Welch's t-test vs Student's t-test: When Equal Variances Fail and Which One to Report

Welch's t-test and Student's t-test both compare the means of two independent groups, but Welch's version does not assume the two groups have equal spread, so it stays accurate when variances differ. Use Welch's by default. If your groups have unequal sizes and unequal variability, Student's t-test can give you a misleading p value, sometimes a falsely "significant" one. The two tests can disagree on the same data, so choosing the wrong one could change your thesis conclusion.

Key Takeaways

  • Student's t-test assumes both groups have equal variances. Welch's t-test does not.
  • When group sizes and variances are both unequal, Student's t-test can inflate or deflate the false-positive rate well beyond 5%.
  • When variances really are equal, Welch's t-test performs almost identically to Student's, so it costs you very little.
  • R's t.test() runs Welch's test by default (var.equal = FALSE). Methodologists such as Delacre, Lakens and Leys (2017) recommend Welch as the default for psychology research.
  • In APA 7, a Welch result is reported with fractional degrees of freedom, e.g., t(9.73) = 1.86, p = .093.

What is the difference between Welch's t-test and Student's t-test?

Both tests ask the same question: is the difference between two group means bigger than random noise would produce? They differ only in how they estimate that noise.

Student's t-test pools the two groups' variances into one shared estimate. This is efficient when the groups truly vary by the same amount. It goes wrong when they don't.

Welch's t-test keeps each group's variance separate. It then adjusts the degrees of freedom downward using the Welch–Satterthwaite formula to reflect the extra uncertainty. That is why Welch results often show non-whole degrees of freedom like 37.4.

The intuition: pooling assumes one shared "wobble" for both groups. If one group is far more variable, pooling averages that wobble away and misjudges how noisy your mean difference really is.

When do equal variances fail, and why does it matter?

Unequal variances are only dangerous in combination with unequal sample sizes. The direction of the error depends on which group is noisier:

  • Smaller group has the larger variance: Student's test underestimates the noise. It becomes too liberal, producing false positives far more often than 5%.
  • Larger group has the larger variance: Student's test overestimates the noise. It becomes too conservative and misses real effects.

With equal group sizes, Student's test is fairly robust even if variances differ. Real thesis data rarely has perfectly balanced groups, though. Clinical samples, dropout, and convenience sampling all create imbalance.

Welch vs Student: key differences

Feature Student's t-test Welch's t-test
Assumes equal variances Yes No
Variance estimate Pooled across groups Separate for each group
Degrees of freedom n₁ + n₂ − 2 (whole number) Welch–Satterthwaite (often fractional)
Unequal n and unequal variances Can be badly biased Stays accurate
Equal variances Optimal Nearly identical power
Default in R No Yes
SPSS output row "Equal variances assumed" "Equal variances not assumed"

Should I use Levene's test to decide?

Probably not. The traditional workflow runs Levene's test first and picks Student's if Levene's is non-significant. This two-step approach has two problems:

  1. Levene's test is underpowered in small samples. It often misses real variance differences exactly when they matter most.
  2. Making one test conditional on another distorts your overall error rate.

The simpler rule is to report Welch's t-test unless you have a strong design-based reason to believe variances are equal. You lose almost nothing and avoid a known failure mode. If you're unsure whether a t-test is even right for your data, see our guide on Mann-Whitney vs the t-test.

Worked example: when the two tests disagree

Suppose you compare anxiety scores between a clinical sample and a community sample:

  • Clinical: n = 10, M = 30.0, SD = 10.0
  • Community: n = 40, M = 24.0, SD = 4.0

The mean difference is 6.0 points. Note the pattern: the smaller group has the much larger variance. This is the exact scenario where Student's test is too liberal.

Step 1: Student's t-test (pooled).
Pooled variance = [(9 × 100) + (39 × 16)] / 48 = 31.75.
Standard error = √[31.75 × (1/10 + 1/40)] = 1.99.
t(48) = 6.0 / 1.99 = 3.01, p = .004. That looks like a clear, significant effect.

Step 2: Welch's t-test (separate variances).
Standard error = √(100/10 + 16/40) = √10.4 = 3.22.
t = 6.0 / 3.22 = 1.86.
The Welch–Satterthwaite degrees of freedom come out to 9.73, close to the clinical group's n − 1, because that group dominates the uncertainty.
t(9.73) = 1.86, p = .093.

Step 3: Interpret.
The same data give p = .004 under Student's test and p = .093 under Welch's. Student's test treated the noisy clinical group as if it were as stable as the community group, and that shrank the standard error. Welch's 95% confidence interval for the difference is [−1.21, 13.21]. It includes zero, so the evidence is suggestive but not conclusive.

The effect size uses the average of the two variances as the standardiser: d = 6.0 / √[(100 + 16)/2] = 0.79.

How do I report Welch's t-test in APA 7?

Report the fractional degrees of freedom to two decimals. Italicise t, p, and d, drop the leading zero on p, and include the effect size and confidence interval:

An independent-samples Welch's t-test showed that the clinical group (M = 30.0, SD = 10.0) did not score significantly higher on anxiety than the community group (M = 24.0, SD = 4.0), t(9.73) = 1.86, p = .093, d = 0.79, 95% CI [−1.21, 13.21].

Name "Welch's" explicitly in your Methods section so reviewers know why your degrees of freedom aren't whole numbers.

How do SPSS, R, JASP, and jamovi handle this?

  • SPSS prints both rows ("Equal variances assumed" / "not assumed") next to Levene's test, and you choose which to report.
  • R runs Welch by default in t.test().
  • JASP and jamovi offer a "Welch" checkbox in the independent-samples t-test menu.
  • StatRyx checks your group sizes and variances, selects Welch's test when it's the safer choice, and writes the APA 7 sentence for you, including the fractional df, effect size, and CI.

If you've ever had to figure out which SPSS row to copy, StatRyx is built for that problem. It's an AI-powered statistical analysis tool that replaces manual SPSS workflows with automated, APA 7-formatted reporting.

Stop calculating this by hand — run it free in StatRyx → Try StatRyx

Frequently Asked Questions

Is Welch's t-test always better than Student's t-test?

For practical purposes, almost always. When variances are equal, Welch's test has nearly the same power as Student's. When they're unequal, Welch's avoids serious error-rate problems. The main exception is very small samples where you have strong theoretical grounds to assume equal variances.

Why does Welch's t-test give decimal degrees of freedom?

The Welch–Satterthwaite formula estimates effective degrees of freedom from each group's variance and size. The result is rarely a whole number. Report it to two decimals in APA 7, e.g., t(37.42).

What should I report from SPSS: equal variances assumed or not assumed?

Report the "Equal variances not assumed" row, which is Welch's t-test. Don't base the choice on Levene's test, because it often lacks the power to detect real variance differences in small samples.

Can I use Welch's t-test if my data aren't normally distributed?

Welch's test corrects for unequal variances, not non-normality. With moderate or large samples it tolerates mild skew reasonably well. With small, heavily skewed samples, consider a Mann-Whitney U test or a bootstrapped comparison instead.

Does Welch's t-test work for paired data?

No. Welch's and Student's independent-samples tests are for two separate groups. For before-and-after or matched measurements, use a paired-samples t-test, which analyses the difference scores directly.

Stop calculating this by hand. Upload your dataset and StatRyx's AI runs the correct test and returns copy-paste-ready APA 7 output in seconds — no SPSS license, no syntax.

Run your data through StatRyx free →
← All posts