Normality Tests in Six Sigma: How to React When Data Isn’t Normal

In Six Sigma, data drives every decision. You use it to find root causes, validate improvements, and prove results. But before running any statistical test, one critical question often goes unnoticed: Is your data normal? Normality isn’t just a textbook concept. It determines whether your statistical conclusions are valid or not.

Many Six Sigma tools assume data follow a normal distribution—the familiar bell curve. When that assumption breaks, your p-values, control charts, and capability indices can all become misleading.

This guide explains what normality means, how to test for it, and what to do when your data isn’t normal. You’ll learn about key normality tests, their interpretation, and practical workarounds when your data refuses to behave.

Why Normality Matters in Six Sigma

Most Six Sigma analyses rest on statistical assumptions. One of the biggest is that your data follows a normal distribution.

A normal distribution is symmetric. Most values cluster around the mean, and the probability of extreme values drops evenly on both sides. This pattern appears naturally in many processes due to random variation. However, not all data behaves this way.

When the data deviates from normal, several problems arise:

Problem	Impact
Biased capability indices (Cp, Cpk)	Cp and Cpk assume normal data. Non-normal data distorts capability results.
Invalid hypothesis test results	Tests like t-test and ANOVA rely on normality. Non-normal data affects p-values and confidence intervals.
Unreliable control charts	Individual (I) and X̄ charts assume normality. Non-normal data increases false alarms or hides real issues.
Wrong conclusions	You may think a process is in control or capable when it isn’t.

Many Six Sigma practitioners forget to verify normality before running tests. That oversight can lead to costly errors in interpretation.

The Central Limit Theorem (CLT) helps sometimes—it states that the mean of many samples tends toward a normal distribution, even if individual data points don’t. However, the CLT only applies under certain conditions, such as large sample sizes and independent observations. You can’t assume it will always rescue your analysis.

Recognizing Non-Normal Data

Before running formal tests, visual inspection helps. Graphs often reveal the problem faster than numbers alone.

Common signs your data isn’t normal

Indicator	Description
Skewness	The distribution leans left (negative skew) or right (positive skew).
Outliers	Extreme values distort the average.
Heavy or light tails	More (or fewer) extreme values than expected under a normal curve.
Multiple peaks	The data has more than one mode, indicating a mixed process.
Asymmetry in plots	Histograms or box plots show uneven spread.

Useful graphical tools

• Histogram with normal curve overlay – Quickly shows if data follow the bell curve.
• Box plot – Displays outliers and skewness.
• Q-Q (quantile-quantile) plot – If points deviate from the straight line, data are non-normal.

Visuals should always come first. But to confirm, you need statistical tests.

Common Normality Tests

Several formal tests check whether data follow a normal distribution. Each has strengths, weaknesses, and preferred sample sizes.

Test	Ideal Sample Size	Focus	Strengths	Limitations
Shapiro–Wilk Test	n ≤ 2000	Detects skewness and kurtosis	Very powerful for small samples	Overly sensitive for large samples
Anderson–Darling Test	n ≤ 5000	Detects deviations in tails	Strong sensitivity in tails	Requires software; sensitive for large n
Kolmogorov–Smirnov (K-S) Test (with Lilliefors correction)	n > 50	Compares empirical vs. theoretical distribution	Works for large data sets	Less sensitive to tail differences
Jarque–Bera Test	n > 200	Based on skewness and kurtosis	Easy to calculate; interpretable	Less sensitive for small samples
D’Agostino–Pearson Test	n > 50	Combines skewness and kurtosis measures	Balanced power	Not ideal for tiny samples

Each test starts with the null hypothesis (H₀): data come from a normal distribution.
If the p-value < 0.05, reject H₀ — your data are likely not normal.

How to Interpret Normality Test Results

The key is context. A significant result (p < 0.05) means data deviate statistically from normality. But in Six Sigma, you also need to decide whether the deviation is practically important.

For small samples (n < 30):
Normality tests have low power. They may not detect real deviations. Combine test results with visual inspection.

For large samples (n > 500):
Tests are too sensitive. Even tiny, harmless deviations produce small p-values. Don’t reject normality automatically—check plots and skewness values.

Rules of Thumb for Skewness and Kurtosis

When checking for normality in Six Sigma, two quick numerical indicators can help you judge whether your data behaves normally — skewness and kurtosis. They summarize how symmetric and how peaked your data is compared to a normal distribution.

Understanding these values allows you to catch non-normal data early, even before running formal normality tests like Anderson-Darling or Shapiro-Wilk.

