In the Analyse Phase of DMAIC, statistical tools help improvement teams move from plausible explanations to verified root causes. One of the most useful decisions in this phase is determining whether several process groups have comparable variation before comparing their averages.

That is where Bartlett’s Test becomes important.

Suppose you want to compare the average cycle time of three machines, the average patient waiting time across four clinics, or the average transaction-processing time across several teams. A one-way ANOVA may be the appropriate method for testing whether the group means differ. However, standard ANOVA relies on an important assumption: the groups should have approximately equal variances.

Bartlett’s Test assesses that assumption before you interpret the ANOVA result.

It is not a formality. If the variation differs substantially between groups, particularly when sample sizes are also unequal, standard ANOVA may produce an unreliable conclusion. A disciplined analyst checks the assumption, understands the result, and selects an appropriate alternative such as Welch’s ANOVA when required.

What Is Bartlett’s Test?

Bartlett’s Test is a statistical test for determining whether three or more independent groups have equal population variances.

The hypotheses are:

• Null hypothesis, H₀: All group variances are equal.
• Alternative hypothesis, H₁: At least one group variance is different.

Variance describes how widely individual observations spread around their group mean. Two departments may have the same average cycle time but very different consistency. One may complete every task close to the target, while the other produces a wide range of results.

Bartlett’s Test compares the separate sample variances with a pooled estimate of variance. Its test statistic approximately follows a chi-square distribution with _k − 1 degrees of freedom, where k is the number of groups.

The fundamental purpose is straightforward:

Before comparing group means with standard ANOVA, check whether the groups have sufficiently similar variation.

Bartlett’s Test is especially powerful when the data in each group are approximately normally distributed. It is also sensitive to non-normality. If the data are heavily skewed or contain influential outliers, Levene’s Test or the Brown–Forsythe method may provide a more robust assessment.

The National Institute of Standards and Technology explanation of Bartlett’s Test provides the formal definition, equations, and a worked engineering example.

Why Equal Variances Matter for ANOVA

ANOVA evaluates whether the differences between group means are large relative to the natural variation within the groups.

Conceptually:

[
F = \frac{\text{variation between group means}}{\text{variation within groups}}
]

Standard ANOVA estimates within-group variation by pooling information from all groups. This pooling is most defensible when the group variances are approximately equal.

For example, imagine the following average cycle times:

The means appear relatively similar, but the consistency of the machines is not. Machine A is highly stable, while Machine C produces much more variable results.

If the sample sizes were equal and the variance differences were modest, standard ANOVA could remain reasonably robust. However, large variance differences, unequal sample sizes, non-normal data, or a combination of these conditions increase the risk of an unreliable inference.

This is why Bartlett’s Test acts as a gatekeeper. It helps determine which version of the mean-comparison analysis is appropriate.

Worked Example: Calculating Bartlett’s Test Step by Step

Consider three process groups, each containing 10 observations. Assume their sample variances are:

• Group A: (s_1^2 = 1)
• Group B: (s_2^2 = 9)
• Group C: (s_3^2 = 25)

We will test the equality of variances at:

[
\alpha = 0.05
]

There are:

• (k = 3) groups
• (n_1 = n_2 = n_3 = 10)
• Total sample size (N = 30)
• Degrees of freedom for each group: (n_i – 1 = 9)

Step 1: Calculate the pooled variance

The pooled variance is a weighted average of the group variances:

[
s_p^2 =
\frac{\sum (n_i-1)s_i^2}{N-k}
]

Substitute the values:

[
s_p^2 =
\frac{(9)(1)+(9)(9)+(9)(25)}{27}
]

[
s_p^2 =
\frac{9+81+225}{27}

\frac{315}{27}

11.6667
]

The pooled variance is therefore approximately 11.6667.

Step 2: Apply the Bartlett statistic

A commonly used form of the Bartlett statistic is:

[
T =
\frac{
(N-k)\ln(s_p^2)

\sum (n_i-1)\ln(s_i^2)
}{
1+
\frac{1}{3(k-1)}
\left[
\sum\frac{1}{n_i-1}

\frac{1}{N-k}
\right]
}
]

First, calculate the logarithmic component:

[
(N-k)\ln(s_p^2)

27\ln(11.6667)
\approx 66.332
]

Next, calculate the sum of the group components:

[
\sum (n_i-1)\ln(s_i^2)

9\ln(1)+9\ln(9)+9\ln(25)
]

[
\approx 0+19.775+28.970

48.745
]

The uncorrected numerator is:

[
66.332-48.745=17.587
]

Now calculate the correction factor:

[
1+
\frac{1}{3(2)}
\left[
\frac{1}{9}+\frac{1}{9}+\frac{1}{9}

\frac{1}{27}
\right]
\approx 1.04938
]

Therefore:

[
T =
\frac{17.587}{1.04938}
\approx 16.76
]

Step 3: Compare the statistic with the critical value

Bartlett’s Test has:

[
df=k-1=2
]

At (\alpha=0.05), the chi-square critical value for 2 degrees of freedom is approximately:

[
\chi^2_{0.95,2}=5.991
]

Our result is:

[
T=16.76
]

Because:

[
16.76>5.991
]

we reject H₀.

