A process capability study is a structured check that confirms a measurement system is trustworthy, proves the process is stable on a control chart, verifies the data is roughly normal, collects enough subgrouped data, and then computes capability indices so you can say whether the process can hold the specification. The math is the last step, not the study.

Most bad capability numbers come from skipping the setup work and jumping straight to a Cpk. If the gauge cannot repeat, or the process is unstable, or the data is skewed, the index that comes out is fiction dressed as a decision. This guide is the full procedure: the order of operations, how much data you need, how to read short-term against long-term, and what to actually do with the result. For what the indices mean once you have them, see Cp and Cpk explained.

What is a process capability study?

A process capability study compares what a stable process actually produces against what the specification demands, and expresses the answer as indices such as Cp, Cpk, Pp, and Ppk. It is not a single calculation; it is a sequence of gates. Each gate has to pass before the next one means anything: a trustworthy gauge, a process in statistical control, a distribution you understand, and a sample large enough to estimate spread and center with reasonable confidence.

The capability study pipelineFive gates before the number means anything1. GaugeMSA2. Stabilitycontrol chart3. Normalitycheck shape4. Collectsubgroups5. Compute + actCp, Cpk, PpkFail a gate, and every number downstream is unreliable. Fix it, then continue.
A capability study is a pipeline of gates. The indices at the end are only as honest as the gauge, stability, distribution, and sample behind them.

How do you run a process capability study, step by step?

The sequence below is the standard order. It is deliberately front-loaded: the cheap checks that can invalidate everything come first, so you never burn a full data collection on a process that was never going to yield a meaningful index.

  1. Validate the measurement system first. Run a gauge study (repeatability and reproducibility) before collecting a single capability data point. If measurement variation eats a large share of the tolerance, the study measures your gauge, not your process, and no downstream step can rescue it.
  2. Establish stability on a control chart. Plot the data on a control chart and confirm the process is in statistical control, with no points beyond the limits and no non-random patterns. Capability math assumes one predictable spread and center; an out-of-control process has neither.
  3. Remove special causes and re-check. If the chart shows special-cause signals, find and eliminate them, then confirm control again. Do not compute capability over a process you have not stabilized.
  4. Check the distribution shape. Test the data for normality with a histogram and a normality test. Skewed or bounded characteristics (flatness, runout, anything with a hard zero) break the ppm predictions and need a transformation (Box-Cox or Johnson) or a non-normal capability method.
  5. Collect enough subgrouped data under normal conditions. Gather at least 100 to 125 individual values across roughly 25 to 30 rational subgroups, spanning the routine sources of variation such as shifts, operators, and material lots. Fewer than about 20 subgroups gives an index with too much uncertainty to trust.
  6. Use rational subgroups. Form subgroups from consecutive parts made under the same conditions, so within-subgroup variation captures only short-term noise. Subgrouping by arbitrary time buckets can hide or exaggerate variation and corrupt the short-term sigma.
  7. Compute short-term and long-term indices. Calculate Cp and Cpk from within-subgroup (short-term) sigma, and Pp and Ppk from the overall (long-term) sigma of all the data. Report them together.
  8. Act on the result. Read the pattern: low Cp means the spread is too wide, so attack variation; a healthy Cp with a low Cpk means the process is off center, so re-center it; a wide gap between Cpk and Ppk means the process is not as stable as the control chart suggested.

How much data does a capability study need?

Enough that the estimate of spread is not itself mostly noise. The widely used guidance is at least 100 to 125 individual measurements, collected across 25 to 30 subgroups, with 20 subgroups treated as a practical floor before any index is calculated. Smaller samples do not just lower confidence; they systematically mislead, because a standard deviation estimated from 30 parts has a wide swing around the truth. Treat a Cpk from a handful of parts as a rough hint, never a verdict.

QuantityCommon guidanceWhy
Individual values100-125 minimumStable estimate of process spread
Subgroups25-30 (20 floor)Enough to judge stability and within-variation
Conditions spannedShifts, operators, lotsCapture the real sources of variation
PreconditionsGauge validated, process stableOtherwise the index is not meaningful
Sample-size rules of thumb for a defensible study. The numbers vary by customer standard, but the logic does not: estimate spread with enough data to trust it.

Why short-term and long-term sigma tell different stories

The single most useful output of a capability study is the comparison between the short-term and long-term indices. Cp and Cpk use within-subgroup variation, the process at its best behavior between disturbances. Pp and Ppk use the overall standard deviation of every data point, including drift, shift changes, and lot-to-lot wander. When the two agree, the process is stable and the between-subgroup noise is small. When Cpk sits well above Ppk, the process has more long-term wander than the control chart let on, and the short-term number is flattering.

Short-term versus long-term variationWithin-subgroup vs overall variationtightshort-term σ (within)→ Cp, Cpklong-term σ (overall, includes drift) → Pp, PpkA big Cpk-minus-Ppk gap is drift the short-term number never saw.
Short-term sigma is measured inside each subgroup; long-term sigma spans the whole run. The gap between the two indices is your instability, quantified.

What do you do with the result?

The indices point to different fixes, and the value of running the full study is knowing which problem you actually have before spending money:

For the full breakdown of how the four indices differ and interlock, see Cp, Cpk, Pp, and Ppk compared.

What are the common mistakes that wreck a study?

Most capability studies fail for the same handful of reasons, and every one of them traces back to skipping or rushing a gate:

Consider a concrete case. A supplier reports a Ppk of 0.9 on a bore diameter and panics about the tooling. A gauge study reveals the bore gauge alone accounts for a large slice of the observed variation. After the gauge is fixed and the study repeated, the true Ppk is 1.6. The process was fine all along; the measurement system was the problem. That is why the gauge is gate one, not an afterthought.

The standards behind the study

Capability studies are governed by industry standards, which is why customers can demand a specific index and expect a comparable method behind it.

Why capability studies quietly expire

The most common capability failure is not a flawed study; it is a good study nobody repeats. The indices get computed at launch, framed in a PPAP binder, and quoted for years while tooling wears and the mean wanders off center. A capability number is a snapshot, not a property of the process, and it goes stale the moment conditions change. When dimensional data and check weights are captured digitally at the station instead of on clipboards, a fresh study stops being an annual archaeology project: the subgroups for a new Cpk are already in the system, next to the downtime log and the scrap reasons. That continuous view is what Harmony's quality intelligence and paperwork digitization provides, turning capability from a certificate in a binder into a number you can trust today. See it on a real floor in our CLS case study.