distribution of a continuous measure and can reveal skewness, multiple populations, truncation, or unusual tails. A box plot compares medians, spread, and outliers across groups. Run charts and time-series plots show trends, cycles, shifts, and serial dependence. A scatter plot examines the relationship between two quantitative variables; the pattern may be linear, curved, clustered, or absent. A dot plot is valuable for small samples because it preserves individual observations.

Graphical interpretation requires caution. A few extreme observations can create an apparent relationship, and overlapping groups may conceal an important effect. Axes should not exaggerate differences, and units must be visible. Outliers should be investigated for data errors, measurement problems, or legitimate unusual conditions; they should not be removed merely because they weaken a desired conclusion.

A scatter plot with a fitted line can support a regression hypothesis, but correlation does not establish causation. A third factor may affect both variables. For instance, overtime hours and errors may be positively correlated because both increase during peak demand; demand intensity may be the more fundamental driver.

Statistical Validation of Suspected Causes

Comparing continuous outcomes across groups: ANOVA

Analysis of variance (ANOVA) evaluates whether the means of a continuous response differ across two or more groups. It partitions total variation into variation between groups and variation within groups. The central test asks whether the observed group differences are greater than would reasonably be expected from within-group variation alone.

For a one-way ANOVA, the hypotheses are:

• H0: all group means are equal.
• H1: at least one group mean differs.

The F statistic is the ratio of between-group mean square to within-group mean square. A small p-value provides evidence against equal means, but it does not identify which groups differ or establish practical importance. Post-hoc comparisons, such as Tukey-type procedures, can identify pairwise differences while controlling the overall error rate. Effect sizes and confidence intervals should accompany p-values.

ANOVA assumptions include independent observations, approximately normal residuals, and reasonably homogeneous variances. ANOVA can be fairly robust to moderate departures, but severe imbalance, dependence, heteroscedasticity, or non-normality may require transformation, a generalized model, or a nonparametric alternative. Repeated observations from the same machine, patient, customer, or operator are not automatically independent.

A factorial ANOVA can examine several factors simultaneously and estimate interactions. An interaction exists when the effect of one factor changes according to the level of another. For example, a new method may reduce cycle time only on a particular machine. Main effects alone would not adequately describe this process.

Regression analysis

Regression models the relationship between a response and one or more explanatory variables. In simple linear regression,

Y = β0 + β1X + ε

where Y is the response, X is a predictor, β0 is the intercept, β1 is the expected change in Y for a one-unit change in X, and ε represents unexplained variation. Multiple regression extends the model to several predictors and can help adjust for confounding.

Interpretation should focus on the estimated coefficient, its confidence interval, practical meaning, and the conditions under which the relationship applies. R-squared describes the proportion of sample variation explained by the model, but a high R-squared does not prove causation or guarantee useful prediction. Residual plots should be examined for nonlinearity, unequal variance, dependence, and influential observations. Multicollinearity can make individual coefficients unstable when predictors contain overlapping information.

Regression is especially useful for questions such as whether temperature predicts strength, whether staffing predicts waiting time, or whether travel distance predicts damage claims. When the response is binary, logistic regression may be appropriate; when it is a count, Poisson or negative-binomial models may be more suitable. The model must match the data-generating process.

Prediction and explanation are related but different. A variable may improve prediction without being a controllable root cause. Conversely, a controllable cause may be important even if its statistical coefficient is modest. Teams should combine model output with process experiments or intervention evidence whenever possible.

Contingency analysis for categorical data

When both the response and suspected factor are categorical, a contingency table summarizes counts by category. For example, defect status can be cross-tabulated by supplier, or approval outcome by application channel. The chi-square test evaluates whether the distribution of one categorical variable is independent of the other.

The null hypothesis is that the variables are independent. A significant result indicates association, not necessarily causation. Expected cell counts should be sufficiently large for the usual approximation; sparse tables may require exact methods or category consolidation based on process logic. Standardized residuals can identify cells contributing strongly to the overall association.

Percentages must be calculated in the correct direction. Comparing defect rates within each supplier is different from comparing the composition of all defects by supplier. The denominator must match the managerial question. Measures such as risk ratio, odds ratio, or difference in proportions can make the practical effect more interpretable.

Proportion and nonparametric comparisons

Many Six Sigma responses are proportions, rates, or non-normal measures. A comparison of two defect proportions may use a two-proportion procedure, while a confidence interval estimates the plausible range of the difference. For skewed cycle times, rank-based methods or a transformation may be appropriate. These methods do not eliminate the need to understand sampling, independence, and practical significance.

