Paired vs Unpaired t-Test: Which One Should You Use?
A paired t-test and an unpaired t-test answer the same question — do these two means differ? — but they treat the data differently, and on the same numbers they can give opposite answers. This guide runs one dataset through both tests, shows why pairing is more powerful, and gives a decision table for choosing the right one.
The short version: if each measurement in one group has a natural partner in the other group — the same student before and after a course, the same patient on two treatments — the data are paired, and you should use a paired t-test. Pairing removes the variation between subjects, so the test sees the difference far more clearly. Use the unpaired test only when the two groups are genuinely independent.
That distinction is the whole guide. We define pairing, explain both tests in a paragraph each, run one dataset — eight students, before and after a course — through both tests and get opposite answers, show the maths that makes pairing more powerful, list the mistakes that cost power, and finish with a decision table. The two tests are run on the same numbers, so the only thing that changes is how the data are treated.
What makes data paired?
Data are paired when each observation in one group has a natural partner in the other group, and the pairing is part of the design — not something you discover after collecting the numbers. The classic cases are before-and-after measurements on the same subjects (a student’s score before a course and after it), matched pairs (twins, or patients matched on age and sex), and two measurements on the same unit (blood pressure in the left arm and the right arm, or the same sample tested by two methods).
What matters is that the two values in a pair are linked. A student’s “before” score and “after” score move together: a strong student scores high both times, a weaker student scores lower both times. That link — the correlation between the paired values — is what the paired test exploits and the unpaired test ignores.
The test for pairing is a design question, not a numbers question. Ask: “Is there a reason these two values belong together?” If the answer is yes — same person, same unit, matched on purpose — the data are paired. If the two groups are just two separate collections of people or things with no link between them, the data are unpaired (independent). Equal sample sizes do not make data paired: two groups of 8 with no link between them are unpaired, however equal their n. We come back to that myth in the mistakes section.
When the data are paired, the analysis works on the differences. Each pair gives one difference (after − before), and the paired t-test is simply a one-sample t-test on those differences: it tests whether the mean difference is zero. That is why the paired test is so powerful — it throws away the between-subject variation entirely and looks only at the changes. The five steps of a hypothesis test are the same for both tests; our hypothesis testing explained guide works them through with z, t and chi-square examples.
The paired t-test in one paragraph
The paired t-test takes the differences within each pair, then asks whether their mean is different from zero. For n pairs, compute each difference dᵢ, then the mean difference d̄ and its standard deviation s_d. The statistic is t = d̄ / (s_d / √n) with df = n − 1. Because it works only with the differences, the paired test never sees the variation between subjects — a student who scores 80 before and 82 after contributes a difference of 2, exactly like a student who scores 68 before and 70 after. The between-student spread (68 to 80) is irrelevant. Our paired t-test calculator runs the test, draws the histogram of the differences, and gives the confidence interval for the mean difference. The method follows the NIST/SEMATECH e-Handbook, 7.3.1.1. Analysis of paired observations.
The unpaired t-test in one paragraph
The unpaired (independent-samples) t-test compares the means of two separate groups that have no link between them. It takes the difference of the two sample means, x̄₁ − x̄₂, and divides by the standard error of that difference, which combines the spread of both groups: SE = √(s₁²/n₁ + s₂²/n₂) for Welch’s version, which does not assume equal variances. The statistic is t = (x̄₁ − x̄₂) / SE, with degrees of freedom from the Welch–Satterthwaite formula (or n₁ + n₂ − 2 for the pooled version). Because the standard error carries the full between-subject spread of both groups, the unpaired test needs a much bigger mean difference to reach significance. Our two-sample t-test calculator runs both the Welch and pooled versions. The method follows the NIST/SEMATECH e-Handbook, 1.3.5.3. Two-Sample t-Test for Equal Means.
