DMAIC: Analyse and Root-Cause Validation

Introduction

The Analyse phase of DMAIC converts collected data into defensible knowledge about process performance. Define identifies the problem, Measure establishes a reliable baseline, and Analyse determines why the gap exists. The central question is not simply, “What appears to be wrong?” but, “Which causes demonstrably produce the observed variation, and how certain are we?”

Root-cause analysis is therefore both a practical and statistical activity. Qualitative tools such as Pareto analysis, the 5 Whys and cause-and-effect diagrams help teams generate and organise possible causes. Quantitative methods—including stratification, probability distributions, confidence intervals, hypothesis tests, correlation, regression and ANOVA—help determine whether proposed causes are supported by evidence.

A cause should not be accepted because it is familiar, senior managers believe it, or it appears on a brainstorming list. A credible root cause should have a logical mechanism, a measurable relationship with the problem, evidence that is reproducible, and a practical connection to process control. The Analyse phase ends when the important causes have been validated sufficiently to guide the Improve phase.

Learning objectives

By the end of this chapter, the learner should be able to:

1. Explain the purpose and outputs of the Analyse phase.
2. Use Pareto analysis, 5 Whys and cause-and-effect diagrams appropriately.
3. Stratify process data to reveal meaningful patterns.
4. Select and interpret suitable probability distributions.
5. Construct and interpret confidence intervals.
6. Formulate and evaluate hypothesis tests.
7. Distinguish correlation from causation.
8. Use regression and ANOVA to investigate relationships and differences.
9. Validate root causes using converging qualitative and quantitative evidence.
10. Recognise statistical, operational, ethical and governance limitations.

1. The logic of root-cause validation

A process output, Y, is influenced by inputs and conditions commonly represented as X₁, X₂, ..., X_k. A useful conceptual model is:

Here, f represents the process mechanism and ε represents unexplained variation, measurement error and other unmeasured influences. The notation does not imply that every process can be expressed exactly by a known equation. It provides a disciplined way to ask which inputs affect the output and how large those effects are.

A suspected cause is stronger when:

• the cause occurs before the effect;
• there is a credible physical, technical or organisational mechanism;
• the relationship is observed repeatedly;
• the relationship remains after relevant factors are considered;
• changing or controlling the cause changes the outcome; and
• the measurement system is capable of detecting the difference.

The last point is essential. If measurement error is large, apparent differences may be artefacts. The Measure phase should therefore provide evidence of operational definitions, measurement-system analysis, sampling logic and data integrity.

Common outputs of Analyse

The Analyse phase normally produces:

• a refined process understanding;
• a list of suspected causes;
• stratified evidence showing where and when the problem occurs;
• statistical estimates and tests;
• validated root causes and rejected hypotheses;
• quantified cause effects, where possible; and
• a prioritised input list for Improve.

A “root cause” is not necessarily the deepest philosophical cause. It is a controllable or manageable process condition that, when addressed, materially reduces the problem and prevents recurrence.

2. Qualitative discovery tools

2.1 Pareto analysis

Pareto analysis ranks categories by their contribution to a defined problem. It is based on the practical observation that causes are often unevenly distributed, not on a universal law that exactly 80% of effects always arise from 20% of causes.

For category i, its percentage contribution is:

where c_i is the count, cost, time or other measure for category i, and ∑c_i is the total across all categories. Categories are sorted from largest to smallest, and a cumulative percentage is calculated.

The measure must match the decision. A defect-count Pareto may prioritise frequent minor defects, whereas a cost Pareto may identify rare but expensive failures. A time-loss Pareto may differ again. Teams should state the unit, period, population and inclusion rules.

Example. A service centre records 200 complaint cases: waiting time 90, incorrect information 50, billing 35, system access 15 and other 10. Waiting time contributes 45%, incorrect information 25%, and the first two categories together contribute 70%. This suggests an initial focus, but it does not prove that waiting time is the underlying cause. It may itself result from staffing, scheduling or system constraints.

Pareto analysis is descriptive and prioritising. It does not establish statistical significance or causality.

2.2 The 5 Whys

The 5 Whys is an iterative questioning method. Starting with a clearly defined problem, the team asks why it occurred and then asks why the answer occurred. Five is a guide, not a required number. The questioning stops when the team reaches a process condition that is specific, evidence-based and actionable.

