What is a Welch’s t-Test Calculator?
Welch’s t-test (also called the Welch test or the unequal-variances t-test) compares the means of two independent groups without assuming they share a variance. It divides x̄₁ − x̄₂ by √(s₁²/n₁ + s₂²/n₂) and judges the result against a t distribution whose degrees of freedom come from the Welch–Satterthwaite formula, usually a non-integer. If p ≤ α, the two population means differ.
What the Welch’s t-Test Calculator does
This Welch’s t-test calculator runs the test from summary statistics or raw data. It shows each group’s variance of the mean, the standard error, the t statistic, the Welch–Satterthwaite degrees of freedom with the full formula, the p-value, the critical value, a 95% confidence interval for μ₁ − μ₂ and Cohen’s d. It also reports the pooled t-test p-value so you can see whether the equal-variance assumption would have changed the answer.
Welch’s test is the default two-sample t-test in R, where the output is headed “Welch Two Sample t-test”, and it is recommended by many statisticians for routine use: when the variances really are equal it loses almost nothing, and when they are not it keeps the false-positive rate close to α. It matters most when the group sizes differ and the smaller group has the larger spread.
How to use the Welch’s t-Test Calculator
- Choose the alternative hypothesis for μ₁ − μ₂: two-tailed (≠) unless you predicted a direction before seeing the data, otherwise left (<) or right (>).
- Set the significance level α, usually 0.05.
- Enter each group’s mean, standard deviation and size under Summary statistics, or switch to Raw data and paste both samples (commas, spaces or one value per line). The groups can have different sizes.
- Read t, the Welch df and the p-value, then the decision. The step-by-step solution shows how the df formula combines the two variances, ready to copy into an assignment.
What is Welch’s t-test?
Welch’s t-test is a two-sample t-test for independent groups that drops the equal-variance assumption of Student’s pooled t-test. Instead of pooling the two sample variances into one estimate, it keeps s₁²/n₁ and s₂²/n₂ separate in the standard error, and it replaces the usual n₁ + n₂ − 2 degrees of freedom with an approximation that shrinks when one group is small and noisy.
The test is named after B. L. Welch, who published it in Biometrika in 1947; the degrees-of-freedom approximation it uses was set out by F. E. Satterthwaite in 1946, which is why it is called the Welch–Satterthwaite df. You will also see it written as the Welch t-test, Welch’s test, the Welch two sample t-test or, in SPSS output, the “equal variances not assumed” row (and occasionally misspelled as the Welsh or Welches t-test).
The null hypothesis is H₀: μ₁ = μ₂. The test needs two independent random samples from populations that are roughly normal; it does not need equal variances or equal sample sizes.
How to interpret Welch’s t-test results
To report it, give the test, t, the fractional df, p and the interval, for example: “Welch’s t-test, t(13.67) = 2.65, p = .019, mean difference 4.58, 95% CI [0.87, 8.30].”
- p-value: if p ≤ α, reject H₀ and conclude the population means differ (or differ in the direction you chose). If p > α, the data do not show a difference; that is not proof the means are equal.
- t statistic: the gap between the sample means in standard-error units. Its sign tells you which group is higher: positive means x̄₁ > x̄₂.
- Welch df: always between min(n₁, n₂) − 1 and n₁ + n₂ − 2. A df close to the lower end means one group’s variance dominates the standard error, so that group’s size matters most.
- 95% confidence interval for μ₁ − μ₂: the plausible size of the difference. If it excludes 0 the two-tailed test is significant at α = 0.05; its width shows how precise the estimate is.
- Cohen’s d: the difference in standard-deviation units (about 0.2 small, 0.5 medium, 0.8 large). A significant p with a tiny d is a real but probably unimportant difference.
- Pooled t-test p: shown for comparison. If the two p-values lead to different decisions, the unequal variances matter and Welch’s answer is the one to report.
How to calculate Welch’s t-test by hand
- Find each sample’s mean x̄, standard deviation s and size n.
- Divide each variance by its sample size: s₁²/n₁ and s₂²/n₂.
- Add the two and take the square root to get the standard error.
- Divide the difference in means, x̄₁ − x̄₂, by the standard error to get t.
- Plug the two variances of the mean into the Welch–Satterthwaite formula (shown in full below) to get df.
- Find the p-value from the t distribution with that df, using a P-Value from t Calculator because a printed t-table cannot handle fractional df, and compare it with α.
Welch’s t-test formula
| Symbol | Meaning |
|---|---|
| x̄₁, x̄₂ | sample means |
| s₁, s₂ | sample standard deviations |
| n₁, n₂ | sample sizes |
| df | approximate degrees of freedom, between min(n₁, n₂) − 1 and n₁ + n₂ − 2 |
This calculator uses the exact (fractional) df, as R and SciPy do. Some textbooks round df down to a whole number or use the conservative min(n₁ − 1, n₂ − 1); Excel’s Data Analysis ToolPak (“t-Test: Two-Sample Assuming Unequal Variances”) rounds it to a whole number. Those choices give slightly different p-values. You can check the df on its own with the Degrees of Freedom Calculator.
