Cp, Cpk, Pp, Ppk — which one do I report?
Four indices, two questions.
Cp vs Cpk — Cp asks "is the spread narrow enough to fit the spec?", Cpk asks "and is it also in the right place?". A perfectly precise process aimed at the wrong value has a high Cp and a low Cpk. Cp is the potential, Cpk is the reality.
Cp/Cpk vs Pp/Ppk — this is the pair people confuse. Cp and Cpk use the short-term spread (variation within a subgroup, the noise you cannot remove). Pp and Ppk use the total spread of everything you measured, which also contains the drift between subgroups. So if Cpk is 1.8 but Ppk is 1.1, the process is precise minute to minute but wanders across the day — look for setup differences, tool wear or a temperature cycle. Chasing more precision would be the wrong fix.
The assumption nobody checks: all four are built on plus/minus 3 sigma, which only means "99.73% of parts" if the data is roughly normal. Life data, surface roughness, contamination counts and anything bounded at zero are usually skewed. On a right-skewed process with an upper spec limit the normal formula underestimates how far the tail reaches and reports a better Ppk than reality — the direction that gets bad parts shipped. SenSight runs an Anderson-Darling test first and, when the data is not normal, gives you the percentile-based figures (ISO 22514-2) computed from the distribution that actually fits.
Cpm is the odd one out: it penalises being off target even when you are still inside the spec.
Cp vs Cpk — Cp asks "is the spread narrow enough to fit the spec?", Cpk asks "and is it also in the right place?". A perfectly precise process aimed at the wrong value has a high Cp and a low Cpk. Cp is the potential, Cpk is the reality.
Cp/Cpk vs Pp/Ppk — this is the pair people confuse. Cp and Cpk use the short-term spread (variation within a subgroup, the noise you cannot remove). Pp and Ppk use the total spread of everything you measured, which also contains the drift between subgroups. So if Cpk is 1.8 but Ppk is 1.1, the process is precise minute to minute but wanders across the day — look for setup differences, tool wear or a temperature cycle. Chasing more precision would be the wrong fix.
The assumption nobody checks: all four are built on plus/minus 3 sigma, which only means "99.73% of parts" if the data is roughly normal. Life data, surface roughness, contamination counts and anything bounded at zero are usually skewed. On a right-skewed process with an upper spec limit the normal formula underestimates how far the tail reaches and reports a better Ppk than reality — the direction that gets bad parts shipped. SenSight runs an Anderson-Darling test first and, when the data is not normal, gives you the percentile-based figures (ISO 22514-2) computed from the distribution that actually fits.
Cpm is the odd one out: it penalises being off target even when you are still inside the spec.
Try it in the app
Try: SPC tab, enter LSL/USL, press Capability. Then load a skewed column (a life or roughness measurement) and watch the normal and percentile figures separate.