Intended use of the StepChange Analyzer
In short: the StepChange Analyzer supports process improvement and the retrospective analysis of historical data. It helps you see whether, and when, a business, operational or service-level measure changed. It is not a medical device, and must not be used for diagnosis, treatment or any other decision about an individual patient.
What it is for
The analyser looks for step changes in a series of measurements over time, using the Bootstrap CUSUM method (Taylor, 2000). It has two main uses.
Process improvement. It helps people working on improvement to:
- see whether a measure has shifted, and roughly when;
- check whether a change they made coincided with a change in results;
- spot unplanned shifts worth investigating;
- ask better questions about the process that produced the data.
Retrospective analysis. It helps people look back over historical data to:
- find when a measure changed, and how many distinct periods of performance there have been;
- review whether past initiatives or policies were followed by a change;
- describe the history of a measure for reports and reviews.
Suitable data
Business and operational measures recorded at regular intervals, one value per period: for example sales, throughput, output, cycle times, complaints, defects, waiting times or service performance. This includes aggregate healthcare service data, such as the share of patients seen within a target time or the rate of adverse events across a service.
What it is not
It is not a medical device. It has not been designed, validated or approved for clinical use, and it must not be used for the diagnosis, monitoring or treatment of an individual patient, or for any decision about one person's care.
It does not explain causes or prove that an intervention worked. A detected change shows that the data shifted; why it shifted needs knowledge of the process, and other evidence.
It is not a substitute for professional judgement or specialist analysis when important decisions depend on the result.
Using the results responsibly
- What the confidence level tells you. It answers one question: could ordinary ups and downs easily produce a pattern this clear? If nothing had really changed, a 99% result would happen only about 1 time in 100. So a high confidence level is good evidence that something did change. But it isn't the exact chance that the change is real, and it can't tell you what caused it. It also assumes each value's ups and downs are independent of the last: trends, seasonal patterns and values that follow on from one another can make it look higher than it should.
- The method assumes neighbouring values are independent and that each stage has a stable average. Trends, seasonal patterns and values that follow on from one another can produce changes that are not real. Use the checks provided and look at the chart.
- A short series may show no change even when one happened. "No change found" is not proof that nothing changed.
- Compare detected dates with what you know. Change dates are estimates and can be a period or more out.
- A staircase of steps all in one direction may be one gradual change. The analyser can only show change as flat stages, so an improvement that builds up over time appears as several steps. See One gradual change, or several steps?
- In historical data, decide what you are looking for before you look. Long histories contain many events, so some event will usually fall near any detected change by coincidence. Write down the dates of the initiatives you want to test first.
- Check for changes in how the data were recorded. New definitions, systems or counting rules can produce a step change that reflects the data, not the process.
- Report the settings you used. Trying different settings or date ranges until a change appears or disappears makes the result unreliable. Exports record the settings.
See Checking your StepChange results for guidance on each of these.
Your data
The analysis runs entirely in your web browser. Your file is not uploaded to our servers.
Statement version 1.0, November 2026.