A Disciplined Validation Workflow

A robust Analyze phase can proceed through the following integrated workflow:

1. Clarify the response. Define the operational measure, unit, time window, target, and data source.

2. Confirm data quality. Recheck definitions, missing values, duplicates, measurement reliability, and sampling bias.

3. Describe the problem. Use Pareto charts, histograms, run charts, and summary statistics.

4. Stratify intelligently. Examine process, product, time, supplier, and customer dimensions suggested by the workflow.

5. Generate hypotheses. Use process mapping, cause-and-effect diagrams, 5 Whys, interviews, and technical knowledge.

6. Prioritize hypotheses. Consider frequency, severity, controllability, mechanism, and available evidence.

7. Select an analytical method. Match the method to response type, factor type, design, and assumptions.

8. Test and estimate. Report effect sizes, confidence intervals, p-values, and diagnostic evidence.

9. Investigate alternatives. Examine confounding, interactions, measurement effects, selection bias, and temporal dependence.

10. Validate operationally. Confirm that the suspected cause occurs when the problem occurs and that controlling it is feasible.

11. Document the conclusion. Record validated causes, rejected hypotheses, uncertainty, and implications for Improve.

The Improve phase should not begin with a long list of untested solutions. It should receive a focused set of verified causal levers and a clear statement of the evidence supporting each one.

Applied Case: Reducing Damage in a Distribution Centre

Background

The following is an illustrative case. A regional distribution centre experienced a rise in damaged electronic products. The monthly damage rate had increased from approximately 1.8% to 3.4%, producing rework, replacement cost, and customer dissatisfaction. The Define phase established the problem as damage identified before shipment or reported within the agreed delivery window. Measurement-system review confirmed that damage classification was reasonably consistent among inspectors.

Situation and initial analysis

A Pareto chart showed that cracked screens represented 62% of recorded damage, followed by dented housings and missing accessories. The team initially suspected careless picking. Stratification showed that the highest raw number of incidents occurred on the day shift, but the day shift also processed 70% of all orders. When rates were calculated per 1,000 units, the night shift had the highest rate.

A time-series plot showed that damage increased after a new carton supplier was introduced. However, the supplier change coincided with a product-model mix change, so the apparent relationship required further analysis. A cause-and-effect diagram identified packaging material, carton size, conveyor transfers, handling method, product design, and inspection as possible causes.

Statistical analysis

The team cross-tabulated damage status by carton supplier, controlling the analysis by product family. Within each product family, the newer supplier’s cartons had higher damage proportions. A chi-square test indicated an association between supplier and damage status, while the effect size showed a practically important increase in risk. A contingency analysis by shift showed that the night-shift rate was higher, but the difference diminished substantially after adjusting for the greater proportion of fragile products processed at night. Product mix was therefore a confounder in the shift comparison.

The team then measured the internal clearance between product and carton wall for a sample of shipments. A scatter plot suggested that damage probability increased as clearance decreased below a critical range. Logistic regression, with product family and supplier included as predictors, found that insufficient clearance and the newer carton material were associated with higher damage probability. Residual and influence checks did not reveal a single observation driving the result.

Process observation showed that the new cartons had greater dimensional variation. When cartons were assembled, some products contacted the side wall during conveyor transfers. A controlled pilot using cartons with tighter dimensional tolerance and a specified protective insert reduced contact events and damage in comparable product families. The pilot was not treated as final proof of permanent performance, but it strengthened the causal argument because the intervention changed the suspected mechanism.

Approach and solution

The validated causes were not simply “careless handling” or “night shift performance.” The evidence pointed to a combination of packaging dimensional variation, insufficient clearance for selected product families, and an unstandardized protective insert. Shift differences were treated as a secondary condition requiring workload and product-mix management rather than employee blame.

The Improve team specified carton tolerances, revised packaging instructions, introduced an incoming carton check, and defined product-specific insert requirements. A follow-up control plan included damage rate by product family and supplier, periodic carton-dimension audits, and escalation rules for abnormal contact observations.

Outcome and learning points

The illustrative case demonstrates several analytical lessons. Raw counts can misdirect attention when exposure differs. Stratification can reveal confounding. A significant association is more persuasive when it agrees with a plausible mechanism and is supported by an intervention. Root-cause validation may identify a system condition rather than an individual failure. Finally, the analysis generated control measures that were specific enough to monitor in the Control phase.

Common Errors, Limitations, and Risks

The most common error is confusing correlation with causation. A second is selecting a statistical test after viewing many comparisons and reporting only the favorable result. Multiple testing increases the chance of false positives; exploratory findings should be labeled as such and confirmed where possible.

