Process Capability Analysis: Step-by-Step

How do I run a capability analysis whose result would survive an IATF or customer audit?

1. Confirm the process is in statistical control (control charts).
2. Verify measurement system capability (MSA / Gage R&R).
3. Check normality of the distribution.
4. Calculate Cp, Cpk, Pp, Ppk.
5. Predict PPM and sigma level.
6. Compare against the customer or internal target.

Skipping the stability and MSA checks invalidates the indices. A high Cpk on an unstable process is meaningless. A high Cpk on a process with a poor gage is just measurement noise.

1. Define the characteristic, USL/LSL/target and one-sided vs two-sided type.
2. Define the sampling plan: subgroup size, frequency, and total window.
3. Run a Gage R&R that covers the actual operating range of the characteristic.
4. Collect data without adjusting the process during the study window.
5. Plot control charts (X̄-R, X̄-s, or I-MR) and confirm stability.
6. Test normality (Anderson-Darling or probability plot).
7. Compute Cp, Cpk (within-subgroup σ) and Pp, Ppk (overall σ).
8. Compute predicted PPM at nearest tail; disclose the convention.
9. Compare against the required target and record deviation with reaction plan.
10. Archive raw data, MSA, control charts, and study metadata alongside the index.

• Subgroups spanning shifts, operators, or setups — inflated within-subgroup σ.
• Data collected during process tuning; the study captures the transient, not the steady state.
• MSA done at a single value rather than across the tolerance.
• Applying normal-based PPM to a clearly skewed distribution (roundness, flatness, torque).

• The largest source of bad Cpk numbers in industry is not the process — it is bad subgrouping.
• A capability study where the operator adjusts the machine mid-run measures the operator, not the process.
• The gap between Cpk and Ppk quantifies between-subgroup variation. A near-zero gap suggests the study window is too short to detect long-term variation.
• Normality tests fail more often on large samples — a p-value alone should not disqualify a study; inspect the histogram and physical mechanism.

• Processes that have never been stabilized — the number is not predictive.
• Characteristics without a valid tolerance or with a symbolic tolerance only.
• Destructive tests where subgrouping is impossible without redefining the sampling plan.

• Analysis output feeds Cp vs Cpk, Cpk vs Ppk, and PPM/sigma conversions.
• Analysis inputs come from control charting (stability) and MSA (measurement adequacy).

1. Prove stability first, then calculate capability.
2. Run MSA before interpreting the output.
3. Split mixed streams before pooling the data.
4. Use the free Process Capability Calculator to compute Cp, Cpk, Pp, Ppk, sigma level, and predicted PPM from your measurement data — no sign-up required.

• Store the sampling plan and σ method with the result — the number alone is not evidence.
• Never "improve" Cpk by removing outliers without a documented physical cause.

• ✓Stability proven on the chart appropriate to the data
• ✓MSA covers the tolerance range and passes the acceptance rule
• ✓Rational subgrouping documented
• ✓Normality checked or non-normal method disclosed
• ✓Cp/Cpk and Pp/Ppk both reported with σ method
• ✓PPM convention (nearest-tail vs two-sided) declared
• ✓Result compared to the governing requirement

• Process condition: statistical stability is required.
• Distribution assumption: use a distribution model justified for the data.
• Confirm process stability and measurement-system adequacy before interpreting a capability index.
• Use a justified distribution model or non-normal method when the normal model is unsuitable.