What Skewness Tells You

Skewness measures how asymmetric your data is around the mean.

• A skewness of 0 means the data is perfectly symmetric.
• Positive skewness means the right tail is longer — more high values stretch the distribution to the right.
• Negative skewness means the left tail is longer — more low values pull the distribution to the left.

Skewness Value	Shape of Distribution	Interpretation
0	Perfectly symmetric	Normal
0 to ±0.5	Slight skew	Approximately normal
±0.5 to ±1.0	Moderate skew	Possibly non-normal
> ±1.0	Strong skew	Definitely non-normal

Example:
A Six Sigma team measures response times at a call center and finds a skewness of 1.25. That means the data is right-skewed — a few unusually long calls pull the average higher.

What Kurtosis Tells You

Kurtosis measures how peaked or flat the data distribution is compared to normal. A normal distribution has a kurtosis of 3 (this is sometimes adjusted to “0” when using excess kurtosis).

• Kurtosis > 3: The distribution has heavy tails and a sharp peak — known as leptokurtic.
• Kurtosis < 3: The distribution is flatter and more spread out — called platykurtic.

Kurtosis Value	Distribution Type	Interpretation
≈ 3	Mesokurtic	Normal
> 3	Leptokurtic	Heavy tails, more outliers
< 3	Platykurtic	Flatter, fewer outliers

Example:
A process measuring fill volumes shows a kurtosis of 4.6. That means extreme high or low fill levels occur more often than expected — a sign of heavy-tailed data.

What to Do When Data Isn’t Normal

Failing a normality test doesn’t mean your project is doomed. It just means you must adjust your approach. Several methods can help, depending on the situation.

1. Transform the Data

Transformations reshape the data to make them more symmetric and stabilize variance.

Transformation	When to Use	Formula / Example
Log transformation	Right-skewed data (e.g., time, cost)	Y' = ln(Y)
Square-root transformation	Count data (e.g., defects per unit)	Y' = √Y
Reciprocal transformation	Strong right skew	Y' = 1/Y
Box–Cox transformation	Unknown skew direction	Finds optimal λ to best normalize data
Johnson transformation	Complex or extreme non-normality	Uses fitted curves (SL, SU, SB)

After transformation, retest for normality. If p-value > 0.05, proceed with normal-based tools (e.g., t-tests, Cp/Cpk).

Example:
Cycle times often show right skew. Applying a log or Box-Cox transformation usually straightens the tail. You can then run capability analysis on transformed data and convert results back to original units for reporting.

2. Use Non-Parametric Methods

If transformations fail, switch to methods that don’t assume normality. These tests work on ranks instead of raw data.

Purpose	Parametric Test	Non-Parametric Alternative
Compare two samples	t-test	Mann–Whitney U test
Compare multiple samples	ANOVA	Kruskal–Wallis test
Paired comparison	Paired t-test	Wilcoxon signed-rank test
Correlation	Pearson	Spearman rank correlation

Non-parametric methods are robust and simple. They may be slightly less powerful but provide reliable conclusions when data violate assumptions.

3. Apply Robust Statistics

Robust methods minimize the influence of outliers and non-normality.

Example	Description
Median instead of mean	Reduces effect of extreme values.
Trimmed mean (e.g., 10%)	Ignores smallest and largest data points.
Robust regression (M-estimator)	Fits models that resist outlier impact.

These methods keep your conclusions stable, even when data behave unpredictably.

4. Use Bootstrap or Resampling

Bootstrap resampling builds a distribution by repeatedly sampling (with replacement) from your data.
It lets you estimate confidence intervals and p-values without assuming normality.

Steps:

1. Draw thousands of random samples from your dataset.
2. Compute your statistic (mean, median, difference).
3. Use the distribution of those statistics to find confidence limits.

Bootstrapping is common in Minitab, R, and Python. It’s powerful when you can’t transform or classify data easily.

5. Collect More Data or Subgroup

Sometimes non-normality stems from small sample size or mixed sources.
Increasing your sample or separating data by conditions often helps.

Example:

• Combining data from two shifts with different setups may create a bimodal curve.
• Splitting data by shift often restores normality within each subgroup.

Always check for special causes before assuming the data are inherently non-normal.

Example 1: Manufacturing Dimensions

Imagine a Six Sigma team measuring shaft diameters from a new supplier.
The specification is 10.00 ± 0.50 mm, and they collect 25 samples.

Step 1 – Visual Check

A histogram shows a slight right skew. The box plot has a couple of high outliers.

