Problem Solving & Quality · Observe

Stratification

Split the data by shift, line, operator, supplier, raw-material lot or product — a flat average often hides one guilty stratum.

Time45 min
FormatSolo
StageObserve

Stratification: what it is and why it works

Stratification splits a data set by the factors that could separate it (shift, line, operator, supplier, raw-material lot, product, season) and compares the resulting groups side by side. The aim is to find a stratum that behaves differently: a higher level, a wider spread or a different trend. For this to work, every data record must carry the labels of the factors of interest, ideally added when the data is collected. The team splits by one factor at a time, plots each stratum, and takes any suspicious group back to the floor to ask what is specific about it.

An overall average blends good and bad performance and can look acceptable while one line, one supplier or one product carries most of the loss. Stratification is often the fastest way to reduce a big, diffuse problem to a specific one, and it needs little statistics: side-by-side histograms, box plots or run charts usually suffice. It is one of the seven basic quality tools and works hand in hand with check sheets, which should be designed with the right columns, and with Is / Is-Not, whose entries often come straight from strata. Scatter diagrams benefit in the same way when points are colored by group. When a difference is small relative to the noise, a formal comparison test helps confirm it is real.

What you need

Data on the indicator of interest over a representative period
Labels on each record: shift, line, supplier, lot, product and so on
A list of factors that could plausibly separate the data
Simple plotting tools: spreadsheet, statistics software or graph paper

What you get

Side-by-side plots of the indicator by stratum
Identification of any stratum with a different level, spread or trend
A narrower problem scope
Questions to take back to the floor about the suspicious stratum
Labeling requirements for future data collection

When to use it

When the overall figure looks acceptable but complaints keep coming from somewhere.

How to do it, step by step

List the factors that could separate the data: shift, line, operator, supplier, raw-material lot, product, season.
Make sure each data record carries those labels; add them at collection if missing.
Split the indicator by one factor at a time and plot each stratum side by side.
Look for a stratum that behaves differently — higher level, wider spread, different trend.
Take the suspicious stratum back to the floor and ask what is specific about it.

Worked example: Underweight bags of mortar mix

Illustrative scenario — figures are realistic but not from a real company.

A producer of 50-lb bags of dry mortar mix averages 50.9 lb per bag, comfortably above the label weight, yet a large distributor keeps rejecting pallets for underweight bags. The plant runs two packing lines on three shifts, fed from two cement silos.

The team listed candidate factors: line, shift, silo, bag supplier and time since the last packer calibration. Checkweigher records already carried line, time and silo; bag supplier was added by linking to the packaging lot log.
Split by line: line 1 averaged 50.9 lb with a standard deviation of 0.2 lb; line 2 had the same average but a standard deviation of 0.6 lb, three times the spread.
Split by silo within line 2: the spread was normal on silo A and wide on silo B. Shift and bag supplier showed no difference.
Split by time: silo B's spread widened in the hours after each fresh cement delivery was blown into the silo.
On the floor, operators explained that freshly delivered cement is heavily aerated and flows unevenly for a while, which upsets the packer's fill control.

Result. A settling period after each delivery and a fill-control adjustment for silo B cut line 2's standard deviation to 0.25 lb, and the distributor's rejections stopped. The plant-wide average had hidden the fact that almost all underweight bags came from one line, one silo and a few hours after each delivery.

Open it in the FieldKit Problem Solving app → Get the Problem Solving template pack — join the waitlist

Common pitfalls and how to avoid them

Collecting data without labels.Decide the strata before collection and record the labels with every data point.
Splitting by many factors at once.Split by one factor at a time first; combine factors only once a suspicious stratum appears.
Comparing averages only.Look at spread and trend too; many problems live in variation, not in the average.
Over-reading small groups.Check how many points each stratum holds; a difference based on a handful of points may be chance, so collect more or use a formal test.

Frequently asked questions

What is stratification in quality control?

Stratification is the practice of separating data into groups according to factors such as machine, shift, operator, material lot or supplier, and comparing those groups. It shows whether a problem is concentrated in one part of the process rather than spread evenly. It is one of the seven basic quality tools and is often the quickest way to narrow down a broad problem.

How is stratification different from a Pareto chart?

A Pareto chart ranks categories of problems (defect types, causes, cost items) by size. Stratification splits the data for one indicator by where it comes from (line, shift, supplier) to see which source behaves differently. The two are often combined: the Pareto chart picks the defect to attack, and stratification shows where that defect comes from.

What if no stratum stands out?

Then the factors chosen may not be the relevant ones, or the problem may be common to the whole process. Try other factors, such as raw-material properties, ambient conditions or time since maintenance, or a finer time resolution. If nothing separates the data, the cause is probably systemic, and a process map, a fishbone diagram or a designed experiment is a better next step.

Origin

Stratification — one of the seven basic quality tools of Japanese quality control (Kaoru Ishikawa, JUSE).

Used in these playbooks

Chronic scrap reduction month 1 month

One month against a loss everybody has learned to live with: count at the source, rank the losses, split the data, list the causes and prove the real one before spending money.

Related methods

More in “Observe”

Go where the problem happens and collect facts, counts and timelines — not opinions.