A control chart is a time-ordered plot of process data with a center line and statistically derived upper and lower control limits, used to tell routine variation from real process changes. Which chart you use depends on two things: what kind of data you have (measurements or counts) and how you collect it.
Picking the wrong chart type is the most common technical mistake in statistical process control and it is entirely avoidable. The decision comes down to a handful of questions you can answer in a minute. This guide gives you the selection framework, the full chart-type table, and how to read the chart once it is running.
Which control chart should you use?
Answer these five questions in order and the chart picks itself.
- Is the data variable or attribute? Variable data is measured on a scale: weight, diameter, temperature, time. Attribute data is counted: defective units, number of flaws. Variable data carries far more information per sample, so prefer it when you can get it.
- For variable data: how do you sample? One measurement at a time (slow processes, batch parameters, destructive tests) points to an I-MR chart. Small subgroups of 2-9 consecutive pieces point to X̄-R. Subgroups of 10 or more point to X̄-S, where the standard deviation replaces the range because the range wastes information in large subgroups.
- For attribute data: are you counting defective units or defects? A unit is defective (pass/fail, one verdict per unit); a unit can carry multiple defects (scratches on a panel, errors on a form). Defective units point to p or np charts, which model yes/no outcomes; defect counts point to c or u charts, which model rates of occurrence.
- Is your sample size constant or varying? Constant sample size allows the simpler np (defectives) or c (defects) charts. Varying sample size requires p (proportion defective) or u (defects per unit), which adjust the limits for each sample's size.
- Sanity-check the subgroup logic. A subgroup should capture only short-term, within-stream variation: consecutive parts from one line, not a mix of four cavities or three shifts. Get this wrong (the classic "rational subgrouping" error) and the chart's limits will be too wide or too narrow no matter which type you picked.
The seven chart types at a glance
| Chart | Data type | Subgroup / sample | Typical use |
|---|---|---|---|
| X̄-R | Variable | 2-9 per subgroup | Machined dimensions, fill weights: small samples of consecutive parts at intervals |
| X̄-S | Variable | 10+ per subgroup | High-volume automated measurement where large subgroups are cheap |
| I-MR | Variable | 1 (individuals) | Batch parameters, destructive tests, slow processes: one value per time period |
| p | Attribute (defectives) | Varies | Proportion of defective units when lot or shift sizes change |
| np | Attribute (defectives) | Constant | Count of defective units in a fixed-size sample |
| c | Attribute (defects) | Constant inspection unit | Defects per fixed unit: flaws per panel, errors per form |
| u | Attribute (defects) | Varies | Defects per unit when the amount inspected changes |
Two practical notes. First, every variables chart is really two charts: one tracking location (X̄ or the individual value) and one tracking spread (R, S, or moving range). Read the spread chart first; if the process spread is unstable, the limits on the location chart cannot be trusted. Second, when in doubt with variable data, I-MR is the safe workhorse: it handles nearly any measurement collected one value at a time, at the cost of being slower to detect small shifts than a subgrouped chart.
How do you read a control chart?
A process is in control when its points look like random noise around the center line, inside the limits, with no patterns. Anything else is a signal. The classic pattern tests come from the Western Electric Statistical Quality Control Handbook (1956), which divided the chart into zones at one, two, and three standard deviations from the center line and defined four rules:
- Rule 1: a single point beyond 3 sigma (outside a control limit).
- Rule 2: two of three consecutive points beyond 2 sigma on the same side.
- Rule 3: four of five consecutive points beyond 1 sigma on the same side.
- Rule 4: eight consecutive points on one side of the center line.
Rules 2-4 exist because a modest process shift, say one sigma, can hide inside the 3-sigma limits for a long time; runs and clusters reveal it sooner. The trade-off is false alarms: each added rule raises the odds of flagging a stable process. A sensible floor default is Rule 1 plus Rule 4, adding the zone rules only on critical characteristics where catching small shifts early is worth chasing occasional ghosts.
Common chart-selection mistakes
Three errors account for most bad charts on real floors:
- Using an I-MR chart on averaged data. Plotting daily averages of dozens of parts on an individuals chart shrinks the visible variation and produces limits so tight that everything signals. Chart the raw values with the right subgrouped chart instead.
- Mixing streams in one subgroup. Sampling one part from each of four cavities into a single subgroup blends four different processes. Chart streams separately, or subgroup within a stream, so the limits reflect one process's short-term variation.
- Using attribute charts when measurements exist. Recording a fill weight as pass/fail throws away the number you already measured. A p chart needs hundreds of units per sample to see what an X̄-R chart sees in five, so keep the measurement whenever the gauge gives you one.
Where control charts come from, and why the limits sit at 3 sigma
Walter Shewhart proposed the control chart at Bell Telephone Laboratories in a 1924 memo and set the limits at three standard deviations as an economic choice, not a sacred number: wide enough that chasing false alarms stays rare, tight enough that real trouble gets flagged (ASQ, Walter A. Shewhart). The key facts worth keeping straight:
- The control chart dates to Shewhart's May 16, 1924 memo, and the full method to his 1931 book Economic Control of Quality of Manufactured Product (ASQ).
- The four zone-based pattern rules were codified in Western Electric's 1956 Statistical Quality Control Handbook still the reference for run rules (ASQ, Control Chart).
- Control limits are computed from process data; specification limits never belong on a control chart. Mixing them is the fastest way to turn a process-behavior tool into a pass/fail tally.
What do you do with a signal?
A signal is an instruction to investigate, not necessarily to adjust. The sequence that works: re-check the measurement, look for an obvious event (setup change, new material lot, tool change), annotate the chart with what you find, and escalate to containment if the cause stays hidden. That trail of annotations feeds root cause analysis and a signal tied to a confirmed cause that could recur deserves a CAPA rather than a quiet fix.
Once a chart shows sustained control, the next question is whether the stable process actually fits the spec, which is where Cp and Cpk take over. And if keeping paper charts current across a plant is the bottleneck, that is a capture problem, not a statistics problem: digitizing checks at the station, the way Harmony's paperwork digitization and live visibility modules do, turns every reading into chartable, searchable data the moment it is written, feeding the same quality trends your QMS reporting needs. The lines and rules on the chart, meanwhile, haven't changed since 1956, because they haven't needed to.