Step 2 – Run Normality Test

Shapiro–Wilk p-value = 0.02 (< 0.05).
Conclusion: Data are not normal.

Step 3 – Try Transformation

Log transformation improves symmetry. New p-value = 0.16.
Now, the data pass normality, so they proceed with the t-test to compare against the target.

Step 4 – Report Results

The team clearly notes the transformation step in their project documentation to maintain transparency.

Example 2: Service Process Cycle Times

A financial process measures the time to approve loan applications.
Times range from 1 minute to 120 minutes, with most approvals under 10 minutes.

Step 1 – Visual Inspection

Histogram shows a long right tail — many fast approvals, few very slow ones.

Step 2 – Anderson–Darling Test

p-value = 0.000 (< 0.05).
Data are strongly non-normal.

Step 3 – Box–Cox Transformation

Software suggests λ = 0.20.
After transformation, p-value = 0.08.
The team proceeds with capability analysis.

Step 4 – Interpretation

Before transformation, Cpk = 0.89 (appeared poor).
After correcting for normality, true Cpk = 1.34 — process meets requirements.

Lesson: Testing normality avoids false negatives in process capability.

Handling Non-Normal Data in Control Charts

Control charts also assume normal data, but some charts tolerate non-normality better.

Chart Type	Normality Sensitivity	Notes
X̄–R / X̄–S Charts	Low	Sample means tend to normality (thanks to CLT).
Individuals (I) Chart	High	Requires near-normal data. Consider transformations.
p, np, c, u Charts	N/A	Designed for count or attribute data; assume binomial or Poisson instead.
Non-normal capability analysis	Moderate	Use Weibull or lognormal models if process follows those patterns.

If your process data stay non-normal, you can use non-normal control charts or percentile-based limits instead of traditional ±3σ limits.

Capability Analysis for Non-Normal Data

When data fail normality, normal-based Cp and Cpk become unreliable.
You can either transform the data or fit a non-normal distribution model.

Common alternatives

Distribution	Typical Application	Notes
Weibull	Reliability, life data, failure times	Fits right-skewed data.
Lognormal	Cycle times, waiting times	Common in service and manufacturing.
Gamma	Skewed continuous data	Flexible for many positive-only datasets.
Beta	Percentages or proportions (0–1)	Great for yield or defect rates.

Most statistical software (like Minitab or JMP) automatically fits these distributions and provides non-normal capability indices such as Pp, Ppk, or Cnpk.

Practical Rules of Thumb

These quick rules help decide whether to proceed or adjust:

Situation	Recommended Action
n < 30	Use visual + Shapiro–Wilk. Consider non-parametric tests.
30 ≤ n ≤ 300	Use Shapiro–Wilk or Anderson–Darling.
n > 500	Focus on shape and skewness, not just p-value.
Outliers present	Investigate causes; consider robust methods.

Normality Test Workflow for Six Sigma Projects

You can use this structured approach whenever you analyze process data.

1. Collect raw data
2. Plot histogram and Q-Q plot
3. Check for outliers
4. Compute skewness and kurtosis
5. Run normality test (Shapiro–Wilk or Anderson–Darling)
6. If p > 0.05 → proceed
7. If p < 0.05 → transform or use non-parametric approach
8. Re-test after transformation
9. Document everything (test used, p-value, decision)

Keeping this workflow consistent helps teams maintain credibility and reproducibility.

Example: Summary Table of Normality Decision Path

Step	Action	Example Result	Next Step
1	Plot histogram	Right skew observed	Run normality test
2	Shapiro–Wilk test	p = 0.01 (non-normal)	Try Box–Cox transformation
3	Re-test after transform	p = 0.12	Proceed with t-test
4	Report	Include both raw and transformed findings	Document in A3 or Control Plan

Key Takeaways

• Always test for normality before running parametric analyses.
• Combine graphical and statistical methods—neither alone tells the full story.
• Don’t panic if data are non-normal. You have many options: transform, use non-parametric tests, or model non-normal distributions.
• Document your approach. Transparency builds trust in your Six Sigma results.

Normality testing might seem like a small step, but it protects the integrity of your project. It ensures every improvement decision rests on solid statistical ground.

Conclusion

Six Sigma depends on data accuracy. But data rarely behave perfectly. Real processes have skew, outliers, and noise.
Understanding how to detect and handle non-normality separates good practitioners from great ones.

Normality tests help you ask, “Can I trust my data?”
If the answer is no, you now know exactly what to do next.