PDCA: the Check is the part everybody skips

PDCA — Plan, Do, Check, Act — is a cycle for making a change and finding out whether it worked. Deming called it a cycle for learning, and the distinction matters: the output of a PDCA is knowledge, not a completed action.

Three of the four steps are done routinely everywhere. Check is skipped almost universally — not deliberately, but because by the time the change is in, everyone has moved on, and because nobody decided in advance what checking would consist of.

Plan

State the change, the prediction, and how you will know.

The prediction is the load-bearing part, and it is what distinguishes PDCA from simply doing things. Not "this should help" — "missed appointments will fall from around 37 a week to under 30, and we will see it within six weeks". A prediction you write down can be wrong, which is the only way the cycle produces learning.

Then decide what you will measure, before the change, and write a data collection plan:

The last row is the one people regret omitting. Recording the clinic alongside the count costs nothing at the time and is impossible to reconstruct afterwards.

And a rule that must be written down: how will you decide? "We will conclude it worked if the run chart shows a shift — six or more consecutive weeks below the median." Decided in advance, this is a test. Decided afterwards, it is a description of whatever happened.

Do

Run the change small and reversible. A trial on one clinic for eight weeks tells you almost everything a full roll-out would, at a fraction of the cost and with the option of stopping.

And here the discipline from data collection applies: collect the data as the plan says, without changing how the work is done. Standing over someone with a clipboard changes what they do — everyone works to the standard while being watched. Use data the process already produces where you can, and where you cannot, tell people what you are recording and why. Observation that people know about is far less distorting than observation they suspect.

Check

Compare what happened with what you predicted. Not with what you hoped, and not with last month.

This is where the run chart earns its place. Two points cannot separate signal from noise; neither can three. The rules — six consecutive points one side of the median, too few runs — are precisely the tools for asking whether the change did anything, and they need enough data to work.

"It did not work" is a successful PDCA. You now know something that cost you eight weeks and one clinic instead of a permanent half-post. The cycle that produces no learning is the one where nobody checked.

Act

Three honest options: adopt it, adapt it and run another cycle, or abandon it.

Abandoning is the option organisations find hardest, because a change that has been announced acquires defenders. This is exactly why the prediction and the decision rule are written down in Plan — they were agreed before anyone was invested in the answer.

The run chart opens with all 24 weeks.

1. Confirm: median 36, twelve runs, no signals. The tool states plainly that no signal is not the same as "the process is fine".
2. Now delete weeks 16 to 24, leaving only the three weeks the clinic looked at plus the pre-change period. Fifteen points, and still no shift — the run of consecutive weeks below the median is three, not six. Even the clinic's own truncated view did not support the conclusion it drew.
3. Paste a series where the change did work: keep weeks 1–12 and replace weeks 13–24 with 24, 22, 25, 21, 23, 20, 24, 22, 26, 21, 23, 22. The shift rule fires immediately, and the runs test with it. This is what a working change looks like, and it does not need interpretation.
4. Compare how the two conclusions feel. The real one is not subtle. If you find yourself arguing about whether a change worked, that is usually the answer.