Statistical Process Control

A metric you watch over time — a processing time, an error rate, a daily count — is always wobbling. The hard question every time it moves: did something actually change, or is this just the normal jitter the process always has? Over-react to noise and you waste effort chasing ghosts (and often make things worse); ignore a real shift and you miss a genuine problem. Statistical process control (SPC) is the decades-old, beautifully practical discipline for telling these two apart — so you respond to real signals and leave the noise alone.

Born in manufacturing (Shewhart and Deming) but applicable to any repeating process, SPC is a close cousin of anomaly detection and streaming monitoring, with a sharp conceptual core. This page is that core: the variation distinction at its heart, the control chart that operationalises it, and the trade-off behind its famous limits.

Real change vs normal noise

The whole field rests on one reframing. Before SPC, people reacted to every up and down — a bad day's numbers triggered a scramble to "fix" something, even when nothing had actually changed. Shewhart's insight was that a stable process still varies, and reacting to that inherent variation as if it were a signal — tampering — usually makes the process worse, not better. The job of SPC is to draw a principled line: this much wobble is normal, leave it alone; that is a genuine signal, investigate it.

Common cause vs special cause

The foundational distinction — and arguably the single most useful idea in operational analysis:

Common-cause variation — the natural, inherent jitter of a stable process. Many small, ever-present influences (slight differences in conditions, materials, timing). It's predictable in its range, even if any single value isn't, and it's not worth chasing — it's just how the process is.
Special-cause variation — an unusual, assignable cause: something genuinely changed (a new supplier, a broken tool, a process change). This produces a value (or pattern) outside the normal range, and it is worth investigating — there's a real reason to find.

A process showing only common-cause variation is in control (stable and predictable); one with special-cause variation is out of control (something to act on). The entire machinery of SPC exists to separate these reliably — so you act only when there's genuinely something to act on.

The control chart

The tool that operationalises all this is the control chart: plot the metric over time, with three reference lines — a centre line at the process average, and upper and lower control limits (UCL/LCL) marking the boundary of normal variation.

A control chart. Points wobbling within the limits are common-cause variation — the process is in control, leave it alone. A point breaching a control limit is a special-cause signal — investigate. The limits encode 'how much wobble is normal' for this process.

As long as points bounce around within the limits with no pattern, the process is in control — that's just common-cause noise, and the correct action is none. When a point crosses a limit, that's a special-cause signal worth investigating. The chart turns a vague "that looks high" into a principled, repeatable decision rule.

Why three sigma?

The control limits are conventionally set at the centre line plus or minus three standard deviations of the process:

\text{UCL},\,\text{LCL} = \mu \pm 3\sigma

Why three? It's a deliberate cost trade-off, and recognising that is the key to using SPC well. For roughly bell-shaped data, only about 0.3% of points fall beyond ±3σ by chance — so a breach is very probably a real signal, not luck. Set the limits tighter (±2σ) and you'd catch real shifts sooner, but you'd also get far more false alarms — and reacting to those (tampering) wastes effort and destabilises the process.

Signals beyond a single breach

A point outside the limits is the obvious signal — but a process can drift in ways no single point catches. The Western Electric rules (and Nelson's) add patterns that also flag a special cause even when every point is inside the limits:

A run of several consecutive points all on the same side of the centre line (the process has shifted).
A steady trend of points marching up or down (gradual drift).
Too many points clustered far from the centre, or other non-random patterns.

The logic is the same throughout: a stable process should look random around the centre. Any non-random pattern — a run, a trend, a cycle — is a fingerprint of something systematic, i.e. a special cause, even before a point breaches a limit. SPC is as much about spotting structure as outliers.

Charts for small, slow shifts

The classic Shewhart chart uses only the current point, which makes it great at catching large sudden shifts but slow to notice a small persistent drift. For that, two charts use the history:

CUSUM (cumulative sum) — accumulates the running deviations from target, so even a small consistent bias adds up into a clear signal.
EWMA (exponentially weighted moving average) — a weighted average that emphasises recent points (the same smoothing idea as time series), sensitive to gradual moves.

Both are tuned to catch the slow drifts a Shewhart chart would miss — you pick the chart for the kind of change you're worried about. (There are also chart types for different data: X-bar/R for measurements, p-charts for proportions.)

Where it shows up in my work

Don't chase the noise

Any operational metric or KPI watched over time — a processing time, a volume, an error or compliance rate — invites the question SPC answers: is this a real change or just normal fluctuation? The single most valuable idea is the common-cause vs special-cause distinction: most month-to-month wobble is common-cause noise that shouldn't trigger action, and reacting to it (tampering) makes things worse — a discipline that saves a lot of wasted effort and over-correction.

A control chart turns "that number looks off" into a defensible rule, and the ±3σ choice is the same false-alarm trade-off (and precision/recall tension) that runs through anomaly detection — set it for the cost of crying wolf vs missing a real shift. It pairs naturally with streaming (real-time monitoring) and time series (the underlying signal), and is one of the most genuinely practical tools in the whole toolkit.

Refresh in 60 seconds

The common/special-cause framing, the ±3σ control chart, the Western Electric rules, and CUSUM/EWMA for small shifts reflect current SPC references alongside quality coursework.