Two parts off the same machine are never identical. The question is not whether a process varies — it always does — but which kind of variation you are looking at, because the two kinds demand opposite responses and confusing them is how a competent person makes a process worse.
The distinction
Common cause variation is the process being itself. It is the sum of many small influences, none dominant, all present all the time: slight differences in material, ambient temperature moving through the day, the tool wearing by a micron, the operator's hand. It is inherent, it is predictable in aggregate, and it is what defines the process's natural width.
A process with only common cause variation is in statistical control. That does not mean it is good. It means it is predictable: you can say what tomorrow's output will look like, and be right.
Special cause variation is something happening to the process that is not normally part of it. A tool chipped. The wrong material lot got loaded. A relief operator set the machine differently. The coolant ran low. It is not inherent, it is not predictable, and it is usually traceable to a specific event at a specific time.
A process with special causes present is out of control — not "broken", literally not predictable, because there is more than one process running and you do not know which one produced the next part.
Why confusing them is expensive
This is Deming's point, and it is the practical core of SPC.
Reacting to common cause as though it were special cause — adjusting the machine because the last part came in a little high — is called tampering, and it is worse than doing nothing. Provably worse. The adjustment corrects for a deviation that was going to regress on its own, so the correction becomes a deviation in the other direction, which gets corrected, and the process oscillates with roughly twice the variation it started with.
The operator is behaving conscientiously. The output gets worse. This is not a training problem; it is what happens when a chart is not there to say "that point is normal, leave it alone."
Ignoring special cause as though it were common cause is the opposite failure and it is quieter. A signal appears, someone says "it varies, that's normal", and the assignable cause stays in the process. It will recur, because nobody removed it.
The two errors have a cost structure that makes the control chart worth having: one makes you act when you should not, the other makes you not act when you should, and no amount of experience reliably tells them apart by eye.
What the control chart is actually for
A control chart is not a graph of your data. It is an operational rule for deciding which kind of variation you are looking at, with a known and deliberately chosen error rate.
The limits are set at three sigma of the process's own short-term variation. Inside them, variation is consistent with common cause and the correct action is to leave the process alone. Outside them, or in one of the run patterns, the variation is unlikely enough to be worth investigating as a special cause.
Three sigma is not a statistical derivation — it is Shewhart's economic choice. Tighter limits catch more real signals but generate false alarms that send people chasing nothing and destroy trust in the chart. Wider limits miss real problems. Three sigma is the balance that has held up in practice for a century, which is why it is a convention rather than a calculation.
The run rules exist because a point can stay inside the limits and still be telling you something: nine in a row on one side of the centre line is a shift, six in a row rising is a trend. Both are patterns that common cause variation produces very rarely.
The critical mistake: limits that move
Almost every spreadsheet control chart has this bug, and it is silent.
If your limits are formulas over the same column the chart plots — AVERAGE(B2:B200),
STDEV(B2:B200) — then adding a row recalculates them. A slow drift drags the
limits along with it. The chart stays reassuringly in control while the process
walks away from where it started, and it never signals.
That is not a control chart. A control chart compares what the process is doing now against what it did when it was behaving. So the limits must be frozen to a baseline window — a period when the process was demonstrably stable — and left there. New data is judged against the baseline, not against itself.
Recompute the limits only when you have deliberately changed the process and want a new baseline. That is a decision someone makes, not a side effect of pasting in this week's readings.
Stability comes before capability
Cpk and Ppk are predictions: this is what the process will produce if it carries on behaving as it did. An out-of-control process has no single behaviour to carry on with, so the prediction describes nothing real.
Establish control first. Then measure capability. In that order, always — and the gap between Cpk and Ppk is one useful measure of how much instability was present while you were measuring.
The practical sequence
- Chart it — with limits frozen to a stable baseline, not recalculated.
- Signal? Find the cause. A signal is a question about a specific moment. What changed at 14:20? Log what you found, in the chart, against the point.
- No signal? Do not adjust. However tempting the last reading was.
- Stable but not good enough? That is a common cause problem, and it needs a change to the process itself — different material, tighter fixture, better machine. Not an adjustment. Not an operator conversation.
The chart's most valuable output is often the permission it gives to do nothing.