Other frequent mistakes include using raw counts instead of rates, ignoring process changes over time, treating dependent observations as independent, deleting outliers without investigation, overfitting regression models, and reporting p-values without effect sizes. Small samples can produce unstable estimates, while large samples can make trivial differences appear statistically significant.

Overreliance on brainstorming is also risky. A well-designed fishbone diagram can still contain unsupported assumptions. Conversely, exclusive reliance on statistical software can produce technically correct but operationally meaningless results. Analysts should understand the process, the measurement system, and the decision consequences.

Root-cause work can also create organizational risks. If data are used to rank employees without accounting for workload, product mix, equipment condition, or system design, the analysis may be unfair and may encourage concealment of problems. Confidential information should be protected, and data definitions should be transparent. Safety-related causes require escalation even when their frequency is low. Regulatory records must not be altered to improve apparent performance.

Digital Analytics and Emerging Practice

Modern organizations increasingly use statistical software, process-mining platforms, automated dashboards, and sensor data to support Analyze. Process mining can identify rework loops and waiting pathways in transactional processes. Machine-learning methods can detect nonlinear patterns and interactions in high-dimensional data. Natural-language analytics can classify complaints and identify emerging categories.

These technologies are aids, not substitutes for causal reasoning. Predictive importance does not automatically establish a root cause. Automated models may inherit biased sampling, unstable definitions, or historical decisions that reflect unequal treatment. Analysts should document data lineage, model assumptions, validation results, and access controls. Explainable models are especially valuable when analytical conclusions affect employees, customers, or regulated decisions.

Emerging practice is moving toward integrated causal analytics: combining process knowledge, observational data, designed experiments, digital twins, and real-time anomaly detection. The enduring principle remains unchanged: a cause is valuable when it explains variation and can be acted upon responsibly.

Best Practices for the Analyze Phase

Effective teams generally:

• define the response and denominator before comparing groups;
• verify measurement reliability and data provenance;
• use Pareto analysis for prioritization, not proof;
• stratify by meaningful process conditions;
• write causes in testable language;
• combine graphical exploration with appropriate statistical tests;
• report effect sizes and confidence intervals alongside p-values;
• examine confounding, interaction, dependence, and model diagnostics;
• validate mechanisms through observation, controlled trials, or natural changes;
• distinguish exploratory findings from confirmatory evidence;
• avoid assigning blame when system conditions explain performance;
• document rejected hypotheses as well as validated causes; and
• transfer clear causal knowledge to Improve and Control.

Chapter Summary

The Analyze phase of DMAIC transforms a measured problem into a defensible explanation. Pareto charts prioritize categories, stratification reveals hidden differences, and cause-and-effect diagrams and 5 Whys generate structured hypotheses. Graphical methods expose distributions, trends, clusters, and relationships. ANOVA compares continuous outcomes across groups; regression estimates relationships and adjusts for several predictors; contingency analysis examines associations among categorical variables.

No single tool validates a root cause. Validation requires a clear response measure, reliable data, an appropriate analytical method, attention to assumptions, effect-size interpretation, and examination of alternative explanations. Association must be distinguished from causation, and statistical significance must be balanced with practical importance. The most credible root causes are supported by data, process mechanism, and operational or experimental evidence. Ethical analysis avoids blame, protects confidential information, and recognizes that system design often contributes to human error.

Key Terms

Root cause: An underlying condition whose control or removal prevents recurrence of a problem.

Symptom: An observable manifestation of a process problem.

Pareto analysis: A method that ranks categories by contribution to an outcome to support prioritization.

Stratification: Separation of data into meaningful subgroups to reveal hidden variation.

Cause-and-effect diagram: A structured visual display of possible causes contributing to an effect.

5 Whys: An iterative questioning technique used to move from an event toward underlying conditions.

Response variable: The outcome or performance measure being analyzed.

Predictor: A factor used to explain or estimate variation in a response.

Common cause: A routine source of variation inherent in a process.

Special cause: An unusual, identifiable source of process variation.

ANOVA: A statistical method for testing whether means differ across groups.

Regression: A modeling method that estimates the relationship between a response and one or more predictors.

Interaction: A condition in which the effect of one factor depends on the level of another factor.

Confounding: Distortion of an apparent relationship because another factor is related to both variables.

Contingency table: A table showing counts for combinations of categorical variables.

Chi-square test: A test commonly used to assess independence or association between categorical variables.

Effect size: A quantitative description of the magnitude of a difference or relationship.

Confidence interval: A range of plausible values for a population parameter based on sample evidence.

