Control charts

Every process varies. The same task takes a little longer one day than the next, a few more forms come back one week than the last. A control chart draws that variation over time and puts three lines across it: a centre line at the average, and an upper and a lower control limit, usually three standard deviations either side, calculated from the data.

Points that wander between the limits without a pattern are common cause variation: the noise of a stable process. A point outside the limits, or a pattern that chance would rarely produce, is a signal of special cause variation: something has changed, and it is worth finding out what.

The chart prevents two expensive mistakes. The first is reacting to noise: asking why this week was worse than last when both weeks came from the same process. That produces explanations, meetings and changes that make no difference, or make things worse. The second is missing a real change because it is hidden among the ups and downs.

A stable process is not the same as a good one. A process can be perfectly stable and still too slow for the people it serves. Stability tells you the process is predictable; whether it is good enough is a separate question, answered by comparing it with what is needed.

Look first for points outside the control limits. Then look for patterns inside them. The rules most teams use are these. One point beyond either limit. Eight points in a row on the same side of the centre line (some organisations use seven or nine; agree one and keep to it). Six points in a row all rising or all falling. Two out of three points beyond two standard deviations on the same side.

Each rule is a pattern that would be very unlikely if nothing had changed. Each one you add catches more real changes and raises more false alarms, so use a small agreed set rather than every rule in the book.

Never draw the specification or the target as a control limit. Control limits come from what the process does; a target is what someone wants it to do. They are different lines answering different questions, and confusing them is the most common misreading of all.

The type depends on what you measure and how the data arrives. Measurements on a scale, such as minutes, days or pounds, use one family of charts; counts of failures use another. Find your row in the table.

The np chart is the p chart for when every sample is the same size; it plots the count of failures rather than the share, which is easier to explain to a team. The u chart is the c chart for when the amount of opportunity changes, for example errors per hundred records when the number audited varies; it plots the rate.

The X-bar S chart replaces the range with the standard deviation once groups reach about ten, where the range starts to waste information. EWMA and CUSUM charts weight or accumulate recent points so that a small sustained shift shows sooner than it would on a standard chart; they are worth their extra explanation only where small shifts are costly.

With fewer than about twenty points, the limits are not yet reliable. Start with a run chart, which needs fewer points and no calculation beyond a median, and move to a control chart when the data allows.

Do not use one on data that has been averaged or totalled over a period that hides the variation you care about, and do not mix different processes on one chart. A chart of two wards, two teams or two products together shows the difference between them as noise.

Every control chart, whatever its type, puts its limits the same way. The types differ only in how they estimate the spread.

Control charts belong to Measure, to show how the process behaves before anything changes, and to Control, to hold the gain. Statistical process control is the wider discipline of using them to run a process day to day.