Data Stratification: Find Out Which Group Is Driving Your Average

Data Stratification compares process performance across shifts, machines, suppliers, teams, products or any other factor you record. You pick one outcome to analyse. You pick one factor to split it by. The tool then builds every chart from the worksheet you can see on screen.

It belongs in the Analyse phase of RDMAICS. By that point you have a defined problem and measured data. What you usually do not have is a credible answer to the question every sponsor asks: where is the problem coming from? An overall average hides that. A stratified view exposes it.

Reach for the tool when you suspect the process behaves differently in different places. Three shifts that feel uneven. Four suppliers with the same specification. Two dispatch methods running side by side. Stratification turns that suspicion into a chart with numbers under it.

The example used throughout this guide is fictional. It is the sample data that loads with the tool, and no real organisation is involved.

A pizza business runs three stores called Central, North and South. Deliveries go out across two shifts, Lunch and Dinner. Dispatch happens one of three ways: Manual dispatch, Traffic-aware routing, or Nearest driver. Each delivery is logged with a delivery time in minutes and a late flag.

Forty eight deliveries were recorded across four weeks, from 2026-W1 to 2026-W4. The business question entered into the tool reads: pizza delivery time segmented by dispatch method, shift, store and week. That is a good question because it names the outcome and the suspected splits in one line.

Before stratifying, the overall picture is unhelpful. Times range from 24 minutes to 41.4 minutes. Some deliveries are flagged late and some are not. There is clearly variation. The whole point of the exercise is to find out what the variation is attached to.

The worksheet holds one row per delivery. Each row carries a Delivery ID, a Week, a Store, a Shift, a Dispatch method, a Delivery time in minutes and a Late flag. That shape matters. One row per event, with the group labels sitting beside the measurement, is what makes stratification possible.

The inputs are short. Outcome to analyse is set to Delivery time (minutes), marked numeric. Primary stratification factor is set to Dispatch method, marked categorical. Second factor is set to Shift. Time or sequence is set to Week. Improvement direction is set to lower is better, which tells the tool which end of the scale is good. The fitted normal curve option is switched on.

Note what happens with the second factor and the time field. They are optional. Adding Shift unlocks a two factor heatmap. Adding Week unlocks a trend over time, split by stratum. If you leave them blank the core comparison still runs.

The tool reports 48 complete observations and 0 incomplete rows excluded. Check that line every time. If rows are being dropped for missing values, you want to know before you read anything else.

The headline result is a verified stratification finding: a large practical separation across 3 strata, with eta squared of 0.9255 and p reported as 0. Manual dispatch averages 38.875 minutes. Traffic-aware averages 26.445 minutes. That gap is the finding.

Eta squared is the share of variation in delivery time that is explained by the stratification factor. At 0.9255 almost all of the spread in delivery time lines up with which dispatch method was used. That is unusually clean, and in this fictional data set it is meant to be.

The stratum statistics table gives the exact values behind the charts. Manual dispatch: n=12, mean 38.875, median 38.85, standard deviation 1.7945, range 36.4 to 41.4, 0 outliers. Traffic-aware: n=20, mean 26.445, median 26.15, standard deviation 1.523, range 24 to 29.1, 0 outliers. Nearest driver: n=16, mean 29.8563, median 29.65, standard deviation 1.006, range 28.5 to 31.6, 0 outliers.

Several views sit on top of those numbers. Box plots show the interquartile boxes and medians, with whiskers stopping at the most extreme non-outlier. Mean with 95 per cent confidence interval shows the uncertainty around each group mean. A distribution histogram lets you choose between overall shape, a within-stratum comparison or a single stratum, with a dashed fitted normal curve as a visual reference rather than a normality test. An individual value plot shows all 48 points with slight jitter. The tool also offers an AI review that explains the result in plain language.

The two factor heatmap crosses Dispatch method with Shift. Manual dispatch shows 38.875 with n=12 at Dinner and a dash with n=0 at Lunch. Traffic-aware shows 28.025 with n=8 at Dinner and 25.3917 with n=12 at Lunch. Nearest driver shows 30.7125 with n=8 at Dinner and 29 with n=8 at Lunch.

That empty cell is the single most useful thing on the screen. Manual dispatch never ran at Lunch in this data. So part of the apparent dispatch method effect is tangled with shift. Dinner times are higher for every method that appears in both shifts, by roughly two to three minutes on the means shown.

The stratified trend over time plots mean delivery time within each week for each method, across 2026-W1 to 2026-W4. The lines stay separated across all four weeks. That tells you the gap is stable, not a one week event.

Stratification does not fix anything. It narrows where you look. With this result, a team can take a small number of concrete next steps.

First, scope the improvement to Manual dispatch rather than to delivery time in general. Second, test whether moving Manual dispatch work onto Traffic-aware routing is feasible, since that pair carries the largest gap in the data. Third, design a confirmation study that runs Manual dispatch at Lunch, because the heatmap shows that combination has never been observed.

Fourth, treat Dinner as a separate operating condition. Every method is slower at Dinner in this data. Fifth, use the week by week trend as a baseline. If you change dispatch rules in a later week, the same chart will show whether the lines move.

The most frequent error is reading a large effect size as proof of cause. Eta squared of 0.9255 says the grouping explains the variation. It does not say the grouping created it. Something else may travel with dispatch method, and in this case shift clearly does.

The second error is ignoring unequal group sizes. The strata here hold 12, 20 and 16 observations. That is fine for comparison, but a group with very few records will produce a wide confidence interval, and the tool needs at least two observations in a group to draw one.

The third error is stratifying by a factor you cannot act on. Splitting by week tells you when. Splitting by dispatch method tells you what to change. Prefer factors that map to a decision.

The fourth error is reading the fitted normal curve as a verdict on distribution shape. It is labelled as a visual reference, not a normality test, and it should be treated that way.

Data Stratification will not tell you why a group performs differently. It will show you that Manual dispatch averages 38.875 minutes while Traffic-aware averages 26.445 minutes. It will not tell you whether that is caused by routing logic, driver allocation, order volume, traffic at Dinner, or something nobody recorded. Confounding is invisible to it unless you have captured the confounder as a column.

It will not untangle overlapping factors for you. The empty Manual dispatch at Lunch cell is a real design gap in the data. The tool displays it faithfully and then leaves the judgement to you. No amount of charting recovers an observation that was never made.

It will not validate your measurement system. If delivery times are recorded inconsistently between stores, the strata will differ and the tool will report the difference as real. It assumes the numbers you supplied mean what you think they mean.

It will not forecast, and it will not quantify benefit. Nothing here states what delivery times would be after a change, or what any change would cost or save. The four week trend describes the past only. Finally, the p value reported as 0 is a display of a very small number, not a guarantee. Statistical significance in 48 fictional rows is not the same as a tested improvement in a live process. Use the output to choose where to run that test, then run it.