Data Stratification (AI): Find the Factor That Explains Your Variation

Data Stratification (AI), Measure phase tool guide from Ci Flow

What the tool is for, and when to reach for it

Data Stratification (AI) compares process performance across shifts, machines, suppliers, teams, products or any other factor you record. You give it one outcome measure and one grouping column. It returns box plots, confidence intervals, histograms, an individual value plot, a two factor heatmap, a trend over time and a statistics table. Every chart is calculated from the worksheet you can see.

It belongs in the Measure phase of RDMAICS. By Measure you have defined the problem and you are collecting data. The question that matters at this point is not what the average is. It is whether the average hides groups that behave very differently.

Reach for stratification when a single summary number feels wrong. A mean of thirty minutes can describe a steady process or two processes stuck together. Stratification is how you tell the difference before you spend time in Analyse.

You suspect one shift, site or method behaves differently from the rest.
Your control chart looks noisy and you want to know whether the noise is structured.
A single mean is being quoted in a meeting and nobody has checked underneath it.
You want to show a sponsor where to focus, not just that a problem exists.

The case: pizza delivery times across four weeks

The example used throughout this guide is fictional. It is the sample data built into the tool, and it exists so you can practise the method without risking a real data set.

A delivery operation runs three stores named Central, North and South. Each delivery is logged with a week, a store, a shift of Lunch or Dinner, a dispatch method, the delivery time in minutes, and a late flag. Three dispatch methods are in use: Manual dispatch, Traffic-aware and Nearest driver.

The worksheet holds 48 deliveries across 2026-W1 to 2026-W4. Delivery times run from 24.0 minutes at the fast end to 41.4 minutes at the slow end. The operations lead wants to know what explains that spread. Is it the store? The shift? The week? Or the way the job is dispatched?

That is exactly the question stratification answers. You pick one factor to test first, and the tool tells you how much of the variation that factor accounts for.

Setting up the data and the inputs

The worksheet is the whole input. Each row is one delivery. Each column is one attribute of that delivery. Nothing is pre-aggregated, which matters, because the tool needs individual observations to calculate spread, quartiles and outliers.

In the business question field the context reads: pizza delivery time segmented by dispatch method, shift, store and week. That line is optional, but it keeps the analysis honest. It states what you are comparing before you see the answer.

The inputs are then set as follows. Outcome to analyse is Delivery time (minutes), numeric. Primary stratification factor is Dispatch method, categorical. Second factor is Shift, categorical. Time or sequence is Week, categorical. Improvement direction is set to lower is better. The fitted normal curve option is switched on.

Only the outcome and the primary factor are required. The second factor unlocks the heatmap. The time column unlocks the stratified trend chart. Setting the improvement direction tells the tool which end of the scale counts as good, so a low mean is read as a strong performer rather than simply a small number.

The inputs sit above the worksheet, so you can see exactly which columns feed the outcome, the primary factor, the second factor and the time axis. Note that every row is a single delivery, not a summary.

Reading the result

The headline banner reports a verified stratification finding: a large practical separation across 3 strata, with eta squared of 0.9255. Manual dispatch averages 38.875 minutes, compared with Traffic-aware at 26.445 minutes. The counters confirm 48 rows analysed, 3 strata compared and 0 incomplete rows excluded. The p value is reported as 0.

Eta squared is the share of total variation explained by the grouping factor. At 0.9255 the dispatch method accounts for almost all of the spread in delivery times. That is unusually clean, which is typical of sample data and rare in the field, but the reading method is the same either way.

The statistics table carries the exact values. Manual dispatch has n=12, mean 38.875, median 38.85, standard deviation 1.7945, range 36.4 to 41.4. Traffic-aware has n=20, mean 26.445, median 26.15, standard deviation 1.523, range 24.0 to 29.1. Nearest driver has n=16, mean 29.8563, median 29.65, standard deviation 1.006, range 28.5 to 31.6. No stratum records any outliers.

The box plots show three separated boxes with no overlap. The mean with 95 per cent confidence interval chart shows three intervals that do not meet. The individual value plot shows every one of the 48 observations, so you can confirm the separation is not driven by a handful of points. The histogram with the fitted normal curve shows three distinct humps rather than one. The tool also offers an AI review that explains the result in plain language alongside the charts.