A weak chain is:

• Problem: shipment was late.
• Why? The operator was careless.
• Why? The operator lacked commitment.

This chain blames individuals and does not identify a controllable process mechanism. A stronger chain might be:

• Shipment was late.
• Why? The order was released after the carrier cut-off.
• Why? Approval remained in a shared mailbox.
• Why? No owner or escalation rule existed.
• Why? The workflow design did not define approval service levels.
• Why? Process ownership had not been assigned during system implementation.

The final statement should be checked against records, observations and interviews. The 5 Whys can oversimplify problems with multiple interacting causes; in such cases, a cause-and-effect diagram and data analysis are preferable.

2.3 Cause-and-effect diagrams

A cause-and-effect, or fishbone, diagram organises possible causes around a defined effect. Common categories in manufacturing include man, machine, method, material, measurement and environment. In services, categories may include people, policy, process, technology, information, customer and environment.

The diagram is a hypothesis-generation device, not proof. Each branch should be translated into a measurable variable. For example, “machine problem” may become temperature, vibration, calibration status, software version or downtime history. Vague labels cannot be validated.

Teams should avoid treating all branches as equally likely. After construction, causes should be screened using process knowledge, data availability, risk and potential impact.

3. Stratification: making hidden patterns visible

Stratification separates data into meaningful subgroups before analysis. Useful strata include product family, supplier, shift, operator, machine, location, customer segment, defect type, day of week, temperature range and process revision.

Suppose the overall defect rate is 4%. This average may conceal rates of 1% on day shift and 9% on night shift, or 2% for one supplier and 12% for another. The average alone cannot reveal the operational decision.

For a stratum j, the observed defect proportion is:

where d_j is the number of defects and n_j is the number of opportunities in stratum j. Rates must be compared using compatible denominators. Comparing 20 defects from 100 opportunities with 20 defects from 10,000 opportunities is misleading.

Stratification can also expose confounding. If one machine is used mainly by experienced operators and another mainly by trainees, a difference initially attributed to operators may actually reflect machine condition or assignment policy. Cross-tabulation and, where appropriate, regression or ANOVA can help separate these effects.

Good stratification requires adequate sample sizes and meaningful categories. Excessive subdivision creates small groups, unstable rates and false patterns. Categories should be defined before inspecting results when possible, or the exploratory nature of post hoc findings should be reported.

4. Probability distributions in process analysis

A probability distribution describes the possible values of a variable and their relative likelihood. The correct distribution depends on the data-generating process, not on convenience.

For a Poisson variable X, the probability of observing x events is:

Here, λ is the average event rate for the defined exposure, and e is the base of natural logarithms. The model assumes events occur independently at a stable rate over the relevant interval. Overdispersion, clustering or changing exposure can make the model inappropriate.

For a binomial variable:

where n is the number of independent trials, p is the probability of a defect on each trial, and C(n,x) is the number of combinations of n trials containing x successes. The trials must have a consistent definition and approximately common probability.

The normal distribution is described by mean μ and standard deviation σ. Standardisation is:

This transformation is useful when the normal model is reasonable. Normality should not be assumed merely because the sample is large; large samples improve estimation but do not make a fundamentally unsuitable model appropriate for every purpose.

Distribution analysis helps determine suitable charts, intervals and tests. It also supports risk estimates such as the proportion beyond a specification limit. For skewed data, transformation, robust methods or a distribution designed for positive values may be preferable.

5. Estimation and confidence intervals

A sample statistic estimates a population parameter. Because samples vary, a point estimate should usually be accompanied by an interval estimate.

For a sample mean x̄ with sample standard deviation s and sample size n, a commonly used confidence interval for the population mean is:

Here, α is the significance level, t_(1−α/2, n−1) is the relevant t-distribution critical value, and n−1 is the degrees of freedom. This interval assumes independent observations and is exact under normality; for sufficiently large samples, the central-limit approximation may be useful under less restrictive conditions.

For a proportion, the simple Wald interval is often unreliable with small samples or proportions near 0 or 1. Wilson or exact binomial intervals are generally safer. The chosen method should be stated.

