Real SPC vs Coloured Trend Bars: Reading the Risk Weather

Most safety dashboards show each indicator as a tile: red, amber or green, often with a small up or down arrow. It looks decisive. The trouble is what the colour actually means. A red/amber/green tile encodes exactly two facts, the number moved, and the latest value is above or below a fixed line someone chose when the dashboard was set up.

That is not a signal. It is a comparison against an arbitrary line. Three problems follow directly:

• It cannot see normal variation. Every indicator wobbles month to month. A go-around rate that bounces between 3 and 7 per thousand landings is not improving in the good months or failing in the bad ones, it is just varying. A coloured tile flips between green and amber on that wobble and manufactures alarm where there is none.
• The threshold is a guess. Someone picked the amber line. Set it too tight and every indicator screams; set it too loose and a genuine drift sits green for months. Either way the colour reflects the line someone drew, not the behaviour of the operation.
• It hides the shape of the change. A single bad month and six consecutive months creeping the same direction can show the same colour. One is noise; the other is the operation genuinely drifting. The tile cannot tell them apart, so neither can the person reading it.

The result is a board that someone has to eyeball and interpret. It tells you a number went up. It does not tell you whether the change is real. That is the gap real statistical process control closes.

Statistical process control (SPC) is a long-established method for telling ordinary variation apart from a real change in the thing being measured. It does not ask anyone to pick a line. It builds the line from the data.

For each Safety Performance Indicator (SPI), a quantifiable measure of an aspect of safety, such as unstable approaches per thousand sectors, SPC produces a control chart:

• A centre line: the indicator's typical level, computed from its own history. This is what “normal” looks like for this indicator, not a number borrowed from a template.
• Upper and lower control limits: the boundaries of normal variation, also computed from the indicator's own data. As long as the indicator stays inside these limits and shows no out-of-pattern run, the operation is behaving as it always has.

The point of the limits is that they are earned, not assumed. An indicator with a lot of natural month-to-month spread gets wider limits; a stable one gets tighter limits. The chart adapts to how each indicator actually behaves, which is exactly what a fixed colour threshold cannot do.

ICAO Doc 9859, the Safety Management Manual, points operators to this kind of analysis precisely because it removes the argument about where to draw the line. The data draws it. The only question left is the right one: has the indicator stayed inside its own normal envelope, or has it genuinely shifted?

A control chart on its own is a picture. What turns it into a signal is a fixed set of tests applied to every new month, the Western Electric Rules. They look for the patterns that mean “this is no longer the same process”, each one a different shape of real change rather than noise:

• A point past the limit. A single month falls outside the upper or lower control limit. The strongest single-month signal: a value this extreme is very unlikely to be ordinary variation.
• A run on one side. Several consecutive months all sit on the same side of the centre line. No single month is alarming, but a sustained run on one side is the operation drifting to a new level, the pattern a coloured tile most often misses.
• A steady trend. A run of months stepping the same direction, up or down, month after month. A consistent climb is a real movement even while every individual month still looks within range.
• Clustering near a limit. Several recent months bunch up close to a control limit. The indicator is creeping toward the edge of normal, an early warning before a point actually crosses it.

The discipline of the rules is the whole point. They are applied the same way every month, to every indicator, so an analyst does not have to squint at a chart and decide by feel whether something looks wrong. A single bad month that is just variation does not trip a run rule; a quiet, sustained drift that no tile would flag does. That is the difference between a number going up and an indicator having genuinely shifted, a real special-cause signal.