P-Charts for Validating Studies
Use a p-chart with defined subgroups, comparable denominators and suitable control limits.
Updated:
A p-chart monitors a proportion across time-ordered subgroups. It can reveal changes worth investigating. It does not certify unbiased sampling, correct categories or achievement of a productivity target.
Required data
Define a binary characteristic within a clear taxonomy, such as direct-work reading versus other states. Retain each subgroup's count xᵢ and size nᵢ; pᵢ=xᵢ/nᵢ.
The classical construction assumes independent trials with a common reference probability for the process being considered. People in one round may share states. Mixing products, shifts or causes may produce overdispersion. Examine these conditions before interpreting alerts and use rational subgrouping reflecting the intended comparison.
There is no universal minimum of 20 or 30 readings ensuring normal approximation. Both nᵢp̄ and nᵢ(1−p̄) matter. Rare events or low counts may need binomial limits or another method; declare the method used.
Center line and limits
With a suitable reference phase:
p̄ = Σxᵢ / Σnᵢ.
With unequal sizes this is generally not the unweighted mean of subgroup percentages. Classical three-sigma limits for each subgroup are:
UCLᵢ=min(1,p̄+3√(p̄(1−p̄)/nᵢ))
LCLᵢ=max(0,p̄−3√(p̄(1−p̄)/nᵢ)).
Limits change with nᵢ. Record the reference source, subgroup count and justified exclusions. Do not delete points merely because they are inconvenient.
For a normal variable with known parameters,±3standard deviations has theoretical coverage near 99.73%. A discrete binomial p-chart with estimated p does not guarantee that exact coverage per point, much less across a sequence. Additional runs rules increase false-alarm opportunities; prespecify the rules.
Simulated example with variable subgroup sizes
These are teaching data, not client results. The state is defined as direct work and the subgroup is a shift:
| Shift | nᵢ | xᵢ | pᵢ |
|---|---|---|---|
| 1 | 42 | 18 | 0.429 |
| 2 | 38 | 15 | 0.395 |
| 3 | 40 | 19 | 0.475 |
| 4 | 41 | 16 | 0.390 |
| 5 | 39 | 20 | 0.513 |
| 6 | 40 | 17 | 0.425 |
| 7 | 43 | 19 | 0.442 |
| 8 | 37 | 14 | 0.378 |
| 9 | 40 | 18 | 0.450 |
| 10 | 41 | 17 | 0.415 |
| 11 | 39 | 16 | 0.410 |
| 12 | 40 | 19 | 0.475 |
| 13 | 42 | 18 | 0.429 |
| 14 | 38 | 15 | 0.395 |
| 15 | 40 | 18 | 0.450 |
Σxᵢ=259 and Σnᵢ=600, giving p̄=259/600=0.4316667. For the first shift:
√(p̄(1−p̄)/42)≈0.076428; LCL≈0.20238; UCL≈0.66095.
Differences from0.203/0.661 result from early rounding. Calculate at full precision, round for presentation and repeat using each nᵢ.
No point in this table exceeds its classical limits. That means the rule detects no signal, not definitive proof of stability. Fifteen shifts may be insufficient to characterize a future reference, and coverage, independence and classification remain to be checked. The43.17% mean describes these600 readings, not every plant or period.
Responding to a signal
Check recording errors, product changes, shift composition, incidents and dictionary changes. An outlying point may reflect a special cause, an incorrect assumption or a false alarm; it does not automatically identify its cause.
Real variation does not invalidate the entire study. Describe the observed period, analyze segments or establish justified new references after a change. Do not project one mean uncritically as a constant process.
A process without detected signals can still perform unacceptably. Common-cause improvement involves changing the system; “in control” does not mean “needs no improvement.” A special cause may be unavoidable or favorable; alerts do not automatically call for punitive correction.
Confidence intervals, taxonomy and OEE
A confidence interval estimates a period/population proportion under a sampling design. Control limits describe expected variation between subgroups under a reference model. Sample-size planning does not require z=3 merely because a chart uses three-sigma limits.
Review MECE classification through definitions, cases and observer agreement; a stable chart does not validate the taxonomy. Complete OEE requires A=run time/planned production time, P=ideal cycle×total count/run time and Q=good count/total count. State charts alone do not provide all those inputs.
WorkSamp supports visit recording and study preparation. Do not assume automatic p-charts or SPC alerts. Cronometras supports durations, Induly working-time records and ASETEMYT specialist resources.