A 95% confidence interval does not mean there is a 95% probability that a fixed population parameter lies inside this particular calculated interval. In repeated sampling, a procedure that produces 95% intervals will contain the true parameter in approximately 95% of comparable samples. In practice, the interval expresses estimation uncertainty under the model and sampling design.

Worked example. A process records 36 cycle times with mean 12.4 minutes and standard deviation 1.8 minutes. Using an approximate 95% t critical value of 2.03 for 35 degrees of freedom:

12.4 ± 2.03 × (1.8/√36) = 12.4 ± 0.609 minutes.

The interval is approximately 11.79 to 13.01 minutes. It describes uncertainty in the population mean, not the range in which individual cycle times will fall.

Precision improves as n increases, but the standard error decreases in proportion to 1/√n. More observations cannot correct biased sampling or poor measurement.

6. Hypothesis testing

Hypothesis testing evaluates whether data are sufficiently inconsistent with a specified null hypothesis. A typical procedure is:

1. Define the practical question and population.
2. State the null hypothesis H₀ and alternative hypothesis H₁ before examining results.
3. Select a test, significance level and assumptions.
4. Calculate the test statistic and p-value or confidence interval.
5. Assess practical significance, not only statistical significance.
6. Make a process decision and document limitations.

The p-value is the probability, assuming H₀ is true and the model assumptions hold, of obtaining a result at least as extreme as the observed result. It is not the probability that H₀ is true, nor is it the probability that the result occurred by chance in an informal sense.

A two-sample comparison of independent means may use:

This is the Welch t statistic, which does not require equal population variances. The degrees of freedom are estimated using the Welch–Satterthwaite method. Independence and an appropriate measurement scale remain important.

For paired observations, such as the same machines measured before and after a change, analyse the differences d_i rather than treating the observations as independent:

where d̄ is the mean difference and s_d is the standard deviation of the differences.

A statistically significant result may be too small to matter operationally. Conversely, a practically important effect may fail to reach significance because the sample is small or variation is high. Report effect size, confidence interval and process consequences together.

Multiple testing increases the chance of false positive findings. If many causes, strata or outcomes are tested, teams should predefine primary analyses, control the family-wise error rate where appropriate, or use methods such as false-discovery-rate control. Exploratory findings should be labelled as such.

7. Correlation and regression

Correlation measures the strength and direction of linear association between two quantitative variables. Pearson’s sample correlation is:

Its value lies between −1 and +1. A value near +1 indicates a strong positive linear association, a value near −1 a strong negative linear association, and a value near 0 little linear association. A nonlinear relationship can have a correlation near zero.

Correlation does not establish causation. A third variable may influence both measured variables, the direction may be reversed, or the relationship may arise from time trends or selection effects. Scatterplots should precede numerical correlation, and outliers should be investigated rather than automatically removed.

Simple linear regression models a response as:

where y_i is the response, x_i is the predictor, β₀ is the intercept, β₁ is the expected change in the response for a one-unit increase in the predictor, and ε_i is the error term. Ordinary least squares estimates coefficients by minimising the sum of squared residuals:

Regression assumptions commonly include appropriate linear form, independent errors, constant error variance and, for small-sample inference, approximately normal errors. Residual plots are essential. A high R² does not prove causality and does not guarantee useful predictions outside the observed data range.

Multiple regression can adjust for several predictors:

Coefficients are interpreted conditionally: β₁ estimates the expected change in y for a one-unit change in x₁, holding the included predictors constant. Multicollinearity can make coefficients unstable and difficult to interpret. Regression is most persuasive when supported by process logic, temporal order and intervention evidence.

8. ANOVA for comparing several conditions

Analysis of variance (ANOVA) tests whether the means of three or more groups differ. It partitions total variation into variation between groups and variation within groups:

where SST is total sum of squares, SSB is between-group sum of squares and SSW is within-group sum of squares.

For k groups and total sample size N:

The F statistic compares between-group mean square with within-group mean square. Under the null hypothesis that all group means are equal, and assuming independent observations, approximately normal residuals and equal variances, the statistic follows an F distribution with k−1 and N−k degrees of freedom.

