SPC Guide
Which Control Chart Should I Use?
Two questions decide it: what kind of data do you have, and how is it grouped? Follow the tree.
The decision tree
Is the data measured (variables) or counted (attributes)?
│
├─ MEASURED (length, weight, temperature, time…)
│ │
│ ├─ Subgroups possible (n ≥ 2 per sampling)?
│ │ ├─ n = 2–8 → X̄-R chart
│ │ └─ n ≥ 9 → X̄-S chart
│ │
│ └─ Individual values only (n = 1)?
│ ├─ Default → I-MR chart
│ └─ Watching for slow drift → EWMA or CUSUM
│
└─ COUNTED
│
├─ Counting DEFECTIVE UNITS (pass/fail per unit)?
│ ├─ Sample size varies → p-chart
│ └─ Sample size fixed → np-chart
│
└─ Counting DEFECTS (flaws per unit/area)?
├─ Inspection unit fixed → c-chart
└─ Inspection unit varies → u-chart10 real scenarios, mapped
| Scenario | Right chart |
|---|---|
| Shaft diameter, 5 parts measured every hour | X̄-R chart → |
| Batch chemistry — one pH reading per batch | I-MR chart → |
| Daily % of boards failing final test (varying lot size) | p-chart → |
| Number of solder defects per board (fixed inspection unit) | c-chart → |
| Defects per square meter of fabric (varying area) | u-chart → |
| Detecting a slow 0.5σ drift in furnace temperature | EWMA chart → |
| Catching small sustained shifts in fill weight | CUSUM chart → |
| Subgroups of 12 wafers from an automated tester | X̄-S chart → |
| Destructive test — only 1 sample affordable per lot | I-MR chart → |
| Go/no-go gauge results, fixed sample of 50 per shift | np-chart → |
The mistake that invalidates everything
Treating measurement data as attribute data. If you can measure the dimension, don't chart it as pass/fail — you throw away most of the statistical power. A p-chart needs hundreds of samples to detect a shift an X̄-R chart catches in five subgroups. Measure when you can; count only when you must.
FAQ
What's the difference between defectives and defects?
A defective is a bad unit (a board that fails — count it once, use p/np charts). A defect is a flaw (one board can have five solder defects — count each, use c/u charts). Mixing these up is the most common chart-selection mistake.
When should I use X̄-S instead of X̄-R?
When subgroup size is 9 or larger. The range statistic loses efficiency for large subgroups; the standard deviation (S) uses all the information. Below n=9 they perform almost identically, and X̄-R is easier to compute by hand.
Why not just use EWMA or CUSUM for everything?
They're tuned for small sustained shifts (0.5–1.5σ) and react slower to large sudden shifts than a Shewhart chart with Western Electric rules. Standard practice: Shewhart charts as the default, EWMA/CUSUM when your failure mode is slow drift.
Not sure? Let the wizard decide
Answer 3 questions about your data and get the right chart, pre-configured.
Open the Chart Wizard →