The corresponding p-value is approximately 0.0002, which is well below 0.05.

Step 4: Interpret the result

There is statistically significant evidence that the three process groups do not have equal variances.

This does not identify which specific group differs. It tells you that the assumption of common variance is not supported. You should then investigate the process conditions, inspect box plots, review measurement-system information, and select a mean-comparison method that accommodates unequal variances.

What Happens When Variances Are Unequal?

When Bartlett’s Test indicates unequal variances, the preferred alternative for comparing several means is often Welch’s ANOVA.

Welch’s ANOVA differs from standard ANOVA in two important ways:

1. It gives groups with smaller variance greater influence in the calculation of the test statistic.
2. It adjusts the degrees of freedom to account for unequal variances and, where relevant, unequal sample sizes.

This makes Welch’s ANOVA more reliable when the equal-variance assumption is not reasonable.

The practical decision path is:

• Bartlett’s p-value > 0.05: Do not reject the equal-variance assumption. Standard ANOVA may be appropriate, provided normality, independence, and sampling conditions are also reasonable.
• Bartlett’s p-value < 0.05: Evidence indicates unequal variances. Consider Welch’s ANOVA.
• Strong non-normality or influential outliers: Consider Levene’s or Brown–Forsythe testing and assess whether a transformation, robust method, or Kruskal–Wallis procedure is more suitable.

Remember that a p-value above 0.05 does not prove that the variances are identical. It means the sample does not provide sufficient evidence to conclude that they differ. Always combine the test result with a visual review of the data and practical process knowledge.

The Minitab guide to tests for equal variances also highlights an important point: the effect of unequal variances depends on factors such as sample-size balance, the statistical model, and the underlying distribution.

Bartlett’s Test in the Analyse Phase of DMAIC

In DMAIC, the Analyse Phase identifies and verifies the root causes of process performance problems. Bartlett’s Test supports that objective by helping you determine whether variation changes across process conditions.

A practical workflow is:

1. Define the response variable
   Select a measurable output such as cycle time, defect percentage, resolution time, or delivery accuracy.

2. Identify the factor and groups
   Compare machines, shifts, suppliers, sites, materials, software versions, or operating conditions.

3. Check the measurement system
   An unreliable measurement process can create artificial differences in both means and variances.

4. Visualise the data
   Use box plots, individual value plots, histograms, and standard-deviation comparisons.

5. Assess distributional shape
   Bartlett’s Test is most suitable when the groups are approximately normal.

6. Run Bartlett’s Test
   Record the test statistic, degrees of freedom, and p-value before selecting the ANOVA method.

7. Select standard ANOVA or Welch’s ANOVA
   Match the analysis to the evidence rather than forcing every dataset into the same procedure.

8. Translate the result into process knowledge
   If one machine has much greater variation, investigate calibration, maintenance, setup conditions, operator methods, or incoming material.

9. Carry verified factors into Improve
   Use the findings to design controlled trials, optimise critical inputs, and reduce variation.

This sequence connects statistical analysis to the broader Lean Six Sigma principle of making decisions from evidence. A statistically significant result is not automatically a business solution, but it can identify where the team should focus its improvement effort.

Common Mistakes to Avoid

When using Bartlett’s Test and ANOVA, avoid these frequent errors:

• Treating a non-significant result as proof of equality.
• Ignoring normality, even though Bartlett’s Test is sensitive to non-normal data.
• Using standard ANOVA automatically when variances and sample sizes are substantially different.
• Confusing statistical significance with practical significance.
• Skipping visual analysis, which may reveal outliers or different distribution shapes.
• Interpreting unequal variance as a root cause without investigating the process mechanism.
• Testing repeatedly until a preferred result appears, rather than defining the analysis plan in advance.

Reliable Analyse work combines statistical tests with process observation, stratification, measurement-system knowledge, and subject-matter expertise.

Build Stronger Statistical Judgement

Bartlett’s Test is a small step in a larger analytical discipline. It helps you ask a critical question before interpreting ANOVA:

Are the groups comparable in their level of variation?

When the answer is reasonable, standard ANOVA can provide a powerful comparison of group means. When the answer is no, Welch’s ANOVA offers a more suitable path. In both cases, the improvement team makes a stronger decision because the analysis matches the data.

Professionals developing these capabilities benefit from structured Lean Six Sigma Green Belt training, where statistical tools are connected to project selection, root-cause analysis, experimentation, and measurable business outcomes. You can also explore the Lean Six Sigma Hub blog and the practical Pareto Analysis in the Analyse Phase guide to strengthen your DMAIC toolkit.

Pursue Lean Six Sigma Green Belt training and certification to apply ANOVA, Bartlett’s Test, Welch’s ANOVA, and other analytical methods with confidence in real improvement projects.

Kaizen. Kai-Care. Kai-Done. ( Lean Six Sigma)