Two kinds of variation, and how a control chart tells them apart

Amara Start with the thing people get wrong on day one.

Nadia That variation is a fault. It is not. Every process varies, and the only useful question is which of two kinds you are looking at, because the two have completely different owners.

Amara Name them.

Nadia Common cause is the process being itself. Ambient temperature, material lot to material lot, the ordinary play in a mechanism. Always present, predictable within a range. Special cause is something new: a worn tool, a failed heater, a different operator, a delivery from a different supplier.

Amara Why does the distinction matter so much?

Nadia Because it decides who acts. Special cause belongs to the person at the machine, now, and it can be found. Common cause belongs to management, because reducing it means changing the process itself. Asking an operator to fix common cause variation is asking them to do something the job does not permit.

Amara And the chart tells them apart.

Nadia That is its entire purpose. Results in time order, a centre line at the average, limits at three standard deviations calculated from the process while it was behaving.

Amara Why three?

Nadia It is a compromise, not a law. Tighter and the chart cries wolf and people hunt causes that are not there. Wider and real signals slip past. Three has held for a century because it balances those two costs about as well as anything does.

Amara Most people only know one rule.

Nadia A point outside a limit, and on its own that catches large shifts only. A process can drift a very long way with every point still inside the lines.

Amara So what else are you looking for?

Nadia Runs and trends. Seven consecutive points on one side of the centre line means the average has moved. Seven rising or falling means something is drifting progressively, and that is what tool wear looks like before it produces a single reject.

Amara That is the useful one.

Nadia It is the whole prize. A trend caught at point five is a tool change scheduled at the end of the shift. The same trend caught when a point finally breaches the limit is a quarantined batch and a conversation with a customer.

Amara You also watch for points that are too neat.

Nadia Data hugging the centre line unnaturally closely. Real processes are not that tidy. It usually means somebody is averaging, rounding, or writing down what they expect to see rather than what the gauge said.

Amara Control limits and specification limits. Make the distinction hard to forget.

Nadia Control limits come from the process and describe what it does. Specification limits come from the customer and describe what is acceptable. They are unrelated quantities that happen to be drawn in the same units.

Amara What goes wrong when they are confused?

Nadia A process can be beautifully in control and outside specification on every unit. Stably producing rubbish. It can also be well inside specification and completely out of control, which means it is getting away with it today and you have no idea what it does tomorrow.

Amara And on the chart itself?

Nadia If the tolerance is drawn on a control chart, that chart is training people to tamper. Every time a point wanders towards a line it was never meant to be compared with, somebody reaches for an adjustment.

Amara Define tampering.

Nadia Adjusting in response to common cause variation. The part reads slightly high, the operator corrects, the next reads slightly low, they correct again. The output becomes measurably more variable than if nobody had touched the machine at all.

Amara That is counterintuitive enough that people will resist it.

Nadia The arithmetic is simple. The natural swing is still there, and now your correction is added on top of it. Two sources of variation where there was one. It is not a failure of skill, it is what happens when you respond to noise.

Amara How do you stop it without telling people to care less?

Nadia You never tell them to care less. You give them a rule that says when a reading is a signal and when it is not. That is precisely what the chart is for, and it is why an unread chart on a wall is worse than no chart, because it looks like control.

Amara Sampling. How much is enough?

Nadia Take consecutive units so the sample reflects one moment of the process, and space the samples through the shift so the gaps between them reveal drift. A sample assembled from units picked at random across twelve hours blends the two and can hide a shift that happened at lunchtime.

Amara And frequency?

Nadia Follow the rate of change. A process that drifts with tool wear needs sampling often enough to see the drift while correcting it still costs a tool change rather than a batch.