1. Define the routine process and comparable records
Describe the method version, materials, instruments, operators, environment, execution frequency, and output being monitored. Decide whether changes in lot, site, instrument, matrix, or method require annotation, stratification, bridging, or a new process population.
Preserve all run identities and statuses. Combining unlike products or method versions can manufacture variation that a control chart then mislabels as process behavior.
2. Choose a measure and chart that match the data
Select the measure from the process question: one result per run, a meaningful subgroup mean and spread, a count, a proportion, or a statistic designed for smaller sustained changes. Define units, transformation, time order, missing data, repeats, and uncertainty.
An individuals chart, mean/range or mean/SD chart, attribute chart, CUSUM, or EWMA answers a different statistical question. Choose during process design, not after viewing which chart produces the preferred signal.
| Data structure | Decision before charting |
|---|---|
| One independent value per run | Dispersion estimator and assumptions for an individuals chart |
| Meaningful within-run subgroup | Subgroup definition and separate center/spread chart |
| Count or proportion | Denominator stability and suitable attribute chart |
| Small sustained change | CUSUM or EWMA parameters and response threshold |
3. Establish the baseline and control limits
For an illustrative individuals chart, eight ordered control-recovery results are 99, 101, 100, 102, 98, 100, 101, and 99%. Their mean is 100.0%. The seven moving ranges average 2.0 percentage points; using d₂ = 1.128 gives estimated σ = 2.0 ÷ 1.128 = 1.773.
The illustrative three-sigma limits are therefore 100.0 ± 3 × 1.773, or LCL 94.68% and UCL 105.32%. Retain the included records, estimator, constants, calculated values, version, and approval. This individuals-chart calculation is not interchangeable with a subgroup, attribute, CUSUM, or EWMA chart.
Individuals chart: σ̂ = MR̄ ÷ 1.128; limits = mean ± 3σ̂| Run | Recovery | Moving range |
|---|---|---|
| 1 | 99% | — |
| 2 | 101% | 2 |
| 3 | 100% | 1 |
| 4 | 102% | 2 |
| 5 | 98% | 4 |
| 6 | 100% | 2 |
| 7 | 101% | 1 |
| 8 | 99% | 2 |
| Baseline | Mean 100.0% | Average 2.0 |
4. Detect a predefined change signal
Apply the approved chart and rule versions to ordered data. If Run 9 is 107.0%, it exceeds the worked UCL of 105.32% and triggers the predefined one-point-beyond-three-sigma signal.
Retain Run 9, its exact value, baseline and rule versions, active filters, calculation time, and signal. The flag suggests that the process may have changed; it is not a diagnosis or automatic product decision.
5. Investigate common and assignable causes
Compare material lots, instruments, analysts, cell banks, timing, preparation, plate effects, maintenance, software and method changes, deviations, and neighboring runs. Use process knowledge and evidence to distinguish a plausible assignable cause from baseline common-cause variation.
Keep the original data and signal intact. Record the scope, evidence, causal conclusion, affected results, and any reason an assignable cause could not be established.
6. Control the process and verify the response
Define notification, triage, disposition, correction, corrective or preventive action, approval, and follow-up before routine monitoring. Verify whether the action restored the intended process without rewriting the earlier baseline.
Periodically assess false-signal burden, detection delay, investigation value, and process changes. Version any approved change to the chart, rule, baseline, or response procedure.
Limits and statistical review
Autocorrelation, unequal uncertainty, censoring, recalculation, non-normality, small samples, and changing reference materials can defeat a simple chart. Obtain statistical review before using an automated signal for a quality decision.
This page explains SPC method selection. The control-chart guide covers implementation records; the Levey–Jennings and Westgard pages cover two narrower charting conventions.
Frequently asked questions
Can SPC replace per-run system suitability?
No. Run-level suitability evaluates the current execution; SPC monitors behavior across a defined sequence. They answer related but distinct questions.
Should failed or invalid runs be removed from charts?
Define handling by chart purpose. Excluding them can hide process behavior, while including incomparable invalid results can distort limits. Preserve them and document whether they contribute to each calculation.