A significant omnibus ANOVA indicates that at least one mean differs; it does not identify which groups differ. Planned contrasts or multiplicity-adjusted post hoc comparisons, such as Tukey’s method, are required. If variances are unequal, Welch ANOVA or suitable robust methods may be preferable.

Two-way ANOVA can assess two factors and their interaction. An interaction means the effect of one factor changes depending on the level of another. This is often operationally important: a new material may perform well on one machine but poorly on another. Ignoring interaction can produce misleading main-effect conclusions.

9. A disciplined validation workflow

A practical Analyse workflow is:

1. Restate the problem operationally. Define the response, unit, defect opportunity, time window and target.
2. Verify the data. Check missing values, duplicates, coding, measurement capability and sampling bias.
3. Explore visually. Use run charts, histograms, box plots, scatterplots and Pareto charts.
4. Stratify. Compare meaningful process segments and inspect exposure denominators.
5. Generate hypotheses. Use process maps, 5 Whys, fishbone analysis and subject expertise.
6. Select an analytical method. Match the method to data type, design, distribution and question.
7. Quantify uncertainty. Report confidence intervals, effect sizes and assumptions.
8. Check alternative explanations. Examine confounding, measurement error, time effects and selection.
9. Validate critical causes. Use controlled trials, natural experiments, confirmation samples or prospective monitoring where feasible.
10. Prioritise for Improve. Select causes that are influential, controllable, economically important and ethically acceptable.

A cause-and-effect matrix can help rank inputs by their relationship to critical outputs, but rankings are provisional until supported by data. Failure mode and effects analysis may then assess severity, occurrence and detectability; however, risk priority scores should not replace evidence or professional judgement.

10. Illustrative case: late laboratory reports

Background

A regional diagnostic laboratory has a target that 95% of reports should be released within 24 hours. Over three months, performance falls to 86%. The issue affects patient communication and creates overtime costs.

Problem definition

The team defines a late report as a report released more than 24 hours after specimen receipt. The clock starts at electronic receipt, and the denominator is all eligible specimens. This definition excludes emergency tests with a separate service-level agreement.

Initial analysis

A Pareto chart of 1,200 late reports shows: analyser waiting, 420; incomplete clinical information, 300; verification backlog, 240; specimen recollection, 150; other, 90. The largest category is analyser waiting, but the team does not assume it is the root cause.

Stratification shows that analyser waiting is 8% of specimens on the day shift and 22% on the evening shift. However, evening specimens are concentrated on one analyser. When the data are stratified by analyser, analyser B has a late rate of 19%, compared with 7% for analyser A. This suggests a machine, scheduling or workload issue rather than a generic shift problem.

Hypotheses and tests

The fishbone diagram identifies calibration status, batch size, reagent lot, staffing, maintenance and routing rules. The 5 Whys indicates that evening specimens queue because the routing rule sends them to analyser B until a batch reaches a minimum size. The rule was introduced to reduce reagent waste.

The team compares cycle time for batches run under the old and revised routing rules. Because observations are collected in matched time blocks, a paired comparison is considered. The revised rule reduces average waiting time by 18 minutes, with a 95% confidence interval of 11 to 25 minutes. A practical review confirms that the reduction is large enough to improve the service-level target.

Regression is then used to examine waiting time against batch size, workload and analyser. The model suggests that analyser identity and batch size remain important after workload is considered, but residual plots show higher variance during peak periods. The team therefore avoids claiming a complete causal model.

Approach and solution

A controlled pilot changes the routing rule for two weeks while maintaining staffing and reagent monitoring. A comparable two-week baseline is retained. Late reports fall from 14.2% to 7.1%; reagent cost rises modestly but remains below the cost of overtime and service failure. A follow-up ANOVA across day, evening and night periods finds no material difference after the routing change, although peak workload remains a secondary risk.

Outcome

The revised routing rule is standardised, a peak-load escalation trigger is introduced, and analyser performance is monitored weekly. The team also documents that the pilot was not a fully randomised experiment and that seasonal demand may limit generalisation.

Learning points

• Pareto analysis prioritised the visible category but did not prove the cause.
• Stratification revealed that the overall shift effect concealed an analyser-specific pattern.
• Qualitative tools explained the mechanism; statistical analysis quantified it.
• A pilot intervention provided stronger evidence than observational correlation alone.
• Cost, service performance and resource use were evaluated together.