The box plots, confidence intervals and statistics table all tell the same story, which is the check you want. Notice that the three confidence intervals do not overlap.

The second factor and the time view

The two factor heatmap crosses dispatch method with shift. Manual dispatch shows 38.875 minutes at Dinner with n=12, and an empty cell at Lunch with n=0. Traffic-aware shows 28.025 at Dinner with n=8 and 25.3917 at Lunch with n=12. Nearest driver shows 30.7125 at Dinner with n=8 and 29.0 at Lunch with n=8.

That empty cell is the most useful thing on the screen. Manual dispatch is never used at Lunch in this data. So the Manual dispatch mean is a Dinner only mean, and Dinner runs slower for every method. Part of the gap you are looking at is a shift effect riding along with the method.

The stratified trend chart plots the mean for each method across 2026-W1 to 2026-W4. The three lines stay separated across all four weeks. The ranking does not change. That tells you the difference is persistent rather than a one week event.

Read the heatmap before you write your conclusion. A large eta squared on the primary factor is a starting point, not a verdict.

The full screen, from inputs through to the heatmap and trend chart. The n=0 cell for Manual dispatch at Lunch is the detail most people miss on first read.

The decisions this supports

Stratification narrows a project. Before this analysis the improvement question was broad: how do we reduce delivery time. After it, the question is specific and testable.

With the numbers in front of you, several moves become reasonable to propose and defend.

None of these are conclusions the tool draws for you. They are decisions you make, using evidence the tool has laid out from the worksheet.

Scope the project around dispatch method rather than around stores, because the method strata separate cleanly.
Investigate the Dinner shift separately, since Manual dispatch appears only at Dinner and every method runs slower there.
Use Traffic-aware as the comparison point at 26.445 minutes, since it holds the lowest mean under a lower is better direction.
Stop treating week to week movement as a signal, because the trend lines hold their order across all four weeks.
Collect Manual dispatch data at Lunch if you want to separate the method effect from the shift effect.

Common mistakes

The first mistake is stratifying on the factor you already blame. If you only ever test one grouping, you will confirm your assumption and learn nothing. Run the outcome against store, against shift and against week as well, then compare the eta squared values.

The second is confusing separation with cause. The three dispatch methods differ by a large margin here. That does not prove the dispatch logic causes the difference. In this data the methods are not evenly spread across shifts, and the heatmap shows it.

The third is reading small strata as if they were large ones. Each cell count is printed for a reason. A mean from n=8 carries more uncertainty than a mean from n=20, and the confidence interval chart is where you check that.

The fourth is feeding in pre-summarised rows. If you upload weekly averages instead of individual deliveries, the standard deviations, quartiles and outlier counts all describe the summaries, not the process.

Honestly, on the limits of the tool

Data Stratification (AI) will not tell you why a stratum differs. It reports that Manual dispatch averages 38.875 minutes and Traffic-aware averages 26.445 minutes. The reason sits in the process, not in the file. You still need to go and look.

It will not prove causation. Eta squared of 0.9255 measures how much variation lines up with the grouping. It cannot separate the method from anything else that travels with the method, such as the shift it runs on. The n=0 cell in the heatmap is a confounding warning, not a data entry error.

It will not fix a biased sample. If the 48 rows are not representative of how the process actually runs, every statistic on the screen is a faithful description of a skewed picture. It will also not judge whether your grouping column is meaningful. If you stratify on a label that was applied inconsistently, the tool will still produce clean boxes.

It will not choose your target, cost your improvement or predict what happens after a change. It reports no outliers here, which is a statement about the 1.5 times IQR rule applied to this data, not a statement that nothing unusual ever happens. And the fitted normal curve is a visual reference, not a normality test.

Used inside Measure, with those limits understood, the tool does one job well. It shows you where the variation lives, so you arrive at Analyse with a narrower and better evidenced question.

data stratification toolMeasure phase RDMAICSstratified box plot analysiseta squared stratificationprocess variation by factorLean Six Sigma stratificationCi Flow data stratification

Run this method inside Ci Flow

RDMAICS, A3, Just Do It and Kaizen projects with the full Lean Six Sigma toolset, AI interpretation and tollgate presentations built from your real evidence.