Paired t Test Calculator
Test whether the mean change within the same people, objects, locations, or matched units differs from zero. Paste aligned before-and-after values, choose the direction of the difference, and receive the inferential result together with the actual row-level changes that produced it.
Enter matched observations
Paired result
The worked example shows a positive mean second-minus-first change.
Difference strip
Pair 1: +3.0000 | Pair 2: +2.0000 | Pair 3: +3.0000 | Pair 4: +0.0000 | Pair 5: +3.0000 | Pair 6: +3.0000 | Pair 7: +2.0000 | Pair 8: +3.0000
t = 2.3750 ÷ (1.0607 / √8) = 6.3333
What the paired t test is answering
A paired t test asks whether the population mean of a set of within-pair differences is a specified value, normally zero. The word “paired” describes the data structure, not a special two-sample formula. For every unit, compute a difference: follow-up minus baseline, method B minus method A, left side minus right side, or another direction chosen before examining the result. The resulting difference scores form one sample. Their mean is compared with zero using their sample standard deviation and their count.
This design is valuable because stable differences between units can cancel. Suppose eight employees complete the same timed task before and after training. Some employees are consistently fast and others consistently slow, but the research question concerns each employee’s improvement. Comparing independent group averages would mix person-to-person variation into the error term. Pairing keeps the employee attached to both measurements and analyzes only the change. The same reasoning applies to twins matched by characteristics, instruments tested on identical specimens, or plots measured in two seasons.
The calculator defines the displayed effect in the direction selected above. Reversing first and second changes the signs of the mean, t statistic, confidence limits, and dz, but it does not change the two-sided p-value. Always write the direction in your report. “Second minus first” is more informative than an unlabeled statement that the mean difference was 2.38.
How the calculation works
1. Build differences
For pair i, calculate di = secondi − firsti, unless the reverse direction was selected. The count n is the number of complete pairs. An observation without its partner cannot contribute to this standard calculation.
2. Summarize change
Compute the mean difference d̄ and the sample standard deviation sd. The estimated standard error is sd/√n. This standard error describes uncertainty in the sample mean difference, not the variability of individual first or second measurements.
3. Form the t statistic
For a null mean difference of zero, t = d̄/(sd/√n), with n − 1 degrees of freedom. A large absolute t indicates that the observed mean change is large relative to its estimated sampling uncertainty.
4. Infer and estimate
The chosen alternative determines whether one or both tails of the t distribution count toward the p-value. The confidence interval is d̄ ± t* × SE for a two-sided confidence level. The standardized paired effect is dz = d̄/sd.
Reading every result field
| Output | Practical reading | What it does not prove |
|---|---|---|
| Mean difference | The average signed change in the original measurement units. | It is not the average absolute change and does not show how many pairs moved in each direction. |
| t and df | The mean difference expressed in estimated standard-error units, with df = n − 1. | A large t does not by itself establish a useful or causal effect. |
| p-value | Under the null model, the probability of a t result at least as unfavorable to the null as observed, according to the selected alternative. | It is not the probability that the null hypothesis is true and not the chance that the result will replicate. |
| Confidence interval | A range of mean differences compatible with the sample and model at the selected confidence procedure. | It is not a range containing a chosen percentage of individual differences. |
| dz | The mean change divided by the standard deviation of paired changes. | It is not directly interchangeable with an independent-groups Cohen’s d that uses another denominator. |
Worked example from the calculator
The default data contain eight aligned pairs. Defining change as second minus first produces differences of 3, 2, 3, 0, 3, 3, 2, and 3. Their sum is 19 and their mean is 19/8 = 2.375. The sample standard deviation of those changes is approximately 1.0607, so the standard error is 1.0607/√8 = 0.3750. Dividing the mean by that standard error gives t = 6.3333 with 7 degrees of freedom.
For a two-sided 5% test, the two-sided p-value is small and the 95% confidence interval is approximately 1.4882 to 3.2618. The interval excludes zero and remains positive. In an applied report, the original units and context must be restored: “The second measurement was on average 2.38 units higher than the first, 95% CI [1.49, 3.26], t(7) = 6.33.” Add the p-value and preregistered decision rule when appropriate. The paired effect dz is about 2.24, but its meaning depends on the difference-score variability in this sample.