Statistical significance: Evidence that an observed result is unlikely under a specified null hypothesis.

Operational significance: The practical importance of a result for performance, cost, risk, or customers.

Review Questions

Short Answer Questions

1. Why does a Pareto chart prioritize investigation but not validate a root cause?

2. Distinguish a symptom, proximate cause, and root cause using a process example.

3. What is the purpose of stratification in the Analyze phase?

4. State the null hypothesis tested in a one-way ANOVA.

5. Why must effect size be considered alongside statistical significance?

Long Answer Questions

1. Explain how cause-and-effect diagrams and the 5 Whys can be combined to generate testable root-cause hypotheses.

2. Describe the assumptions, interpretation, and limitations of regression analysis in Six Sigma.

3. Compare ANOVA and contingency analysis in terms of data type, purpose, and interpretation.

4. Present a complete workflow for validating root causes from problem definition through transfer to Improve.

5. Discuss how confounding and interaction can distort conclusions about process causes.

Analytical / Application Questions

1. A department records 240 late orders on the day shift and 160 on the night shift. The day shift processes 12,000 orders and the night shift processes 4,000. Which shift has the higher late-order rate, and what additional analysis would you perform before identifying the shift as a root cause?

2. A regression model shows that operator experience predicts lower defects, but experienced operators are assigned mainly to newer machines. Explain how you would investigate confounding and avoid an unsupported conclusion.

3. A chi-square test finds a significant association between supplier and defect status, but the practical difference in defect rates is very small. How should a Six Sigma team interpret this result and decide whether action is warranted?

22. Multiple Choice Questions with Answers

Question 1

What is the primary purpose of Pareto analysis in the Analyze phase?

A. To prove that one factor causes a defect

B. To prioritize categories for further investigation

C. To replace measurement-system analysis

D. To determine whether data are normally distributed

Correct Answer: B. To prioritize categories for further investigation

Question 2

A team compares defect counts across two machines, but one machine produces twice as many units. Which measure is generally more appropriate for comparison?

A. Raw defect count only

B. Defect rate using a relevant denominator

C. The largest individual defect size

D. The cumulative percentage of all defects

Correct Answer: B. Defect rate using a relevant denominator

Question 3

In a one-way ANOVA, a statistically significant result indicates that:

A. Every group mean differs from every other group mean

B. At least one group mean differs from another

C. The factor definitely causes the response

D. Within-group variation is zero

Correct Answer: B. At least one group mean differs from another

Question 4

Which situation best illustrates confounding?

A. A response has two identical measurements

B. A third factor influences both the suspected cause and the response

C. Two factors have an interaction effect

D. A sample contains an outlier

Correct Answer: B. A third factor influences both the suspected cause and the response

Question 5

Which analysis is most directly suited to examining association between supplier category and defect status?

A. Chi-square contingency analysis

B. One-sample t-test

C. Simple linear regression only

D. Histogram analysis alone

Correct Answer: A. Chi-square contingency analysis

Question 6

What does an interaction mean in a factorial analysis?

A. The response has no variation

B. The effect of one factor depends on the level of another factor

C. All factor means are equal

D. The sample was collected randomly

Correct Answer: B. The effect of one factor depends on the level of another factor

Question 7

A team removes an extreme observation solely because it weakens the suspected relationship. What is the principal analytical concern?

A. The graph may have too many labels

B. The team may be introducing biased conclusions

C. The response has become categorical

D. The sample size has necessarily doubled

Correct Answer: B. The team may be introducing biased conclusions

Question 8

Which statement best distinguishes statistical significance from operational significance?

A. They always mean exactly the same thing

B. Statistical significance concerns evidence against a null hypothesis, while operational significance concerns practical impact

C. Operational significance can be assessed only with a p-value

D. Statistical significance applies only to qualitative data

Correct Answer: B. Statistical significance concerns evidence against a null hypothesis, while operational significance concerns practical impact

Question 9

In the illustrative distribution-centre case, why was the day-versus-night comparison potentially misleading?

A. No defects occurred during the day

B. The shifts processed different volumes and product mixes

C. The team used a cause-and-effect diagram

D. The cartons had no dimensional variation

Correct Answer: B. The shifts processed different volumes and product mixes

Question 10

Which evidence most strengthens the claim that a suspected factor is causal?

A. The factor sounds plausible to the team

B. The factor appears in a fishbone diagram

C. The factor is associated with the response, has a credible mechanism, and changes in a controlled or comparable intervention

D. The factor is mentioned by one employee

Correct Answer: C. The factor is associated with the response, has a credible mechanism, and changes in a controlled or comparable intervention