Understanding variation
What is the pattern telling us—and what response fits?
Every process varies. One good or bad result is not enough to judge an improvement. Plot observations in time order, understand how they were collected and use a method suited to the measure and its patterns.
Inspect the spread as well as the average
Before asking whether performance has changed, look at how much the measure varies over time. Is it fairly consistent, or does it move around a lot? Check gaps, unusual values, trends, seasonal patterns and changes in how it was recorded.
How much data might you need? If you are looking for a particular improvement, the wider the usual spread of observations, the harder that improvement may be to distinguish from ordinary variation. A sustained change may therefore need more observations before it becomes detectable. How many depends on the size and persistence of the change, the pattern in the data and the method used.
Before an experiment, compare the improvement you hope to achieve with the variation already present and the time available to collect data. If the expected improvement is small relative to the spread, consider whether the test can give a useful answer in time—or whether you first need a more informative measure, a different test or a better understanding of the sources of variation. Do not assume that reducing variation is always necessary before testing an improvement.
No detected change does not mean no improvement occurred. Consider what matters in practice, whether the change was implemented and whether more data would help. More observations do not necessarily resolve problems such as seasonality, related observations or changes in measurement. A small SD by itself does not prove the process is stable.
Look beyond the overall average. Does the measure combine different wards, shifts, patient groups or pathways? Compare useful groups and follow the work through its stages. Check definitions, case mix and the number of observations: differences are clues, not proof of cause, and small groups can be especially variable.
If the spread remains too wide for the purpose, consider a planned change to the system. Record your prediction, check implementation and balancing measures, and look for a change in the average, the variation, or both. Stratifying, examining process components and testing changes can inform one another; they need not follow a fixed sequence.
On the step-change chart, Show 3SD Limits adds each estimated stage average plus or minus three stage SDs. These bands describe observed spread; they are not targets, confidence intervals or X-mR control limits. Short stages can give unstable SDs, and stage boundaries are estimates.
Common causes and special causes
- Common-cause variation: the variation expected from the current process when it is statistically stable. If its performance is unacceptable, improve the way the system works rather than adjusting it after every fluctuation.
- Special-cause variation: evidence that the process is behaving differently from its established pattern. Investigate what changed; the chart signals a question, not the answer. Understand harmful changes and learn from beneficial ones—not every special cause should be eliminated.
A predictable process is not necessarily a good or safe one. An unpredictable process may need immediate containment as well as investigation. A serious incident can require action after one event; do not wait for a statistical signal to respond to harm.
Use a chart that fits the question
A run chart shows observations over time and can use defined rules to flag patterns. Statistical process control (SPC) charts add limits calculated for a particular type of measure and process.
Choose the chart and rules to suit the data—for example, individual measurements, proportions or event counts. Check denominators, changing definitions, seasonality and dependence between successive observations. Control limits describe expected process behaviour under the method’s assumptions; they are not targets or safety limits.
Step-change analysis looks for possible changes in average level. No detected step does not prove statistical stability or common-cause-only variation. A detected step does not identify its cause or prove an intervention worked. Different methods answer different questions.
What is an X-mR chart?
An X-mR chart monitors whether a process is behaving consistently over time and signals when something may have changed. It is useful for suitable measurements collected one at a time, in time order.
It is a pair of charts, which should be read together:
- X — individual readings: shows the value recorded each time, with a centre line and calculated limits.
- mR — moving range: shows the size of the difference between each reading and the previous one, ignoring whether it went up or down. It helps us see whether the size of the changes is unusual.
Whatever suitable measure you are tracking, the first chart shows the individual values over time; the second shows how much each value differs from the one before it.
How do you spot a possible issue?
Look for signals under the rules chosen for the chart:
- A reading outside the X chart’s limits: an unusually high or low individual value compared with the baseline.
- A moving range above its upper limit: an unusually large difference between two consecutive readings. This does not, by itself, tell you whether the process has settled at a new level.
- A non-random pattern on the X chart: a sustained run on one side of the centre line can indicate a shift even when every value is inside the limits. For example, one established rule flags eight consecutive values on the same side.
Rule sets vary. Agree the rules beforehand and apply them consistently; do not search through different rules until one produces a signal. The eight-point example is a rule for the individual-values chart, not a blanket rule for the moving-range chart.
A signal and a performance problem are different
- A signal means something may have changed. Check the data and investigate what was different. It could reflect deterioration, improvement or a measurement problem. The signal does not identify its cause.
- No signal means no unusual behaviour was detected by the selected rules. It does not prove stability or satisfactory performance. A process can be consistently too slow, too costly or unsafe.
Control limits describe process behaviour; targets and safety requirements describe what is acceptable. They are not interchangeable.
What should you do next?
- Unexpected behaviour: investigate what changed, including the measurement, operating conditions and possible effects elsewhere.
- Consistent but unacceptable performance: improve how the process works through a planned, evaluated change, rather than repeatedly adjusting it after routine fluctuations.
Use a suitable baseline and check that the method fits the data, including seasonality, changing definitions and dependence between readings. Limits estimated from very little data need particular caution. All points inside the limits do not, by themselves, prove statistical stability.
A chart prompts investigation; it does not replace judgement or action on an immediate safety concern. Do not wait for a statistical signal before responding to harm.
Looking at blue steps and a green CUSUM line instead? Read the step-change chart guide →
Optional technical explanation and chart rules
Avoid tampering—not deliberate improvement
Tampering means adjusting a process in response to routine variation as though each result has a distinct cause. Repeated reactive adjustments can add variation and obscure what is happening.
That is different from a planned test to improve an unacceptable system. Make a prediction, check implementation and possible harm, and agree when to review. Do not keep changing the intervention simply because the latest result is disappointing.
Could the pattern be seasonal?
Recurring patterns—by season, weekday or time of day—can make before-and-after comparisons misleading. A winter-to-spring improvement may reflect a usual cycle, an intervention, or both.
Examine comparable periods across enough cycles and use an approach that accounts for seasonality. One same-month comparison is a useful check, not proof. Do not assume a standard control chart or shuffled-data change-point test automatically handles seasonal patterns.
Could an extreme result be followed by a more typical one?
When cases or periods are selected because an observed result is unusually high or low, later measurements may be less extreme because some of the original extremeness reflected random variation. This is regression to the mean—not a guarantee about the next observation.
A programme launched after an exceptionally bad period may receive credit for that return. Use a longer history, an advance prediction and a credible comparison where feasible. A prediction or change-point test alone does not rule this explanation out.
Could this be a false alarm?
A statistical method can signal a change that is not a real process change. Repeatedly trying different periods, measures, thresholds or rules increases opportunities to find a misleading result.
Choose the approach in advance where possible, record settings, check assumptions and investigate the context. No chart eliminates false alarms or missed changes. Operational or safety concerns still matter even without a signal.
What’s your next useful question?
Are we responding to a meaningful pattern—or just the latest number?
Interpret your StepChange result →Return to Right measures? →
If you are investigating queues and workload rather than interpreting a chart, see Variation, demand and capacity.