Box Plot: See Your Baseline Spread And Find The Rows Behind The Outliers

A mean hides more than it reveals. Two processes can share the same average and behave in completely different ways. One runs tight and predictable. The other swings wide and burns your capacity on rework. If you only report the average, you will not see the difference.

The Box Plot tool exists to describe the baseline distribution, expose stratification and trace unusual points back to their source rows. Those three jobs sit squarely in the Analyse phase of RDMAICS. You have already recognised the problem, defined the scope and measured something. Now you need to understand the shape of what you measured.

Use it early in Analyse, before you reach for anything more complex. A box plot is cheap to build and quick to read. It often tells you which comparison is worth testing formally, and which one is noise. It also tells you whether you have one process or several processes wearing the same name.

Reach for it when you suspect a difference between shifts, sites, operators, product lines or time periods. Reach for it when someone quotes a single average and you want to know what sits underneath.

The worked example here uses the sample data that ships with the tool. It is fictional. No real organisation, site or team is described. The point is the method, not the numbers.

The scenario is straightforward. A team is comparing a standard way of working against a pilot way of working. They have collected twenty four observations in total. Each observation carries a measurement value, a group label of either Standard or Pilot, and a second factor label of either Day or Night.

There are six observations in each of the four combinations. Standard Day runs from 30 to 35. Standard Night runs from 41 to 49. Pilot Day runs from 24 to 28. Pilot Night runs from 29 to 35. The row identifiers follow a plain naming pattern, such as Standard-Night-4 or Pilot-Day-2, so every point on the chart can be traced back to a specific row.

That naming discipline matters more than it looks. When the chart flags an unusual point, you want to know exactly which observation it was without hunting through a spreadsheet.

The tool asks for four columns and nothing more. An observation or row identifier, a measurement, a group or stratum, and a second factor or scenario. In the sample data those map to Observation / Row ID, Measurement, Group / stratum and Second factor / scenario.

The measurement column is the only one that must be numeric. Everything else is a label. Keep the labels consistent. Standard and standard will be read as two different groups, and you will end up with four boxes where you expected two.

The second factor column is optional in principle but it carries most of the insight in this case. Without it you would see two boxes, one for Standard and one for Pilot. With it you can split each group by Day and Night and see where the spread actually comes from.

Twenty four rows is a small data set. It is enough to demonstrate the tool and enough to see a pattern, but it is not enough to make a confident claim about a process. Treat the sample as a teaching example, not a template for sample size.

The settings used for this walkthrough are deliberate. Compare by is set to Group / stratum. Orientation is set to vertical boxes. Sort displayed groups is set to worksheet order, which keeps Standard before Pilot rather than reordering by value.

Individual observations is switched on. This draws every one of the twenty four data points over the boxes. With small samples this is the single most useful option in the tool. Six points per combination is not many, and seeing them stops you from reading a box as though it summarised hundreds of readings.

The mean marker is switched on as well. Showing the mean alongside the median lets you see skew at a glance. When the two sit apart, the distribution is leaning one way.

Density overlay and median confidence notch are both switched off here. Both are useful on larger data sets. On six points per group they add visual weight without adding information. Outlier row labels are switched on, which is what connects a flagged point back to its source row.

Start with position. The Pilot values sit lower than the Standard values across the board. Pilot Day spans 24 to 28. Standard Day spans 30 to 35. There is no overlap between those two ranges at all. Pilot Night spans 29 to 35 and Standard Night spans 41 to 49, again with no overlap.

Now look at spread. Standard Night is the widest set in the data, running from 41 to 49, a range of eight. Pilot Day is the tightest, running from 24 to 28, a range of four. Width tells you about consistency. A wide box means the process behaves differently from one run to the next, and that variation is usually where the cost sits.

Then look at stratification. This is where the second factor earns its place. Within Standard, Day sits between 30 and 35 while Night sits between 41 and 49. Those are not two tails of one distribution. They are two different behaviours. The same split appears in Pilot, where Day runs 24 to 28 and Night runs 29 to 35, though the gap is smaller.

If you had pooled Standard into a single box, you would have seen one wide spread from 30 to 49 and concluded the process was simply unstable. Splitting by the second factor shows the real story. The shift is doing the work. The tool also offers an AI review that explains the result in plain language if you want a second read.

A box plot does not make decisions. It narrows them. In this case the chart points to three practical moves.

The first is scope. If night behaviour differs this clearly from day behaviour in both groups, then any improvement work needs to treat them separately. A single root cause analysis across both shifts will blur the signal.

The second is where to aim. Standard Night carries both the highest values and the widest spread. That combination usually marks the best place to start, because you can attack the level and the variation at the same time.

The third is what to test next. The chart suggests the Pilot sits lower than the Standard, and that night sits higher than day in both. Those are hypotheses, not conclusions. The box plot tells you the comparison is worth a formal test. It does not perform the test.

The most frequent error is pooling data that should be split. If you have a second factor and you ignore it, your boxes will look unstable when the underlying processes are each quite stable. The sample data shows this clearly. Pooling Standard produces a wide box. Splitting it produces two narrower ones.

The second error is treating a flagged outlier as a data error. An outlier is a point that sits far from the rest of its group. Sometimes it is a typing mistake. Often it is a real event worth investigating. The row labels exist so you can go and ask. Do not delete points because the chart marked them.

The third error is over reading small samples. Six observations per combination will produce a box, but the quartile positions are sensitive to a single value. Keeping individual observations switched on is the honest way to display this, because the reader can count the dots.

The fourth error is sorting by value when the order carries meaning. Worksheet order is the right choice here because Standard and Pilot form a deliberate pair. Reordering by median would obscure that pairing.

The Box Plot tool describes. It does not explain and it does not prove. That boundary is worth stating plainly, because box plots are often presented as though they settled an argument.

It will not tell you whether a difference is statistically significant. The Pilot values sit below the Standard values in this sample, and the separation looks clean, but a box plot carries no p value and no confidence interval unless you turn on the notch, and even then a notch is an approximation. If you need a formal answer, run a hypothesis test.

It will not tell you why. The chart shows that night values are higher in both groups. It does not know whether that comes from staffing, lighting, handover timing, equipment or something nobody has written down. Causation comes from going and looking, not from the chart.

It will not fix a flawed measurement system. If the measurement definition drifted between shifts, the box plot will faithfully display that drift as process variation. Confirm your measurement is sound before you read anything into the shape.

It will not handle time order. These twenty four rows are treated as four groups, not as a sequence. If the measurement changed steadily over the collection period, a box plot will hide it. Use a run chart or control chart for that question.

Finally, it will not tell you whether your sample represents the process. Six observations per combination is what the sample data holds. Whether six days, six nights or six weeks of data reflect normal behaviour is a judgement you make, not a calculation the tool performs.