I recently reviewed a supplier's failing capability data. The plant quality manager presented a control chart for a critical machined dimension. A thousand data points were scattered across the chart with no visible trend, structure, or assignable cause. The calculated Cpk sat stubbornly at 0.87, declaring the process entirely incapable.
The manager insisted the machine was calibrated, the raw material carried a valid certificate, and the operators were fully trained. He claimed to have exhausted every troubleshooting avenue. But he had only exhausted every avenue using aggregated data. He was attempting to diagnose a multi-variable manufacturing process using a single, blended population.
Stratification is the mechanism that fractures that blended population. By dividing data into homogeneous layers before analysis, you isolate the specific variables driving variation. In this case, breaking the data down by shift revealed that night shift operators were changing tooling at different intervals than day shift. One simple layering exercise solved a problem months of aggregate analysis could not.
The Mechanism of Stratification
Stratification is the practice of splitting data into distinct subgroups that share a common characteristic before you begin analysing it. It is one of the fundamental seven basic quality tools, yet it is frequently the most neglected. Teams readily deploy Ishikawa diagrams and Pareto charts, but they routinely skip the crucial step of segmenting their raw data into meaningful layers.
Aggregated data actively conceals reality. When you mix results from three shifts, two machines, and five material deliveries into a single histogram, the resulting curve averages out the very variation you need to investigate. A Cpk of 1.33 across an entire plant means very little if a single machine is producing scrap while the others operate at a six-sigma level to compensate for it.
I implemented routing verification KPIs at a major aerospace manufacturer that cut internal lead time by 97 percent. That improvement was not achieved by looking at overall factory output. It was achieved by stratifying the data to see exactly where specific bottlenecks occurred across individual work centres and shifts. You cannot optimise what you have not properly segmented.
Fracturing the Data: A Practical Case
Return to the supplier facility with the 0.87 Cpk. When I asked the quality manager if he had analysed the data by shift, he looked at me as if I had asked him for his password. He insisted all shifts operated identically because they followed the same standard work. Procedure does not guarantee identical process behaviour.
It took twenty minutes to slice the existing data into three shifts. The results immediately invalidated his assumption of uniformity. The day shift was highly capable, the afternoon shift was acceptable, and the night shift was generating the entirety of the process failure.
| Shift | Cpk | Mean | Standard Deviation |
|---|---|---|---|
| Morning (A) | 1.45 | 12.503 mm | 0.008 mm |
| Afternoon (B) | 1.33 | 12.498 mm | 0.009 mm |
| Night (C) | 0.62 | 12.487 mm | 0.021 mm |
The night shift had three times the variability and a drifted process centre. A Gemba walk revealed that the night operator was changing the cutting tool every 50 parts instead of the specified 30, simply to avoid stopping the line. Furthermore, the inspection station relied on natural daylight; at night, the poor artificial lighting caused micrometer misreads. Stratification found the root cause in thirty minutes.
Multi-Level Analysis in Automotive Quality
Averaging data masks the exact failures your customer experiences. On one IATF 16949 project, an OEM reported a 2.3 percent defect rate on plastic connectors. The Tier 1 supplier's internal data showed a defect rate of just 0.4 percent. The supplier was defending their numbers instead of investigating the discrepancy, severely damaging the customer relationship.
We stopped debating the aggregate numbers and started stratifying the data. First, we separated the results by the three production lines. Lines one and two were operating at 0.2 and 0.3 percent respectively. Line three was running at 1.1 percent. We had found a hotspot, but it still did not account for the full 2.3 percent defect rate reported by the OEM.
Next, we stratified line three by material supplier. Material from Supplier A yielded a 0.3 percent defect rate. Material from Supplier B yielded a 2.8 percent defect rate. The supplier had been blaming the OEM's handling, but the data proved the material source was the primary variable.
A final stratification revealed the true failure mode. Connectors from Supplier B were packaged in smaller, unstable containers. The defects were not occurring during moulding; they were occurring during freight. Changing the packaging for that specific material flow dropped the overall defect rate to 0.15 percent in three days.
Executing the Method Step by Step
Effective stratification requires a systematic approach to data capture and analysis. You cannot retrofit context to data after the fact. If you want to isolate variables, you must tag them correctly during collection. Modern MES platforms handle this automatically, but manual Excel tracking requires strict column discipline.
The Stratification Sequence
- 01Define the MetricSpecify the exact dimension, tolerance, and measurement method (e.g., CMM, micrometer).
- 02Identify FactorsList potential variables: shift, machine ID, operator, material batch, method, environment.
- 03Tag the DataEnsure every individual measurement carries metadata identifying its specific origin.
- 04Visualise LayersGenerate box plots or overlay histograms to visually separate the subgroups.
- 05Test SignificanceRun ANOVA or t-tests to confirm the visual difference is statistically valid.
Visualisation is where the investigation gains traction. A box plot is highly effective because it displays the median, quartiles, range, and outliers for each layer simultaneously. Overlay histograms clearly show bimodal distributions, which immediately indicate that you are mixing two distinct process populations into one chart.
However, a visual check is only a starting point. You must confirm the difference using statistical tests. If you are comparing two layers, use a two-sample t-test. If you are comparing three or more layers, use ANOVA. Never act on a visual illusion; human cognition is hardwired to find patterns even in random noise.
Integrating Stratification with Core Quality Tools
Stratification does not operate in isolation. It is the mechanism that connects your descriptive statistics to your causal analysis. When deployed correctly, it amplifies the effectiveness of every other quality tool in your IATF 16949 or AS9100 toolkit.
During 8D problem solving, teams often build an Ishikawa diagram to brainstorm potential causes. Every major bone on that fishbone represents a potential stratification factor. The diagram tells you what could be causing the failure; stratification tells you if it actually is.
Aggregate data will tell you that you have a problem. Stratified data will tell you exactly where to stand to fix it.
Apply the same logic to your Pareto analysis. A single Pareto chart across an entire plant often highlights generic issues. Stratify that Pareto chart by shift or production line, and you will find that the night shift's vital few defects are completely different from the day shift's. This forces you to target your containment actions precisely.
Stratification also feeds directly into your PFMEA. Instead of estimating arbitrary Risk Priority Numbers based on theoretical failures, you can use layered historical data to prove exactly how often a specific failure mode occurs under specific conditions. This transforms your PFMEA from a paperwork exercise into a validated risk model.
Common Failures and Statistical Traps
The most frequent error I see in supplier audits is attempting to stratify with insufficient data. If you have 15 total measurements and you try to break them down across five different shifts, you are left with three data points per layer. That is not statistical analysis; that is guessing. You need a minimum of 30 measurements per layer to establish a valid baseline.
Another critical trap is ignoring interactions between variables. A specific operator might perform flawlessly on machine one but generate scrap on machine two. Single-level stratification will miss this entirely. You must be prepared to use multi-level stratification or Design of Experiments to uncover interacting variables.
Single-Layer vs Multi-Layer Analysis
What teams typically do
- Review a single aggregate Cpk for the department
- Stratify by machine, but ignore the operator
- Stop investigating once one variable looks different
- Rely on visual checks without statistical confirmation
What drives actual yield
- Demand Cpk broken down by individual asset and shift
- Cross-reference machine data with material batches
- Use ANOVA to validate findings across all variables
- Run multi-level analysis to catch hidden interactions
Finally, never wait until a problem occurs to start tagging your data. Retrospective stratification is impossible if your measurement logs only contain a number and a date. Every measurement recorded on your shop floor must carry metadata: machine ID, operator code, material batch, and timestamp. If you collect this context daily, you will save weeks of investigation when a crisis hits.