p Chart: Hold the Defect Rate Steady After the Project Team Leaves

The p Chart monitors the proportion of defective units in variable-size subgroups to detect instability in defect rate. That is the whole job. You count how many units you inspected, you count how many of them were defective, and the chart turns that pair into a proportion. Then it tests whether that proportion is behaving the way a stable process behaves.

The variable-size part matters. Most real counting does not happen in neat, equal batches. One day you inspect 120 units. The next day you inspect 135. A chart built for fixed samples cannot cope with that honestly. The p Chart can, because it recalculates the control limits for each point based on how many units sat behind it. A small sample gives wider limits. A large sample gives tighter limits. The chart adjusts.

Use it in the Control phase of RDMAICS. By that point you have already found the cause, made the change and proved the change worked. What remains is the hardest part. The project team disbands. Attention moves elsewhere. The improvement quietly erodes and nobody notices for six months. A p Chart is the instrument that notices. It gives the process owner a running answer to one question: is the defect rate still where we left it?

The chart is also useful earlier, in Measure, to establish what the defect rate looked like before you touched anything. The same tool serves both purposes. Only the baseline changes.

The tool ships with sample data, and this guide is built entirely around it. The case is fictional. It describes a retail group tracking daily incomplete orders across three stores. An order counts as defective if it goes out missing an item.

There are six records. Store 1 contributes two: 7 defective out of 120 units inspected, then 8 out of 135. Store 2 contributes two: 6 out of 110, then 9 out of 140. Store 3 contributes two: 18 out of 125, then 7 out of 130.

Look at those numbers before any maths happens. Five of the six pairs sit in a similar range. One does not. Eighteen defective out of 125 is visibly different from the rest. The whole value of the chart is that it does not rely on you spotting that by eye. It applies a consistent test, every time, to every point, including the ones that look unremarkable.

The series field carries the label Daily incomplete orders. The group field carries the store name. Those two fields let you keep several streams of data in one place without mixing them up when you read the chart.

The tool asks for four things. Two are required and two are optional. Getting them right takes a few minutes and saves a lot of argument later.

Units inspected is the denominator. This is how many items you actually looked at in that subgroup, not how many existed. If 300 orders went out and you checked 120, enter 120. Defective units is the numerator. This is how many of those inspected units failed. A unit is either defective or it is not. If one order is missing three items, it still counts as one defective unit. If you want to count the three items, you want a different chart.

Group or stratum lets you tag where the data came from. In the sample data this is the store. It does not change the arithmetic. It changes what you can see when you go looking for a pattern. Series or scenario separates distinct measurement streams so you can hold more than one in the same tool.

Define defective in writing before anyone collects a number. This is the step teams skip. If one store counts a substitution as incomplete and another does not, your chart is measuring definitions, not performance.

The tool reports six points and a centre of 0.0724. That centre is the pooled proportion across all six subgroups, not the simple average of the six individual proportions. Pooling is the correct method here, because subgroups with more units inspected carry more information and should pull the centre further.

One point sits out of control. The signals table names it: point 5, value 0.144, triggered by rule 1, one point beyond 3 sigma. The explanation offered is that a single reading this far out is almost never chance. Point 5 is the Store 3 record with 18 defective units out of 125 inspected. That proportion is roughly double the centre line.

The chart applies all eight Nelson rules to the main chart. Rule 1 catches single extreme points. The other seven catch patterns that develop over time, such as runs on one side of the centre or steady drifts. With only six points, most of the pattern rules cannot fire yet. They need more history. That is worth remembering when you start a new chart and see no signals. No signal is not the same as proof of stability.

The tool also notes that limits were calculated from all six points, and that you can set a baseline to judge later data against an earlier period. There is an AI review available that explains the result in plain language if you want a second read.

The first decision is whether to act at all. This is the decision that costs organisations the most money when it goes wrong. A point inside the limits is ordinary variation. Chasing it means changing a process that was behaving normally, which usually makes things worse. Point 5 is different. The chart says this one is worth a conversation. That conversation starts with a question about what was different that day at that store, not with a blanket directive to all three stores.

The second decision is about the baseline. Once your improvement is embedded and you have enough stable points, freeze the limits on that period. Every later point then gets judged against the improved process rather than against a rolling average that quietly absorbs drift. Without a frozen baseline, a slow slide back to the old defect rate can go unflagged, because the limits slide with it.

The third decision is about handover. The Control phase of RDMAICS ends with a process owner accepting responsibility. A p Chart gives that person something concrete to own. Who updates it, how often, and what they do when a signal appears. Write that down. A chart nobody updates is a record of the past, not a control.

Mixing streams into one chart is the most frequent error. The sample data holds three stores in one view. That is fine for a first look, but if the stores genuinely run at different defect rates, combining them inflates the variation and hides real signals. If Store 3 is structurally different, chart it separately.

Changing the definition of defective part way through is the second error. Any change in what counts as a failure will show up as a shift in the chart, and you will spend weeks investigating a process that did not change at all.

The third error is treating every signal as a failure of people. A signal is information about a process on a particular day. Point 5 tells you something unusual happened. It does not tell you who, and it does not tell you anyone was careless.

The fourth error is building limits from too little data. Six points is enough to produce a chart. It is not enough to be confident about where the centre truly sits. Treat early limits as provisional and recalculate once you have a longer stable run.

It will not tell you why. The chart flagged point 5 at 0.144 and named rule 1. It said nothing about staffing, stock levels, a new picker or a system outage. Cause comes from going and looking. The chart only tells you where to look and when.

It will not fix a definition problem. If defective means different things to different people, the chart will faithfully plot the confusion. Rubbish data produces a confident-looking chart that misleads.

It will not work for every kind of count. If you need to count how many faults appear in each unit rather than how many units failed, a p Chart is the wrong instrument. It also struggles when defect rates are extremely low or sample sizes are very small, because the limits become too wide to detect anything useful.

It will not tell you whether 0.0724 is a good defect rate. That is a business judgement about customers and cost. The chart tells you whether the rate is stable. Stable and unacceptable are entirely compatible states, and a process can sit in both for years.

Finally, it will not maintain itself. The p Chart holds the gain only while someone keeps feeding it data and acting on what it says. That habit, not the chart, is the real control.