Statistics

Common cause and special cause variation

In French: Causes communes et causes spéciales

All variation belongs to one of two families: the one the system produces permanently, and the one an identifiable event caused. Getting the family wrong leads to acting exactly at the wrong time.

Whose variation it is

The first family is the doing of the system itself: the way it was designed, equipped, supplied and organised. Nobody at the station is responsible for it, and nobody at the station can reduce it — it requires a decision that bears on the system, hence a management decision.

The second is the doing of a dated event, outside the usual operation: a different batch, a broken tool, a replacement, an intervention. It is treated where it occurs, by those present, and as fast as possible.

That split is not only a classification: it assigns a responsibility. Asking a team to correct a variation that belongs to the system amounts to handing it a problem it has no means of solving, and then concluding that it lacks rigour.

Tampering, or the cost of getting the family wrong

Correcting the setting at every deviation, when the process is merely varying normally, increases the spread instead of reducing it. Each correction adds to the next natural variation, and the result is a process more unstable than if nothing had been touched.

The phenomenon is easily demonstrated: it is enough to aim at a target several times, correcting after each attempt for the deviation observed. The cloud obtained is wider than that of the shooter who never corrects. It is counter-intuitive, and that is what makes the habit so stubborn — each correction is justified taken in isolation.

The symmetrical error also exists and costs just as much: letting a real change run by taking it for the usual noise. A slow drift, an offset after an intervention, a different batch then go unnoticed until the result becomes indisputable, that is, long afterwards.

How it is decided

It is not decided on an isolated value but on a series of measurements in time order. A single value says nothing: the question is not "is this deviation large" but "is this behaviour the one this process usually has".

That is what makes a period of observation indispensable before any conclusion. A process that has never been tracked provides no benchmark, and the first deviation always looks abnormal there — hence the reflex to adjust.

Once the family is identified, the action follows without hesitation: identifiable event, look for what happened that day; usual behaviour, look for nothing in particular and work on the process as a whole. They are two different investigations, and mixing the two is the best way to succeed at neither.

The common mistakes

Opening a cause search at every deviation. On a variation that belongs to the system, the investigation will always find something — something always happens — and the explanation found will be wrong while looking convincing.

Asking the team for an action plan in the face of a variation that comes from the system. That produces local actions with no effect, then a legitimate discouragement, then a refusal to take part in the next projects.

Concluding on an isolated value, with no history. The question never bears on a point but on the behaviour of a series of points.

On the shop floor — Plastic injection moulding

An instruction required recentring the machine as soon as a part departed from the target. Tracking over three weeks with no adjustment at all gave a smaller spread than with the instruction applied. The adjustment was not bad: it was correcting a variation that had no cause to correct.

The other tools illustrated by this trade: plastics and chemicals.

Frequently asked questions

How do you know whether a variation is common or special?

By observing the series of measurements over time rather than an isolated value: a stable behaviour, with no outlier or trend, indicates a variation that belongs to the system. The decision therefore assumes a history — without a prior period of observation, no method makes it possible to decide, and the first deviation that comes along will always look abnormal.

Why does correcting at every deviation make the situation worse?

Because the correction adds to the next natural variation instead of cancelling it. The process already starts from a place offset by the previous adjustment, and the next deviation adds to it. The result is a wider spread than by doing nothing, even though each correction, taken in isolation, looked justified.

Who should act on each type of variation?

A variation linked to an identifiable event is treated as close as possible, by those who were present and know what happened. A variation that belongs to the system requires a decision on the system itself — equipment, supply, design, organisation — and is therefore a matter for management. Confusing the two makes a team carry a problem it cannot solve.

Is a process with no special cause a good process?

No, and it is a frequent confusion: it is only predictable. A process can behave in a perfectly regular way while producing a high proportion of scrap, if its usual variation exceeds what the customer accepts. Stability allows forecasting; it is the comparison with the expected limits that says whether the result is suitable.