Fixing the Non-Normality Warning in Minitab Capability Analysis (Cp/Cpk)

You ran Stat > Quality Tools > Capability Analysis > Normal in Minitab, and instead of a clean Cp/Cpk report you got a red warning icon with the message "Data do not appear to follow a normal distribution." The useful takeaway: this warning is triggered by the Anderson-Darling normality test returning a p-value below 0.05, and it does not automatically mean your process is non-normal. Small samples, outliers, and mean shifts all trip the same flag. Work through the checks below before reaching for transformations or nonparametric methods.

Version assumption: the menu paths and dialog options described here reflect Minitab 19/20/21 desktop. Earlier releases use the same Stat > Quality Tools path but dialog wording may differ slightly.

Why the warning appears

Minitab's normal capability analysis assumes the underlying data follow a normal distribution. To protect that assumption, it runs the Anderson-Darling test on your data and flags the result when p < 0.05. The Anderson-Darling test is a goodness-of-fit test that is especially sensitive to deviations in the tails of the distribution. Three distinct problems produce the same red icon, and each needs a different fix.

Cause and diagnostic table

Ordered checks

1. Check subgroup count first. If you have fewer than about 30 subgroups (or fewer than ~100 individual measurements for subgroup size 1), the test result is not trustworthy. Collect more data before changing anything else.
2. Plot the data. Run Graph > Individual Value Plot and Graph > Boxplot. Look for isolated points (outliers) and for a sudden level change (a shift in the process mean). Either one can fail the normality test even when the underlying process is stable and normal.
3. Run the normality test directly. Use Stat > Basic Statistics > Normality Test and select Anderson-Darling. Confirm the p-value. If you found and legitimately addressed outliers in step 2, re-test on the cleaned data.
4. Try a transformation. If p is still < 0.05 and the data are right-skewed (typical for times, counts, and measurements bounded at zero), use Calc > Calculator to create a log-transformed column, or run Stat > Control Charts > Box-Cox Transformation to find the optimal lambda. Re-run the normality test on the transformed column.
5. Decide the path. If the transformed data pass (p ≥ 0.05), re-run capability analysis on the transformed data. If no reasonable transformation passes, switch to a nonnormal distribution model or a nonparametric approach rather than forcing the normal method.

Fixes tied to each finding

Finding: not enough data

Do not transform, do not delete points. Collect enough subgroups and re-run. This is the most common cause of a spurious warning, and it is the cheapest to fix.

Finding: outliers or a shift

Investigate the assignable cause first. Only remove or correct a point when you have a documented reason (data entry error, known special cause). Then re-run the analysis. Never delete points just to make the p-value pass — that hides the very signal capability analysis is meant to surface.

Finding: true non-normality, transformation works

Re-run the capability analysis on the transformed column. Two cautions apply:

• Capability indices computed on transformed data describe the transformed scale. Document the transformation used (e.g., natural log, or Box-Cox with lambda = 0.3) in your report.
• Specification limits must be transformed consistently, or results must be reported as back-transformed values, so the indices remain interpretable against the original engineering tolerance.

Finding: transformation fails

In the Capability Analysis dialog, choose a nonnormal distribution that fits (Minitab offers Weibull, lognormal, exponential, and others via Stat > Quality Tools > Capability Analysis > Nonnormal), or use a percentile-based (nonparametric) capability method. Use Stat > Quality Tools > Individual Distribution Identification to compare candidate distributions by their Anderson-Darling p-values and pick the best fit.

Verifying the behavior yourself

You can confirm each branch of this guide with simulated data — no production data required:

1. Generate normal data: Calc > Random Data > Normal, 50 rows, mean 0, standard deviation 1. Run normal capability analysis — no warning should appear and Cp/Cpk should be reported.
2. Generate exponential data: Calc > Random Data > Exponential, 50 rows, scale 1. Run the same analysis — you should see the red warning and an Anderson-Darling p-value below 0.05.
3. Log-transform the exponential column (Calc > Calculator, function LN), re-run the normality test and capability analysis on the transformed column, and confirm the warning clears.

This exercise takes a few minutes and gives you a known-answer reference for what each condition looks like in your Minitab version.

When to escalate

Escalate to a quality engineer or statistician when all of the following hold: you have at least 30 subgroups, no outliers or shifts explain the result, and no common transformation (log, Box-Cox) brings the Anderson-Darling p-value to 0.05 or above. At that point the question stops being a software issue and becomes a modeling decision: formal distribution fitting, a nonparametric/percentile capability method, or a tolerance-interval approach instead of Cp/Cpk. Reporting a normal-theory Cpk on data that fail every normality check produces a number that looks precise but does not describe your process.