01 · Define what a verified cause must explain
Define what a verified cause must explain
Deriving main causes is the synthesis step in Analyze. It combines process observation and statistical evidence into a defensible explanation of the customer problem. A cause should explain a specific failure, occur before it, have a plausible mechanism, and support an intervention that changes the outcome. A high correlation, a large Pareto category, or a significant coefficient alone does not meet that standard.
Define the outcome and scope precisely. For a warehouse, wrong item shipments from one zone are different from all shipping defects across the site. State the unit, defect denominator, detection method, and operating conditions. List candidate mechanisms rather than vague labels such as human error. Similar labels beside neighboring stock locations is a testable mechanism; careless staff is an uninformative attribution.
02 · Build an evidence chain
Build an evidence chain
Process maps and direct observation identify where the failure can occur. Cause and effect diagrams, interviews, and repeated why questions generate hypotheses, not proof. Stratified data test predictions of those mechanisms. Regression can assess whether associations persist after suitable adjustment. Experiments test whether deliberately changing the proposed cause changes the response under controlled conditions.
Write a prediction for each candidate before analyzing it. If adjacent lookalike labels drive wrong picks, errors should concentrate at those pairs after accounting for exposure, and improving differentiation should reduce that specific error type. If congestion drives mistakes, expect a relation with crowding that persists within comparable order classes. A useful hypothesis risks being contradicted; one that explains every possible result cannot guide verification.
03 · Worked warehouse cause verification
Worked warehouse cause verification
Suppose fictional baseline records show twenty wrong item shipments among two hundred orders from a zone with similar neighboring labels, a 10% rate. Other zones have lower rates, but the affected zone also handles more complex orders. The cross zone association supports investigation without establishing that labels caused the difference. Audit error reports and observe picking to verify whether the label confusion mechanism actually occurs.
The team then randomizes two hundred eligible orders to the existing label display and two hundred to a clearly differentiated display, using independent assignments without spillover. Existing labels produce twenty wrong item shipments; differentiated labels produce six. Estimated rates are 0.10 and 0.03, giving an absolute reduction of 0.07, or seven percentage points, and a relative reduction estimate of 70%.
For an illustrative independent large sample difference interval, standard error is √[0.10×0.90/200 + 0.03×0.97/200] = 0.02440. The approximate 95% interval for the existing minus differentiated rate is 0.07 ± 1.96×0.02440, or 0.02217 to 0.11783. This supports a positive reduction under the design assumptions but leaves considerable uncertainty in the amount. With only six events in one arm, more accurate binomial difference methods can be preferable to this simple teaching approximation.
04 · Verify the mechanism and rule out alternatives
Verify the mechanism and rule out alternatives
The intervention's evidence is stronger if observed wrong picks decrease specifically at the ambiguous locations and unrelated error categories remain consistent. Check that detection, order mix, training, and staffing were comparable across assigned arms. Record whether operators actually saw the assigned display. If a label change is physically shared across an entire zone, individual order randomization may be impossible; randomize independent locations, sessions, or zones and analyze the cluster assignment.
Inspect temporal order and competing explanations. Low scanner battery may accompany long shifts and congestion, so its association can disappear when the real conditions are accounted for. Conversely, inappropriate adjustment for a mediator can conceal a real total effect. Use a process based causal diagram or written pathway to decide which variables belong in a model. Statistical adjustment addresses measured conditions with adequate overlap, not every imaginable source of bias.
05 · Judge main causes by practical contribution
Judge main causes by practical contribution
A verified cause can be real but account for only a small portion of the problem. Estimate exposure, effect, implementation feasibility, and interactions. If only one tenth of all orders pass the affected locations, applying the estimated seven percentage point reduction to every site order would overstate benefits. Under that fixed illustrative exposure, expected overall reduction is 0.10×0.07 = 0.007, or 0.7 percentage points, before considering transportability and uncertainty.
Avoid adding overlapping cause contributions as though they were independent. Congestion and layout may act along the same pathway; fixing one can reduce the apparent contribution of the other. A factorial experiment can investigate interactions among controllable causes. Rank actionable mechanisms by the expected customer benefit and evidence strength, while stating what remains unverified. The largest observed association is not necessarily the most effective or affordable intervention.
06 · Confirm beyond the discovery sample
Confirm beyond the discovery sample
Use new data for confirmation because candidates were selected from the original evidence. Repeat the intervention across relevant shifts, product mixes, and operators. Prespecify the primary outcome, practical threshold, eligibility rules, and analysis. Verify that the effect persists without creating damage, delay, or workarounds. A beneficial mean effect with harmful subgroup behavior needs a conditional operating rule.
When randomization is impractical, an interrupted time series, controlled before and after study, or other quasi experimental design can strengthen the case. Each requires assumptions, such as no concurrent differential change and suitable trend behavior. A simple before and after difference is vulnerable to seasonality, regression to the mean, and selection of an unusually bad baseline. Document alternative explanations and assess how much the conclusion depends on them.
07 · Separate strong and weak verification
Separate strong and weak verification
Strong verification combines a credible mechanism, reliable measurement, exposure adjusted association, controlled intervention, and independent confirmation. It reports an effect with uncertainty and the conditions under which the cause operates. Weak verification treats brainstorming votes as evidence, names a significant regression term a root cause, or declares a cause removed because a test becomes nonsignificant. Failure to reject a remaining difference does not establish elimination.
A failed intervention is useful evidence. Determine whether the manipulated input actually changed, whether the measurement could detect an effect, and whether the mechanism was wrong. Do not protect a preferred explanation by silently changing the outcome or excluding inconvenient runs. Preserve negative results and distinguish absence of evidence from evidence against a mechanism.
08 · Hand off a cause statement through DMAIC
Hand off a cause statement through DMAIC
The Analyze output should state the mechanism, affected population, evidence, estimated contribution, uncertainty, and recommended intervention. For the label example: ambiguous adjacent labels plausibly produce wrong picks, and a randomized display change reduced that event rate in eligible orders under tested conditions. This is more useful than a generic claim that labeling is the root cause.
Improve implements and confirms the intervention at realistic scale. Control monitors both the mechanism input and the customer outcome, with a response plan if ambiguity or errors recur. Archive the causal reasoning, assignment design, measurement definitions, comparison data, safeguards, and unresolved alternatives. A verified cause is a conditional evidence based explanation that supports action, not an absolute claim that no other cause exists.
09 · State what would change the conclusion
State what would change the conclusion
A useful cause statement includes a falsifiable boundary. If the label intervention fails on new comparable orders despite confirmed exposure and unchanged detection, revisit the mechanism or its operating conditions. If errors return only after stock adjacency changes, that supports a conditional layout and labeling mechanism. Control evidence should refine the causal explanation rather than treating the Analyze conclusion as permanently settled.
Connections
Related methods and phase use
Placed under Analyze because of the outcome it supports. Use the method in other phases when its purpose and assumptions fit the work; the phase list is a guide to selection, not a compulsory sequence.
Sources
Professional foundations and further reading
The explanation and fictional teaching examples are written for LeanTrainingLab. Linked primary sources provide professional foundations; example values and local conventions are stated in the guide. The accepted toolset draws on Six Sigma+Lean Toolset, second edition (Lunau, editor; Springer, 2013). No book text or examples are reproduced here.