Assumptions and data checks
Pairs should be independent of other pairs. Measurements within a pair are expected to be related; that dependence is exactly why pairing is used. However, one person’s change should not determine another person’s change unless the study model addresses clustering. Household members, repeated classrooms, multiple eyes from one patient, and repeated time points beyond two may require a multilevel model, repeated-measures model, or cluster-aware analysis.
The difference scores should be reasonably modeled by a normal distribution, particularly for small n. The test does not require each original column to be normal. Inspect the difference strip for gross data-entry errors, strong skew, and isolated extreme changes. With a larger, well-sampled dataset, the mean can be less sensitive to moderate nonnormality, but dependence and biased sampling remain serious problems at any size.
The outcome should support meaningful differences. Interval or ratio measurements normally fit. A five-point ordinal response may not justify treating every one-step change as equal. When difference scores are strongly nonnormal or only ordinal, a signed-rank or sign-based procedure may be more defensible, although those tests target different population features. Do not select a method only because it produces the preferred significance label.
Missingness deserves an explicit rule. Deleting an incomplete pair is sometimes called complete-pair analysis, but it can be biased when the absence of the second measurement relates to the outcome. Record the number excluded and why. Also confirm that the two pasted lists have equal length and identical ordering; this calculator rejects unequal counts but cannot detect a plausible-looking misalignment.
Planning, reporting, and common mistakes
A nonsignificant result is not automatically evidence of equivalence. If the practical question is whether change is small enough to ignore, define equivalence bounds and use an equivalence procedure designed around them. A wide confidence interval crossing zero may simply reveal insufficient precision. Plan sample size using the expected distribution of paired differences, the smallest meaningful mean change, the chosen alpha, desired power, and anticipated loss of complete pairs.
A one-sided alternative should reflect a scientific claim chosen before inspecting outcomes. If improvement and harm would both alter the decision, a two-sided analysis is usually appropriate. Switching to one-sided after observing the sign understates the evidential cost. Similarly, analyzing several outcomes or subgroups increases the chance of a small p-value; disclose multiplicity and distinguish confirmatory from exploratory analysis.
Report the pairing mechanism, the direction of subtraction, n complete pairs, mean and standard deviation of the differences, mean change with confidence interval, t statistic, degrees of freedom, p-value, and any exclusions. Graph the raw paired trajectories or the distribution of differences when space permits. A result is much easier to audit when readers can see that a few unusual pairs did not create the entire conclusion.
If you need to inspect the spread of one list separately, use the standard deviation calculator. These links are conceptual neighbors, not substitutes for choosing the correct sampling structure.
Frequently asked questions
Can I use unequal sample sizes?
No. A standard paired t test requires one second measurement for every included first measurement. If the columns represent unrelated groups with different counts, use an independent-samples method. If a paired value is missing, address the missing-data process rather than inventing a partner.
Does the order of the two columns matter?
It changes the sign, not the strength of a two-sided test. Second minus first and first minus second produce opposite means, t statistics, intervals, and effect sizes. State the direction so a positive result has an unambiguous practical meaning.
Why test the differences instead of both means?
The hypothesis concerns mean within-unit change. Reducing every pair to a difference removes stable between-unit levels from the comparison. The two original column means alone do not preserve that pairing information.
What if every difference is identical?
The difference standard deviation and standard error are zero, so the conventional t statistic is undefined. The data show a perfectly constant sample change, but a t distribution cannot be estimated from zero sample variability. The calculator reports this condition rather than dividing by zero.
Is dz a percent improvement?
No. It is a standardized mean difference measured in standard deviations of the paired changes. A percent change requires a meaningful baseline and a separate definition. Always report the raw-unit mean difference as well.
Can this establish that the treatment caused the change?
Not by itself. Causal interpretation depends on assignment, control of time trends and other threats, adherence, measurement, and study conduct. Pairing improves a comparison but does not manufacture randomization or remove every alternative explanation.
References
The formulas and interpretation follow the U.S. National Institute of Standards and Technology guidance for analysis of paired observations and the NIST Dataplot t test reference. These sources describe paired testing through the t distribution and distinguish it from independent-sample procedures. This educational calculator supports analysis and checking; it does not replace a study protocol or statistical review.