Welch vs Student t-test
Both tests compare two independent means. Student's pooled test assumes the groups share one variance; Welch's test does not. When that assumption fails, especially with unequal group sizes, only Welch's p-value can be trusted.
Sources checked on November 21, 2026. Results computed by the site's engine and tested against R and SciPy.
The one difference
Both tests divide the same mean difference by a standard error. Student's test first pools the two sample variances into one, weighting by degrees of freedom, and uses n1 + n2 − 2 df. Welch's test keeps the variances separate and estimates the df from the data with the Welch-Satterthwaite formula. That is all; everything else, including how you read the result, is the same.
When they agree and when they do not
With equal group sizes, the two standard errors are identical, so t is the same and only the df differ. With Student's sleep data treated as two groups of 10, both give t = −1.8608; Student uses 18 df (p = 0.0792) and Welch 17.78 (p = 0.0794). The difference is trivial.
With unequal sizes and unequal spreads, they can part ways. In R's own example, one group is the numbers 1 to 10 and the other is 7 to 20 plus one extreme value of 200:
| Data | Welch t, df, p | Student t, df, p |
|---|---|---|
| Without the extreme value | −5.43, 21.98, 1.855 × 10−5 | −5.15, 22, 3.691 × 10−5 |
| With the extreme value | −1.63, 14.16, 0.1245 | −1.33, 23, 0.1979 |
| NIST car mileage (unequal n) | −12.95, 136.87, 1.571 × 10−25 | −12.62, 326, 5.273 × 10−30 |
In the NIST example the US group is about 3.2 times the size of the Japanese group, and Welch's df of 136.87 is far below Student's 326. The conclusion does not change, but the precision claimed by Student's test is not justified.
Why Welch is the safer default
Simulation studies summarized by Delacre, Lakens and Leys (2017) show that when variances differ and the larger group has the smaller variance, Student's test rejects a true null hypothesis far more often than the nominal 5%. When the larger group has the larger variance, it loses power. Welch's test holds close to the nominal rate in both cases, and when variances really are equal it gives up very little. Equal variances are also hard to verify with small samples.
This is why R's t.test has used Welch by default (var.equal = FALSE) for a long time. SciPy's ttest_ind, by contrast, still defaults to equal_var=True, so Python users get Student's test unless they ask for Welch.
Do not test for equal variances first
A common recipe is to run Levene's test, then use Student's test if it is not significant and Welch's if it is. That two-step procedure has worse error rates than simply using Welch's test, because the variance test has low power exactly when it matters, with small samples. Decide before seeing the data, and report Welch. Our result panel shows the ratio of the two SDs for information, never as a switch.
The Welch-Satterthwaite degrees of freedom
df = (s1²/n1 + s2²/n2)² / [(s1²/n1)²/(n1−1) + (s2²/n2)²/(n2−1)]
The df lies between the smaller of n1 − 1 and n2 − 1 and the pooled value n1 + n2 − 2. It is usually a decimal. Compute p with the decimal value; truncating it, as some older tables and tools do, makes the test a little conservative. For R's mtcars example the df is 18.33 and R prints 18.332.
Reporting Welch's test
Give the df to two decimals and say which test you used: "Welch's t-test showed that manual cars had higher fuel economy, t(18.33) = −3.77, p = .001." Many journals now accept Welch without comment. If a reviewer asks for Student's test, the calculator shows it beside Welch's in the same result, so you can report both.
Sources
- Delacre M., Lakens D., Leys C. (2017) Why psychologists should by default use Welch's t-test instead of Student's t-test. International Review of Social Psychology 30(1):92-101 checked November 21, 2026
- R source code, stats package, t.test.R checked November 21, 2026
- SciPy documentation: scipy.stats.ttest_ind checked November 21, 2026