01 · Mechanics and boundaries
Mechanics and boundaries
Data stratification separates observations into relevant groups before interpreting a combined result. Typical strata include route, order size, shift, release wave, stock condition, and equipment state. The purpose is to compare like conditions and discover where an observed problem concentrates. Stratification is an analytical grouping method; it is different from stratified sampling, which designs how observations are selected.
Choose factors grounded in the process and available before the outcome where possible. A subgroup defined as “late orders” cannot explain why orders become late. Groups need clear membership rules and denominators. If one order can belong to several categories, distinguish overlapping flags from a mutually exclusive partition.
02 · Construct the analysis
Construct the analysis
Define the outcome and initial grouping hypothesis before inspecting many splits. Preserve the ungrouped total as a reconciliation check. For each stratum count exposure, outcome, and missing data. Compare rates as well as counts, because a large group can produce more late orders while having a lower late proportion. For continuous data compare distributions and timing, not only means.
Cross-classify a second meaningful factor when the first contrast might be confounded. Keep cells large enough to be interpretable and mark sparse cells explicitly. Decide in advance which small categories may be combined based on mechanisms rather than favorable results. Look for whether the overall and within-stratum comparisons tell the same story.
03 · Worked warehouse-dispatch example
Worked warehouse-dispatch example
The fictional North Quay warehouse dispatches a single parcel-order family. Its boundary starts when an order is accepted for release and ends at confirmed carrier handover. Demand is 240 orders in a 420-minute net working day. Each order contains three lines on average, but the example keeps the parcel order as the unit of flow. No number below describes a real organization or a demonstrated improvement. The purpose is to make the reasoning traceable rather than to present a success story.
A representative order spends 12 minutes awaiting release, 2 minutes being released, 18 minutes awaiting picking, 8 minutes being picked, 24 minutes awaiting packing, 5 minutes being packed, 10 minutes awaiting checking, 2 minutes being checked, 15 minutes awaiting loading, and 3 minutes being loaded. These mutually exclusive intervals total 99 minutes: 20 minutes of processing and 79 minutes of waiting. They describe one complete normal-route trace, not a statistically estimated mean for every order. A separate 100-order diagnostic sample contains 20 late orders, defined as handover more than 120 minutes after acceptance. The normal trace and late-order sample therefore answer different questions.
Three pickers, three packers, two checkers, and two loaders work in parallel within their respective stages. At the illustrated processing times, theoretical stage capacities are 157.5, 252, 420, and 280 orders per day. Picking is below the 240-order requirement even before interruptions or rework. Release has two staff and a theoretical capacity of 420. These capacities assume dedicated resources, identical order mix, and no overlapping assignments. They are not observed outputs and should never be reported as effective capacity.
Separating the 100 orders by replenishment hold gives 8 late of 12 held orders and 12 late of 88 others. The held group accounts for 40% of the late orders but only 12% of observed orders. Both perspectives matter: concentration suggests an investigation, while the other 60% of late orders require additional explanations.
For a fictional deeper split, suppose large orders include 10 held with 7 late and 20 non-held with 6 late; small orders include 2 held with 1 late and 68 non-held with 6 late. Totals reconcile to 12 held, 88 non-held, and 20 late. Large-order late proportions are 70% versus 30%; small-order proportions are 50% versus 8.8%. The direction persists, but two small held orders provide a very uncertain estimate.
| Order size | Hold | Orders | Late | Late proportion |
|---|---|---|---|---|
| Large | Yes | 10 | 7 | 70% |
| Large | No | 20 | 6 | 30% |
| Small | Yes | 2 | 1 | 50% |
| Small | No | 68 | 6 | 8.8% |
04 · Interpretation and decisions
Interpretation and decisions
Interpret the cross-classification as better description, not automatic proof. Held orders are disproportionately large, so the original contrast mixes stock condition and complexity. The within-size comparisons reduce that particular ambiguity but may still reflect release timing, route, or missing-stock type. A further targeted observation should test the actual reserve-search mechanism.
A weak application compares counts between shifts without their workload or combines sparse groups until an attractive pattern appears. A sound application reports each denominator, retains the overall reconciliation, and explains why the strata correspond to different operating conditions. Avoid attributing group differences to people when the work mix or resource conditions differ.
05 · Sound application and adaptations
Sound application and adaptations
For many factors, use a small number of theory-led cross-tabs before fitting a model. Maintain original categories so later investigators can revisit the grouping. Examine missing-category rates by shift; a logging failure can create an apparent performance contrast. Beware of aggregate reversals when the mix changes, sometimes called Simpson's paradox: the overall comparison can reverse the within-group pattern.
Stratification does not eliminate all confounding and cannot create information in tiny cells. Resist progressively slicing the same 100 records into dozens of groups. Collect additional observations for the mechanism that matters, and use a prospective period to check whether the pattern repeats. Keep prediction and causal conclusions proportionate to the observational design.
06 · Evidence and follow-through
Evidence and follow-through
Make the observation auditable before debating a countermeasure. Preserve an anonymous order identifier, acceptance and handover timestamps, route, line count, shift, exception flag, and the observer's definitions. A timestamp from a completed screen action may describe data entry rather than physical completion. Reconcile several records with direct observation and ask the person doing the work what each event actually means. Keep an uncertainty note when clocks differ or two actions overlap; rounding cannot repair an ambiguous boundary.
Observe ordinary work and exceptions over a period that includes opening, replenishment, breaks, and the carrier cutoff. Selecting only easily followed orders produces a clean-looking but biased picture. When an order is unfinished at observation end, keep it visible as unfinished rather than silently deleting it. For multi-line orders, establish whether timing ends at the first line or the complete order. Include all three lines in the latter case, because the customer receives the order only after its final requirement is satisfied.
Turn the analysis into a bounded decision. Describe the proposed change, the mechanism expected to alter performance, the measure that would detect the change, and the conditions under which it would be withdrawn. A small layout change might reduce walking yet increase congestion; an extra release wave might lower one queue yet overload picking. Monitor handover performance, first-pass accuracy, work in process, and staff workload together. Treat the proposed numbers as planning assumptions until a comparable observation period supplies evidence.
07 · Boundary and robustness review
Boundary and robustness review
Before implementing the selected countermeasure, check the boundary one more time. If the analysis starts after order release, an upstream acceptance hold is outside its measured lead time even though the customer experiences it. If the analysis ends at the loading bay, a parcel waiting for the carrier is still incomplete within the stated dispatch outcome. Keep those events inside the customer-facing boundary, or state exactly why a narrower operational measure is also being used.
Assess robustness across normal variation. A proposed operating rule should specify what happens when an urgent order arrives, stock is unavailable, a staff member is absent, or the carrier changes collection time. Those conditions do not invalidate the main analysis; they define where its assumptions need adaptation. Rehearse an exception through the proposed route with the people responsible for each decision. Record which assumptions remain unverified and what evidence would be sufficient to proceed.
Finish with an explicit learning statement: what the evidence currently supports, what it does not establish, and which observation is needed next. A quantified hypothesis can be useful without becoming a guaranteed result. Retain the current-state observations so a later comparison can use the same event definitions and population. When the observed outcome differs from the prediction, revisit the mechanism and resource assumptions rather than adjusting the target after the fact.
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.