11. Limitations, ethics and governance

Root-cause analysis can cause harm when conclusions unfairly blame individuals, conceal system failures or are used for disciplinary action without due process. Data should be collected for improvement with transparent purpose, appropriate access control and respect for privacy. Patient, employee and customer identifiers should be minimised or anonymised where possible.

Analysts must not manipulate inclusion rules, remove inconvenient observations without justification, repeatedly test until a desired p-value appears, or present exploratory findings as confirmed causes. Statistical significance should never be used to conceal a clinically, financially or operationally trivial effect.

AI and automated analytics can assist with anomaly detection, text classification, clustering and visualisation. They do not remove the need for valid definitions, representative data, human review, explainability and governance. Models may reproduce historical bias or identify correlations that have no causal meaning. Any AI-supported root-cause claim should be traceable to data, assumptions and review decisions.

Important limitations include non-independence, confounding, inadequate sample size, unstable processes, measurement error, missing data and changes during data collection. If the process is changing over time, a single pooled analysis may conceal meaningful shifts. Control charts and time-ordered analysis should complement group comparisons.

12. Best practices for professional application

• Begin with a precise operational definition.
• Separate discovery from confirmation.
• Use multiple forms of evidence rather than one attractive chart.
• Preserve time order and process context.
• Report denominators, units, sample sizes and missing-data treatment.
• Display raw data where possible, not only summary statistics.
• Match the method to the measurement scale and study design.
• Check assumptions with plots and residual analysis.
• Report effect size and confidence interval with p-values.
• Distinguish statistical association, practical importance and causation.
• Validate the measurement system before interpreting small differences.
• Confirm important causes through intervention or prospective monitoring.
• Document decisions so that the analysis is reproducible.
• Involve process owners and affected personnel in interpreting results.

Chapter Summary

The Analyse phase of DMAIC identifies and validates the causes of process problems. Pareto charts prioritise categories, while the 5 Whys and cause-and-effect diagrams organise causal hypotheses. Stratification exposes patterns hidden by averages and requires compatible denominators. Probability distributions support appropriate modelling, estimation and test selection.

Confidence intervals express uncertainty around estimates. Hypothesis tests evaluate evidence against a null hypothesis, but p-values do not measure practical importance or prove causation. Correlation describes association; regression quantifies relationships while requiring careful assessment of assumptions and confounding. ANOVA compares several means and can reveal factor interactions, but significant results require suitable follow-up comparisons.

Reliable root-cause validation combines process knowledge, sound measurement, appropriate statistical methods, transparent assumptions and, where feasible, confirmation through intervention. Ethical data use, protection against blame, avoidance of p-hacking and responsible use of AI are essential to credible improvement work.

Key Terms

Root cause: A process condition that materially contributes to a problem and can be controlled or managed.

Pareto analysis: Ranking categories by their contribution to a defined total.

5 Whys: Iterative questioning used to move from a problem symptom toward an actionable process cause.

Cause-and-effect diagram: A structured visual method for organising possible causes of an effect.

Stratification: Separating data into meaningful subgroups for comparison.

Population: The complete set of units or observations about which conclusions are intended.

Sample: A subset of the population used for analysis.

Probability distribution: A model describing the possible values and probabilities of a variable.

Confidence interval: An interval produced by a method designed to contain a population parameter at a stated long-run frequency.

Null hypothesis: A formal statement representing no specified effect or difference.

Alternative hypothesis: A statement representing the effect or difference being investigated.

p-value: The probability, under the null model and assumptions, of results at least as extreme as those observed.

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

Correlation: A measure of the direction and strength of linear association.

Regression: A modelling method relating a response to one or more predictors.

Residual: The difference between an observed value and its fitted value.

ANOVA: Analysis of variance used to compare means and partition variation.

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

Confounding: Distortion of an apparent relationship by another variable related to both the predictor and outcome.

Overdispersion: Variation greater than expected under a chosen probability model.

Review Questions

Short Answer Questions

1. What is the principal purpose of the Analyse phase in DMAIC?
2. Why does a Pareto chart not, by itself, prove causation?
3. What is the purpose of stratifying process data?
4. Interpret the meaning of a p-value in a hypothesis test.
5. Why must a significant ANOVA result be followed by additional comparisons?

