Plan-Do-Study-Act and the Model for Improvement

PDSA is the improvement cycle from Walter Shewhart and W. Edwards Deming. Deming insisted on Study rather than the older "Check", because the point is to learn from the comparison of prediction to result, not merely to inspect.

The Model for Improvement (Associates in Process Improvement; popularized by the Institute for Healthcare Improvement) wraps PDSA in three questions asked first.

Evidence: Mixed. The underlying components are sound: statistical process control is well founded, and explicit prediction before observation is basic scientific method. The applied literature is poor — systematic reviews of published health care PDSA work find that most reports omit predictions, run a single cycle, use no time-series data, and do not meet the method's own criteria. The method is better than most of its practice.

Use when: improving a process incrementally with the people who run it, in short cycles.

Do not use when: the problem is strategic, structural, or requires a decision rather than an experiment.

The three questions

What are we trying to accomplish? A specific, measurable, time-bound aim. "Reduce door-to-antibiotic time for suspected sepsis from 90 to 45 minutes by December."
How will we know that a change is an improvement? The measures, agreed in advance — outcome, process, and balancing.
What changes can we make that will result in improvement? Change ideas, drawn from theory, from bright spots, or from the people doing the work.

The cycle

Plan: state the change, who does what, where, when — and predict the result. The prediction is mandatory; without it, Study degenerates into rationalization.
Do: run it at small scale. Record what actually happened, including the things that went wrong and the departures from plan.
Study: compare result to prediction. Where they differ, ask why. The learning is in the difference.
Act: adopt, adapt, or abandon. Then start the next cycle.

Scale of testing

The most common error is testing too big. Ramp deliberately:

One clinician, one patient, one shift.
One clinician, five patients.
Three clinicians, one week.
One ward, one month.
Spread.

Move up a level only after a successful cycle. "Test of change with one patient tomorrow" gets agreement that "pilot across the directorate" never will.

Three kinds of measure

Outcome: did the thing we care about improve?
Process: is the change actually happening? (Usually the first thing to check when the outcome does not move.)
Balancing: what got worse elsewhere? Non-negotiable. Every improvement in one place can export harm to another.

Run charts and special cause

Plot data over time, not before-and-after bars. Before/after comparison cannot distinguish improvement from normal variation, and it is how most improvement claims are inflated. Learn the basic run chart rules — shift, trend, runs, astronomical point — before claiming an effect. See measurement-and-metrics.

Examples by audience

Health care: PDSA is the lingua franca of clinical improvement. Its most common failure is ritual PDSA: the paperwork is completed, no prediction is recorded, the cycle is run once, and it is called a project. A real PDSA is fast, small, repeated, and often uncomfortable.
Software: the same logic as a hypothesis-driven experiment. "We believe reducing review batch size will cut lead time; we predict a 30% reduction in four weeks; if not, we will conclude the constraint is elsewhere." Attach the prediction to the ticket.
Executive: ask any improvement program for its predictions, recorded before the change. Programs that only report post-hoc successes are not learning; they are marketing.

Limitations

Weak for problems requiring redesign rather than iteration.
Published PDSA work in health care is frequently of poor quality — single cycles, no predictions, no measures.
Small-scale testing does not always predict at-scale behavior.