Using Minitab’s Welch’s Two‑Sample t‑Test to Detect Mean Shifts When Variances Differ

Problem: Detecting a Mean Shift When Variances Are Unequal

An engineering team wants to know whether a change in a machine setting has altered the average diameter of a produced part. They have measurements from the process before the change (Group A) and after the change (Group B). Preliminary inspection shows that the spread of the after‑group data is noticeably larger than the before‑group data, so the assumption of equal variances is questionable. Using a standard two‑sample t‑test that assumes equal variances could give misleading conclusions.

Thesis: Minitab’s Built‑In Welch’s Test Provides a Reliable Alternative

Minitab includes a Welch’s approximate two‑sample t‑test under Stat > Basic Statistics > 2‑Sample t‑test. By checking the “Assume unequal variances” option, the procedure computes the Welch–Satterthwaite degrees of freedom, which adjust for variance disparity and sample‑size imbalance. The output supplies a t‑statistic, an approximate degrees‑of‑freedom (DF) value (often non‑integer), a p‑value, and a confidence interval for the difference of means. Optional graphs (boxplot, individual value plot) let you visualise spread and spot outliers alongside the numbers.

Running the Test in Minitab

1. Prepare the worksheet – place the before‑group measurements in one column (e.g., C1) and the after‑group measurements in another column (e.g., C2). No special permissions are required beyond having Minitab installed.
2. Open the dialog – choose Stat → Basic Statistics → 2‑Sample t‑test….
3. Specify the samples – select “Samples in different columns”, then assign C1 to the first sample and C2 to the second.
4. Enable Welch’s correction – check the box labelled Assume unequal variances. This triggers the Welch–Satterthwaite calculation.
5. Add diagnostic graphs – click the Graphs… button, enable Boxplot (and optionally Individual value plot), then click OK.
6. Run the analysis – press OK in the main dialog. Minitab will display the results in the Session window and, if requested, produce the graphs.

Interpreting the Output

When the dialog finishes, look for the following items in the Session window:

• T‑Value – the test statistic.
• DF – the approximate degrees of freedom. Because Welch’s formula adjusts for variance inequality, this value is usually a non‑integer (e.g., 12.7).
• P‑Value – the probability of observing a difference as extreme as the one seen if the true means were equal. Compare this to your chosen α level (commonly 0.05).
• Confidence Interval for Difference** – a range (e.g., (0.15, 2.85)) that estimates the true shift in means. If the interval does not contain zero, the shift is statistically significant at the chosen confidence level.

The accompanying boxplot shows the median, quartiles, and whiskers for each group. Overlapping boxes suggest similar spreads; a clearly shifted median supports a mean difference. Outliers appear as points beyond the whiskers and should be inspected before drawing final conclusions.

Worked Illustration (Using Placeholder Values)

Suppose you have loaded the before‑group data into C1 and the after‑group data into C2. After following the steps above, Minitab might produce output similar to the following:

Two-Sample T-Test and CI: C1, C2

N   Mean   StDev   SE Mean
C1  20    10.2     1.5      0.34
C2  20    12.5     2.8      0.63

Difference = μ(C1) - μ(C2)
Estimate for difference:  -2.30
95% CI for difference:   (-4.12, -0.48)
T-Test of difference = 0 (vs ≠):
T-Value = -2.34   DF = 12.7   P-Value = 0.036

In this illustrative output:

• The DF value (12.7) is non‑integer, confirming the Welch approximation.
• The p‑value (0.036) is below 0.05, indicating a statistically significant shift.
• The 95 % confidence interval for the difference (-4.12, -0.48) does not contain zero, reinforcing the conclusion that the after‑group mean is higher.
• The boxplot (not shown) would display the two groups with medians around 10.2 and 12.5, respectively, and possibly a few outliers worth reviewing.

Limitations and Practical Checks

While Welch’s test is robust to unequal variances, it relies on certain assumptions:

• Sample size – With very small groups (n < 5 per group) the Welch–Satterthwaite DF approximation can become unstable, potentially inflating Type I or Type II error rates. In such cases consider collecting more data or using a non‑parametric alternative.
• Normality – The test assumes each group is approximately normally distributed. Severe skewness or heavy tails can affect validity. Use Minitab’s Graph → Probability Plot or a Shapiro‑Wilk test to assess normality; if violated, consider a log‑transform or a Mann‑Whitney test.
• Outliers – The dialog does not automatically flag outliers. Always review the generated boxplot (or run Stat → Basic Statistics → Display Descriptive Statistics with the “Outliers” option) before concluding.
• Paired data – If the before and after measurements are paired (same part measured twice), the 2‑Sample t‑test is inappropriate; use Stat → Basic Statistics → Paired t‑test instead.

To verify that Minitab’s DF matches the theoretical Welch–Satterthwaite value, you can compute it manually:

DF = (s1²/n1 + s2²/n2)² / [ (s1²/n1)²/(n1-1) + (s2²/n2)²/(n2-1) ]

where s1, s2 are the sample standard deviations and n1, n2 the sample sizes. Compare the result to the DF reported in the Session window; they should agree within rounding tolerance.

Additionally, you can cross‑check the p‑value and confidence interval with an external tool such as R:

t.test(C1, C2, var.equal = FALSE)

The output should yield nearly identical t‑value, DF, p‑value, and CI (differences only due to rounding).

Actionable Closing

When you need to decide whether a process change has shifted a key metric and the groups show different spreads, follow these steps:

1. Organise your data into two columns.
2. Run Stat > Basic Statistics > 2‑Sample t‑test with “Assume unequal variances” checked.
3. Examine the t‑value, non‑integer DF, p‑value, and confidence interval.
4. Confirm the DF with the Welch–Satterthwaite formula and, if possible, compare to R’s t.test(var.equal=FALSE).
5. Review the boxplot (or individual value plot) for normality concerns and outliers.
6. If assumptions are violated, consider transformations, larger samples, or a non‑parametric alternative.

By using Minitab’s Welch’s test and verifying its output with the checks above, you can make an evidence‑based decision about process shifts while guarding against the pitfalls of variance inequality.