Worked example
Worked example: delivery times from two couriers
Problem. Courier A: n₁ = 12 deliveries, mean 31.4 min, s₁ = 4.2. Courier B: n₂ = 20, mean 27.1 min, s₂ = 9.8. Do the mean delivery times differ at α = 0.05?
- 01Variance of each mean: 4.2² / 12 = 1.47 and 9.8² / 20 = 4.802.
- 02SE = √(1.47 + 4.802) = √6.272 = 2.5044.
- 03t = (31.4 − 27.1) / 2.5044 = 1.7170.
- 04df = 6.272² / (1.47²/11 + 4.802²/19) = 39.3380 / 1.4101 = 27.8975.
- 05Two-tailed p-value = 2 · P(T ≥ 1.7170) ≈ 0.0971; critical value ±2.0487.
Answer. p ≈ 0.097 > 0.05, so fail to reject H₀. The 4.3-minute gap is not statistically significant; the 95% CI for the difference, (−0.8309, 9.4309), includes zero. The pooled t-test gives p ≈ 0.161 here: because the larger group has the larger spread, pooling overstates the standard error and makes the test too conservative.
Worked example 2: raw data with very different spreads
Two teaching methods are compared on a 40-point quiz. Method A (10 students): 34, 31, 33, 29, 32, 35, 30, 33, 31, 32. Method B (12 students): 22, 30, 27, 35, 19, 28, 25, 38, 24, 31, 21, 29. These are the calculator’s raw-data defaults, so switch to Raw data to follow along. Use α = 0.05, two-tailed.
- Summaries: x̄₁ = 32, s₁ = 1.8257, n₁ = 10; x̄₂ = 27.4167, s₂ = 5.6481, n₂ = 12. Method B’s SD is about three times Method A’s.
- Variance of each mean: 1.8257² / 10 = 0.3333 and 5.6481² / 12 = 2.6585.
- SE = √(0.3333 + 2.6585) = √2.9918 = 1.7297.
- t = (32 − 27.4167) / 1.7297 = 2.6498.
- df = 2.9918² / (0.3333²/9 + 2.6585²/11) = 13.6688, far below the pooled df of 20 because almost all the uncertainty comes from Method B.
- Two-tailed p = 0.0193 < 0.05, critical value ±2.1497, so reject H₀. 95% CI for μ₁ − μ₂: (0.8651, 8.3016); Cohen’s d ≈ 1.09.
Welch’s t-test vs Student’s t-test
Yes, Welch’s t-test is a two-sample t-test: both versions compare two independent means with a t statistic. They differ only in how they estimate the standard error and the degrees of freedom.
| Welch’s t-test | Student’s pooled t-test | |
|---|---|---|
| Equal variances assumed? | No | Yes |
| Standard error | √(s₁²/n₁ + s₂²/n₂) | sₚ·√(1/n₁ + 1/n₂) |
| Degrees of freedom | Welch–Satterthwaite, usually fractional | n₁ + n₂ − 2 |
| When variances and sizes are unequal | Error rate stays close to α | Too many false positives if the smaller group has the larger variance; too few if it has the smaller |
| When variances are equal | Almost identical to pooled | Exact |
| R | t.test(x, y) (the default) | t.test(x, y, var.equal = TRUE) |
| Excel | T.TEST(r1, r2, 2, 3) | T.TEST(r1, r2, 2, 2) |
| Python (SciPy) | ttest_ind(x, y, equal_var=False) | ttest_ind(x, y) |
Common use cases
- Comparing two groups whose spreads clearly differ, such as a stable process versus an unstable one.
- Experiments with unequal group sizes, where unequal variances distort the pooled test most.
- A/B tests on continuous metrics (revenue per user, session length) where variances are rarely equal.
- Clinical and psychology studies comparing a treatment group with a control group, where many journals now expect Welch’s test by default.
- Any two-independent-sample comparison when you would rather not test variances first.
Assumptions and limitations
- Independence: the two samples must contain different, unrelated subjects. Matched or repeated measurements need the Paired t-Test Calculator.
- Normality: each population should be roughly normal for small samples; with 30+ per group the test tolerates moderate skew. For heavily skewed small samples consider a rank-based test such as Mann–Whitney, and check normality with the Shapiro–Wilk test.
- Testing variances first with an F-test or Levene’s test and then choosing between pooled and Welch is not recommended; the two-stage procedure distorts error rates. Using Welch from the start avoids the problem.
- With equal variances and equal sizes Welch and the pooled t-test give nearly identical answers, so little is lost by defaulting to Welch.
- For three or more groups with unequal variances, use Welch’s ANOVA rather than running several Welch t-tests, which inflates the overall error rate. The ANOVA Calculator covers the standard one-way case.
Frequently asked questions
When should I use Welch’s t-test instead of Student’s t-test?
Is Welch’s t-test a two-sample t-test?
How do you calculate Welch’s t-test?
Why are the degrees of freedom not a whole number?
Is Welch’s t-test the same as an unpaired t-test?
What effect size should I report with Welch’s t-test?
Is it Welch ANOVA or Welch’s ANOVA?
How do I calculate Welch’s t-test in Excel or R?
Cite this page
Statculator. (2026, November 14). Welch’s t-Test Calculator. Retrieved November 14, 2026, from https://statculator.com/hypothesis-tests/welchs-t-test-calculator/