Long Answer Questions

1. Explain how Pareto analysis, the 5 Whys and a cause-and-effect diagram can be used together in a root-cause investigation.
2. Describe the assumptions and practical uses of the binomial, Poisson and normal distributions.
3. Explain the difference between a confidence interval and a p-value, including the limitations of each.
4. Discuss why correlation does not establish causation and explain how regression can improve, but not eliminate, this problem.
5. Describe a complete workflow for validating root causes in a service or manufacturing process.

Analytical / Application Questions

1. A process has an overall defect rate of 5%. When stratified, day shift records 2 defects in 100 opportunities and night shift records 18 defects in 200 opportunities. Calculate each shift’s defect rate and explain why the overall average may mislead decision-makers.
2. A team finds a strong relationship between workload and waiting time. Propose at least three alternative explanations or confounders and design a validation study that would distinguish among them.
3. A three-machine ANOVA produces a statistically significant result, but the largest mean difference is operationally too small to matter. Explain how the team should interpret the result and decide whether improvement action is justified.

Multiple Choice Questions with Answers

Question 1

Which statement best describes the role of Pareto analysis?

A. It proves that the largest category is the root cause.

B. It ranks categories to support prioritisation of investigation.

C. It replaces measurement-system analysis.

D. It determines whether data are normally distributed.

Correct Answer: B. It ranks categories to support prioritisation of investigation.

Question 2

For a Poisson model, λ most directly represents:

A. The population standard deviation

B. The probability of success on one trial

C. The average event rate for the defined exposure

D. The number of comparison groups

Correct Answer: C. The average event rate for the defined exposure

Question 3

A 95% confidence interval is best interpreted as:

A. A 95% probability that the fixed parameter lies in this calculated interval

B. The range containing 95% of individual observations

C. An interval generated by a method that captures the true parameter in about 95% of repeated comparable samples

D. Proof that the null hypothesis is false

Correct Answer: C. An interval generated by a method that captures the true parameter in about 95% of repeated comparable samples

Question 4

Which conclusion is justified by a statistically significant one-way ANOVA?

A. Every group mean differs from every other group mean.

B. At least one group mean differs, subject to model assumptions.

C. The largest group has the largest mean.

D. The factor necessarily causes the response.

Correct Answer: B. At least one group mean differs, subject to model assumptions.

Question 5

What is a major limitation of correlation?

A. It can only be used with categorical data.

B. It always requires equal group sizes.

C. It establishes causation only when the correlation is positive.

D. It describes association but does not by itself establish causation.

Correct Answer: D. It describes association but does not by itself establish causation.

Question 6

In the regression model y_i = β₀ + β₁x_i + ε_i, β₁ represents:

A. The expected change in y for a one-unit increase in x

B. The unexplained error for observation i

C. The average value of x

D. The total sum of squared residuals

Correct Answer: A. The expected change in y for a one-unit increase in x

Question 7

A process records 36 observations with mean 12.4 minutes and standard deviation 1.8 minutes. Using t = 2.03, what is the approximate 95% confidence interval for the mean?

A. 10.37 to 14.43 minutes

B. 11.79 to 13.01 minutes

C. 12.31 to 12.49 minutes

D. 8.80 to 16.00 minutes

Correct Answer: B. 11.79 to 13.01 minutes

Question 8

Which action most directly improves the credibility of a suspected root cause?

A. Accepting the explanation offered by the most senior employee

B. Removing observations that do not fit the explanation

C. Testing the cause through a controlled or prospective intervention where feasible

D. Running many tests until one has p < 0.05

Correct Answer: C. Testing the cause through a controlled or prospective intervention where feasible

Question 9

Why is stratification by denominator important when comparing defect rates?

A. It ensures that rates are based on comparable exposure opportunities.

B. It guarantees normally distributed data.

C. It removes all confounding automatically.

D. It makes every subgroup the same size.

Correct Answer: A. It ensures that rates are based on comparable exposure opportunities.

Question 10

In a two-factor analysis, an interaction means that:

A. Both factors have exactly the same mean.

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

C. Neither factor affects the response.

D. The residual variance is zero.

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