You ran a normal capability study on your process data and got a Cpk that makes no sense. Maybe it says the process is capable when you know parts are failing inspection, or it flags a problem that operators can’t find on the floor. That mismatch usually means your data isn’t normally distributed, and calculating process capability for non normal data the same way you would for a bell curve gives you numbers you can’t trust.
This guide shows you how to fix that. You’ll learn how to test your data for normality, identify which distribution actually fits (Weibull, lognormal, or another), and calculate the right capability indices, typically Ppk based on percentiles rather than the standard sigma-based formulas built for symmetric data.
We’ll walk through the process step by step: checking distribution fit with tools like Anderson-Darling, selecting an appropriate transformation or distribution model, and interpreting results so they reflect what’s really happening on your line. If you’re an engineer or quality manager tired of capability numbers that don’t match reality, this is the hands-on method our Master Black Belts use in the field, not just theory from a textbook.
Why normal capability formulas fail on skewed data
Standard Cpk and Ppk formulas were built on one assumption: your data forms a symmetric bell curve around the mean. The math divides the distance to your spec limits by three standard deviations, then assumes a fixed percentage of parts fall outside each tail based on the normal distribution’s known shape. That assumption works fine for a process like machined shaft diameters, but it breaks down the moment your data skews, which is common with cycle times, particle counts, flatness measurements, or anything bounded at zero.
The tails lie to you
Skewed data doesn’t split evenly around the mean. A lognormal distribution, for example, has a long tail stretching toward high values and a hard limit near zero. If you plug that data into a normal Cpk formula, you’ll underestimate the real number of parts crowding one spec limit while overstating your margin on the other. The index comes out looking healthy, but your scrap reports tell a different story.
A Cpk built on the wrong distribution doesn’t just miss the mark, it points you in the wrong direction entirely.
Where the mismatch shows up
| Data behavior | Normal formula result | What’s actually happening |
|---|---|---|
| Right-skewed (lognormal) | Overstates capability near upper spec | More parts near upper limit than predicted |
| Bounded at zero | Understates true defect rate | Impossible negative values get assumed |
| Bimodal shift | Averages hide two separate populations | Two processes masquerading as one |
This is exactly why percentile-based methods exist. Instead of assuming a shape and calculating from the mean and standard deviation, you fit the actual distribution and read the real percentiles where your spec limits fall. It’s more work upfront, but it’s the only way your capability number matches what’s rolling off the line.
Step 1. Confirm your process is stable
Before you worry about distributions, check whether your process is even stable enough to analyze. Capability indices, normal or not, only mean something if the process is running under consistent conditions. Calculating capability on an unstable process gives you a number that describes a moment in time, not a repeatable system, and any distribution fit you do next will be built on shifting ground.
Run a control chart first
Pull your data into an I-MR or Xbar-R chart before you touch capability software. Look for the classic red flags: points outside control limits, runs of seven or more points on one side of the centerline, or obvious trends and shifts. If your control chart shows any of these, stop. Fix the assignable cause first.
A capability number calculated on an unstable process is a guess dressed up as data.
What stability actually confirms
Stability doesn’t mean your data is normal. It means the variation you’re seeing is consistent and predictable, which is the only condition under which a distribution fit, normal or not, will hold up. Skip this step and you risk fitting a Weibull or lognormal curve to noise from a tool that’s wearing out mid-shift, a mistake that no amount of statistical sophistication downstream can correct.
Step 2. Test the data for normality
Once your control chart confirms stability, run a formal normality test before you calculate anything. Minitab, JMP, and most statistical software give you the Anderson-Darling test by default, and it’s the right choice for capability work because it weighs the tails more heavily than a simple Shapiro-Wilk test. Since your spec limits usually sit out in those tails, you need a test that catches deviations there, not just in the bulk of the data.
Reading the p-value
Set your threshold at 0.05. If the Anderson-Darling p-value comes back below that, reject the normality assumption and move to distribution fitting. If it’s above 0.05, you can proceed with standard Cpk formulas, but don’t stop there. Also plot a histogram and a normal probability plot side by side with the test output.
A p-value alone can’t tell you what shape your data actually takes, only that it isn’t a bell curve.
Don’t skip the visual check
Visual inspection catches things the p-value misses, especially with smaller sample sizes where the test lacks power to flag real skew. Watch for:
- A histogram with a long tail on one side
- Points curving away from the straight line on a normal probability plot
- A hard boundary at zero with data piling up near it
Any of these signals mean it’s time to identify a better-fitting distribution.
Step 3. Identify the best-fitting distribution
Once you’ve ruled out normality, run an individual distribution identification report in your stats software. Minitab’s tool tests your data against eight or more candidate distributions at once and ranks them by fit. Skip the guesswork and let the software do the heavy lifting here.
Reading the fit statistics
Compare the Anderson-Darling statistic across candidates, lower values mean a tighter fit. Pair that number with the p-value for each distribution and check the probability plot visually. A distribution can win on the numbers but still show curved points at the tails, which matters most since that’s where your spec limits sit.
The distribution with the lowest Anderson-Darling statistic and a clean probability plot wins, not the one that’s easiest to explain.
Common distributions in manufacturing data
| Data type | Likely distribution | Typical source |
|---|---|---|
| Cycle times, repair times | Lognormal | Bounded at zero, long right tail |
| Failure rates, particle size | Weibull | Wear-out or fracture mechanics |
| Flatness, surface finish | Weibull or Gamma | Physical boundary at zero |
| Two shift patterns combined | Bimodal, not a single distribution | Two machines or operators pooled together |
If nothing fits well, check whether you’re actually looking at mixed populations instead of one skewed distribution. That’s a data collection problem, not a math one.
Step 4. Calculate and interpret non-normal Ppk
Once you’ve identified the right distribution, calculate Ppk using percentiles instead of standard deviations. Minitab’s non-normal capability analysis does this automatically once you select the fitted distribution, but understanding the logic matters. Instead of measuring three standard deviations from the mean, the software finds the 0.135 and 99.865 percentiles of your fitted curve, the same tail coverage a normal distribution would give you, then measures the distance from those percentiles to your spec limits.
Reading the percentile-based output
Your report will show Ppk split into two one-sided numbers, PPL and PPU, based on the lower and upper spec limits. Take the smaller of the two as your reported Ppk, exactly like you would with normal data. The software also reports observed and expected parts-per-million defective, and for skewed data these two numbers often diverge sharply from what a normal Cpk would have predicted.
Non-normal Ppk isn’t a different formula so much as a different question: how much of your real distribution actually falls outside spec.
What the number is actually telling you
A Ppk above 1.33 still signals a capable process, but now it reflects the true shape of your data. Compare the expected PPM to your actual scrap and rework counts. If they’re close, your distribution fit is solid and you can trust the number for decision-making.
Turning capability data into real process improvement
Getting process capability for non normal data right isn’t an academic exercise. A Ppk built on the correct distribution tells you the truth about defect rates, and that truth drives real decisions: where to focus improvement projects, whether a supplier’s process can actually hold your tolerance, and what to tell a customer asking for capability proof. Get the distribution wrong, and every decision downstream inherits that error.
None of this replaces the judgment that comes from doing it dozens of times across different processes. Software can rank distributions and calculate percentiles, but knowing when a fit is good enough, when you’re actually looking at mixed populations, or when the real fix is upstream of the data entirely takes hands-on experience. That’s the gap between running the analysis and trusting it enough to act on it.
If your team needs that experience, talk to our Lean Six Sigma experts about training your engineers or auditing your current capability studies.