Minitab Normal Capability Analysis Flags Non-Normal Data: Diagnostic Guide

The problem: normal indices on non-normal data

You run Stat > Quality Tools > Capability Analysis > Normal in Minitab and the output warns that the data are not normal. The Anderson-Darling p-value is below your chosen alpha, commonly 0.05, the histogram bars do not match the fitted normal curve, or the normal probability plot points curve away from the reference line. Reporting the normal Cp and Cpk anyway can misstate process capability because those indices assume a symmetric, unbounded distribution.

Useful takeaway: do not treat a normality warning as permission to delete data or force a normal model. Diagnose the shape first, then match the analysis method to the process behavior.

Recognizable condition

Look for one or more of these in Minitab output:

• Anderson-Darling normality p-value below the selected alpha.
• Capability histogram with a fitted normal curve that misses the data bars.
• Normal probability plot with points forming a clear curve instead of falling near the straight reference line.
• Process data bounded at zero or a physical limit, yet modeled with a normal distribution.

Cause and diagnostic table

Observed symptom	Likely cause	Primary check
Right-skewed or left-skewed histogram	Process naturally skewed or bounded	Nonnormal Capability Analysis with distribution fitting
Isolated points far from the main cluster	Outliers or measurement errors	Boxplot and time-order plot review
Many repeated values or few distinct readings across the spec range	Poor gage resolution	Count distinct values; gage repeatability and reproducibility review
Separate clusters by shift, machine, or operator	Multiple process sources pooled together	Boxplot by categorical factor
Drift, trends, or points outside control limits	Process not statistically stable	Control chart before capability analysis

Ordered checks

1. Run the normality test

   In Minitab, go to Stat > Basic Statistics > Normality Test. Select the measurement column and record the Anderson-Darling p-value. Compare it to your alpha, typically 0.05. A p-value below alpha means the normal model is rejected, but it does not by itself explain why.

2. Inspect the capability graphs

   Run Stat > Quality Tools > Capability Analysis > Normal. Enter the upper and lower specification limits. Review the histogram with the fitted curve and the normal probability plot. Look for curvature, outliers, and multimodal peaks.

3. Confirm stability first

   Create a control chart for the data in time order, such as Stat > Control Charts > Variables Charts for Subgroups > Xbar-R, or an individuals chart for single measurements. Capability indices are not meaningful for an unstable process. Establish statistical control before computing Cp or Cpk.

4. Check measurement resolution

   Count the number of distinct measurement values across the specification range. If the gage cannot distinguish enough levelsfor example, only a handful of values across a wide tolerancethe capability indices may be distorted regardless of distribution shape.

5. Test for stratification

   Make a boxplot by shift, machine, operator, or lot using Graph > Boxplot. If the groups show different centers or spreads, pooling them into one normal capability study can create an artificial non-normal shape.

Fixes tied to findings

Finding	Recommended action in Minitab	When it applies
Stable process, skewed but fits a known distribution	Use Stat > Quality Tools > Capability Analysis > Nonnormal. Choose a fitted distribution such as Weibull, lognormal, or exponential, then estimate capability from that distribution.	Process is stable and no assignable outliers exist.
Stable process, no standard distribution fits	Consider a Box-Cox or Johnson transformation in Minitab, then run normal capability analysis on the transformed data. Report that a transformation was used.	Customer or internal requirements allow transformed indices.
Assignable-cause outliers	Remove only invalid points with documented justification and quality approval. Do not delete points simply to improve Cp or Cpk.	Outlier is a measurement error or known one-time event.
Poor gage resolution	Improve the measurement system, such as a finer instrument or more decimal places, before re-running capability analysis.	Distinct value count is too low across the tolerance.
Multiple sources pooled	Analyze each stratum separately by shift, machine, or operator rather than combining all data into one normal capability study.	Boxplots show clear group differences.

Escalation criteria

Escalate to a quality engineer or statistician when:

• The process shows instability on a control chart.
• No distribution fits the data adequately in nonnormal capability analysis.
• Sample size is too small for reliable distribution fitting or capability estimation.
• A customer requires specific capability indices that a transformation would change or invalidate.

Practical verification and limitations

Exact menu names and available distributions vary by Minitab release. Verify the paths in your installed Help or release notes before following them; this guide assumes recent Minitab desktop releases, and web-based or older versions may differ.

To verify your setup, open Minitab, load a sample dataset, run Normal Capability Analysis, then run Stat > Basic Statistics > Normality Test on the same column and confirm that the Anderson-Darling p-value agrees with the capability diagnostic message. This is a configuration check, not evidence that any particular dataset passed or failed.

Limitations: Minitab's normality test can be overly sensitive with large samples, flagging minor deviations that may not matter practically. Small samples may fail to detect real non-normality. Use the probability plot and process knowledge alongside the p-value.

Rollback

Running capability analysis does not change process data, so no rollback is needed for the diagnostic steps. If you remove outliers or transform data, keep the original worksheet and document the change; revert to the original data if quality approval is not obtained.