Histogram: See the Shape of Your Data Before You Blame the Average

A histogram sorts a single measurement column into bins and counts how many values land in each one. The result is a picture of shape. You see where the data clusters, how far it spreads, and whether it leans one way.

That matters in the Analyse phase of RDMAICS because an average hides almost everything. Two processes can share the same mean and behave in completely different ways. One is tight and predictable. The other swings wide. Only the shape tells you which is which.

The tool does two jobs. It builds one histogram for a single measurement column. It also compares grouped and stratified distributions on purpose, so you can see whether a difference between groups is real or just noise in the eye.

Use it when you have a continuous measurement, enough observations to form a pattern, and a question about variation rather than a question about totals.

The example used throughout this guide is fictional. It ships with the tool as sample data so you can click through the whole workflow before you load anything of your own.

The scenario is a three store food service operation measuring a service time in minutes. Each store is measured across three trading periods, Lunch, Dinner and Weekend, over two weeks. That gives eighteen observations.

The question behind the data is simple. The business suspects that one store is slower than the others, but the weekly averages keep arguing with each other. A histogram is the right next step because it shows the full spread rather than a single summary number.

The input table has four columns. The first is an observation or period label, such as Store 1 Lunch. The second is the measurement itself, which is the only required field. The third is the group or stratum. The fourth is the series or scenario.

In the sample data the measurement column runs from 25.8 up to 39.8. The group column holds Store 1, Store 2 and Store 3. The series column holds Week 1 and Week 2. Those two extra columns are what turn a single chart into a comparison.

Keep one row per observation. Do not pre-aggregate. If you paste in averages, you have already thrown away the variation the histogram exists to show.

The labels in the first column are for your benefit, not the maths. They help you trace an unusual bar back to a real event.

Above the chart sit the display controls. Compare by is set to Group or stratum, so the tool splits the eighteen values into three store distributions rather than pooling them into one.

Measure is set to Frequency count, so each bar height is a plain count of observations. Orientation is Vertical columns. Comparison layout is Side by side, which places the three store distributions next to each other instead of stacking or overlaying them.

Fitted normal curve is switched on. That draws a reference curve over the bars so you can judge how far the shape departs from a symmetric bell. Cumulative percentage line is switched off, which keeps the chart clean while you are still looking at shape rather than at percentiles.

Change one setting at a time. If you flip the comparison basis and the layout together, you will not know which change produced the new picture.

With Compare by set to Group or stratum, the tool draws three distributions side by side. Read them in a fixed order every time. Centre first, then spread, then shape, then anything that sits on its own.

Store 2 holds the lowest values in the data set. Its six observations are 25.8, 29.4, 27.6, 26.9, 30.1 and 28.2. They sit in a narrow band. Store 3 holds the highest, with 30.5, 39.8, 34.7, 35.3, 38.6 and 33.9. The gap between its lowest and highest value is wider. Store 1 sits between the two, with values from 28.4 to 36.2.

The fitted normal curve gives you a visual check rather than a verdict. Where the bars sit well away from the curve, the distribution is skewed or the sample is small. With six observations per group, both are plausible, and the curve should be treated as a sketch.

The tool also offers an AI review that explains the result in words. Read your own chart first. Form a view, then compare it with the review, rather than the other way round.

A histogram does not make a decision for you. It narrows the list of sensible next moves.

In this case the stratified view points the investigation at Store 3 rather than at the process as a whole. If one store carries both the highest values and the widest spread, a global process change is unlikely to be the efficient answer. You would go and look at what Store 3 does differently.

The chart also supports a measurement decision. Six observations per group is thin. A reasonable next step is to collect more data before committing to any change, because the shape of a six point distribution can move a long way with one more reading.

Finally, it supports a framing decision. Switching Compare by to the series column would show Week 1 against Week 2 instead. That tells you whether the pattern is stable across time or whether something changed between the two weeks.

The most frequent error is reading a bar as a cause. A tall bar means many observations landed in that range. It says nothing about why.

The second is over-reading small samples. With eighteen rows split three ways, a single value can create or erase a bar. Treat the overall position and width of each distribution as the signal and the individual bars as noise.

The third is mixing populations without meaning to. If you pool Lunch, Dinner and Weekend into one chart, you may see two humps and conclude the process is unstable. The humps may simply be different trading periods. That is exactly what the group and series columns are for.

The fourth is trusting the fitted normal curve too far. It is a drawing aid. It is not a test of normality and it does not confirm that your data follows any particular distribution.

The histogram will not tell you why. It shows distribution, not mechanism. Store 3 sitting to the right of the others is a finding, not an explanation. The explanation comes from going to the process and asking questions.

It will not tell you whether a difference is statistically significant. The tool counts and draws. It does not run a hypothesis test, and the side by side layout can make a modest gap look decisive simply because the groups are drawn apart.

It will not show you time order. Every observation in the sample data is collapsed into a bin regardless of when it happened. A drift, a step change or a one off spike all look the same inside a histogram. If sequence matters, you need a run chart or a control chart alongside it.

It will not detect bad data. If a value was recorded wrongly, the histogram will place it in a bin without complaint. It will also not tell you whether your measurement system is capable of resolving the differences you are looking at.

Finally, it will not make the decision. It will not weigh cost, effort, customer impact or risk. It gives you a shape. What you do with that shape stays with you and the people who run the process.