The same data, both tests
Eight students take a course. Their scores before and after:
| Student | Before | After | Difference (After − Before) |
|---|---|---|---|
| 1 | 72 | 76 | 4 |
| 2 | 75 | 78 | 3 |
| 3 | 68 | 71 | 3 |
| 4 | 80 | 82 | 2 |
| 5 | 74 | 77 | 3 |
| 6 | 71 | 74 | 3 |
| 7 | 79 | 81 | 2 |
| 8 | 76 | 79 | 3 |
| Mean | 74.375 | 77.250 | 2.8750 |
| SD | 4.033 | 3.615 | 0.6409 |
Every student improved — by 2, 3 or 4 points. Now run both tests on these numbers.
| Test | Statistic | df | p-value | Verdict at α = 0.05 |
|---|---|---|---|---|
| Paired t-test | t = 12.6886 | 7 | p < 0.0001 | Reject H₀ — the course clearly helped |
| Unpaired t-test (Welch) | t = 1.5013 | ≈ 13.8 | 0.1558 | Fail to reject — no evidence at all |
Same numbers, opposite answers. The paired test sees a huge effect (t = 12.69) because it works only with the differences, which are tightly clustered (s = 0.6409): every student improved by about 3 points, so the mean difference of 2.875 is more than 12 standard errors from zero. The unpaired test drowns: it compares means of 74.375 and 77.250 against the full between-student spread (s ≈ 4 in each group), so a 2.875-point gap looks like noise (t = 1.50, p = 0.1558). If p-values are new to you, our what is a p-value guide explains what 0.1558 does and does not mean.
The difference is the correlation. Before and after scores move together almost perfectly (r = 0.9919): the student who scored 80 before scored 82 after, the one who scored 68 before scored 71 after. Pairing removes that shared variation; the unpaired test has to fight through it.
In Excel, the paired p-value is =T.TEST(before_range, after_range, 2, 1) and the unpaired (Welch) p-value is =T.TEST(before_range, after_range, 2, 3). The type argument is the whole choice: 1 = paired, 2 = unpaired with equal variances, 3 = unpaired with unequal variances (Welch). You can look up either p-value on our t-distribution calculator.
Why pairing is more powerful
Pairing buys power by removing the variation you do not care about. In the unpaired test, the standard error carries the spread of both groups — the fact that students differ from each other by 4 points either way. In the paired test, the standard error carries only the spread of the differences — how much the improvements vary from student to student (s = 0.6409). When the paired measurements are highly correlated, the differences are far less spread out than the raw scores, so the paired standard error is far smaller and the t-statistic far larger.
The maths: pairing multiplies the effective information by roughly 1/(1−ρ), where ρ is the correlation between the paired measurements. For this dataset ρ = 0.9919, so 1/(1−ρ) = 1/0.0081 = 123.5. That is this dataset’s number — it comes from this dataset’s ρ, and it is not a general claim. The general lesson is the direction: the higher the correlation between the paired measurements, the more power pairing buys. With ρ near 0, pairing buys nothing (the differences are as spread out as the raw scores); with ρ near 1, pairing buys a great deal.
A useful way to see it: the paired test’s standard error is s_d/√n, and s_d² = s₁² + s₂² − 2ρ·s₁·s₂. When ρ is high, the subtraction term nearly cancels the two variances and s_d is small. When ρ = 0, s_d² = s₁² + s₂² and the paired test has no advantage over the unpaired one. So pairing is worth the design effort exactly when the measurements are correlated — same subject, same unit, matched pairs.
Eight students, before and after — every line rises
Each line is one student. All eight lines rise: the smallest gain is +2, the largest +4.
The mistakes that cost power
1. Running an unpaired test on paired data. This is the worked example’s mistake: it turns a clear result (p < 0.0001) into no evidence at all (p = 0.1558). If the data are paired, the paired test is not optional — it is the test that matches the design. Using the unpaired test throws away the pairing information and with it most of your power.
2. Pairing unrelated groups. The reverse error: inventing pairs where none exist — lining up the first person in group A with the first person in group B because the lists happen to be the same length. A pairing must come from the design (same subject, matched on purpose). Pairing by row order is fabrication, and the paired test on fabricated pairs is meaningless.
3. “Equal n means paired.” Equal sample sizes do not make data paired. Pairing is about the design — same subjects, matched pairs — not the counts. Two independent groups of 8 are unpaired, however equal their n. Conversely, paired data can have unequal n only if some pairs are incomplete (a missing “after” score) — and then the analysis uses only the complete pairs, which is why the FAQ below covers it.
Decision table
Use this table to choose:
| Scenario | Test |
|---|---|
| Same subjects measured twice (before/after) | Paired t-test |
| Matched pairs (twins, matched on age/sex) | Paired t-test |
| Two measurements on the same unit (left/right, method A/B) | Paired t-test |
| Two separate, unrelated groups | Unpaired t-test (Welch) |
| Two separate groups, equal variances assumed | Unpaired t-test (pooled) |
| Paired data, differences clearly non-normal, small n | Wilcoxon signed-rank test |
When in doubt, ask the design question: “Does each value in group 1 have a partner in group 2?” Yes → paired. No → unpaired. Choosing between a z-test and a t-test is a different question — see our z-test vs t-test guide.
Frequently Asked Questions
How do I know if my data are paired?
Ask whether each observation in one group has a natural partner in the other, by design. Same subject before and after, matched pairs, two measurements on the same unit — paired. Two separate groups with no link between them — unpaired. If you have to invent the pairing after collecting the data, it is not paired.
Can I run an unpaired test on paired data?
You can, but you should not. It is valid in the sense that it will give you an answer, but it throws away the pairing and with it most of your power — in the worked example it turned p < 0.0001 into p = 0.1558. Use the test that matches the design.
What if the pairs have different n?
Pairs cannot have different n — a pair is two values, one from each condition. What happens in practice is a missing value: a student who took the “before” test but missed the “after” test. That pair is incomplete and is dropped from the paired analysis (the test uses only complete pairs). If many pairs are incomplete, consider whether the missingness itself is informative.
Do paired tests need normal data?
The differences need to be roughly normal — that is the assumption, not the raw scores. With 8 pairs, a clearly skewed set of differences is a problem; with 30+ pairs the central limit theorem covers moderate skew. If the differences are heavily skewed or have extreme outliers, the Wilcoxon signed-rank test is the non-parametric alternative.
What if some differences are zero?
Zero differences are informative — they are pairs where nothing changed — and they stay in the analysis. They pull the mean difference toward zero and increase the standard deviation slightly, which is exactly right: a treatment that helps half the subjects and does nothing for the rest should look weaker than one that helps everyone. Do not drop them.
Is the paired t-test the same as a one-sample t-test?
Yes — on the differences. The paired t-test is a one-sample t-test applied to the within-pair differences, testing whether their mean is zero. If you can run a one-sample t-test, you can run a paired t-test: just feed it the differences. Our one-sample t-test calculator does exactly that if you enter the differences as the data.
When should I use Wilcoxon instead?
When the differences are clearly non-normal and n is small — heavy skew, extreme outliers, or a small sample where you cannot verify normality. The Wilcoxon signed-rank test does not assume normal differences. It is the non-parametric companion to the paired t-test, and it will have its own page in a later batch.
Check Tools
Try the numbers from this guide with the site’s free calculators:
One-Sample T-Test Calculator
Test one mean against a claimed value — or run a paired test by entering the differences.
Try the one-sample t-test calculator →
Paired t-Test Calculator
Test before-and-after differences on the same subjects, with a histogram of the differences.
Try the paired t-test calculator →
Two-Sample t-Test Calculator
Compare two independent groups with Welch or pooled t intervals.
Try the two-sample t-test calculator →
Explore All Statistical Tools
See every calculator on the site in one place.
Browse all our statistical tools →
Check Articles
Go deeper with these guides:
Hypothesis Testing Explained
The five steps of a hypothesis test, with worked z, t and chi-square examples.
Read the guide →
Z-Test vs t-Test
When to use a z-test and when to use a t-test — sigma known or not.
Read the guide →
T-Distribution vs Normal Distribution
How the t-curve differs from the normal curve — and when they agree.
Read the guide →
What Is a P-Value?
What a p-value really means — and what it does not.
Read the guide →