Can different people form a pair?

Yes, when matching is meaningful and incorporated into the design and analysis.

Are paired data always positively correlated?

No; the structural pairing matters even when observed correlation is weak or negative.

Does a nonsignificant paired test establish equivalence?

No; equivalence requires a suitable margin and analysis.

AI Glossary / Statistics and probability / Intermediate

Paired data

Definition

Paired data consist of observations linked in meaningful pairs, such as two measurements from the same person or results from two methods applied to the same item. The pairing is part of the study design or data-generating structure. Analysis must preserve that link and account for dependence between observations within a pair.

What it means

A before-and-after measurement from one employee belongs together because both describe the same employee. Likewise, outputs from two AI models on the same document can be paired by document.
Pairing can also arise from deliberately matched units, such as comparable stores assigned different procedures. Equal sample sizes alone do not create paired data. A valid pair represents a real connection established by the design, not two rows placed beside each other because that is convenient.

How it works

Identify the unit that defines each pair and retain a reliable identifier. Record which observation belongs to each condition, time, method, or matched unit. Check duplicates and missing values before calculation.
For quantitative outcomes, a common summary is the within-pair difference. Specify the direction, such as before minus after when positive numbers mean time saved. If ratios are more meaningful and values are positive, a log ratio may be appropriate, but its interpretation differs from an arithmetic difference.
The variation of paired differences depends on both individual variances and their covariance: Var(A - B) = Var(A) + Var(B) - 2 Cov(A,B). Positive within-pair association can reduce the variability of differences compared with an independent-samples calculation.
Choose analysis for the outcome and target. A paired t procedure addresses the mean difference under suitable assumptions. Other methods address signs, ranked differences, binary disagreements, or more complex repeated-measurement patterns. These methods do not all answer the same question.
The number of independent units is usually the number of independent pairs, not twice that number. Additional clustering, such as multiple documents from each customer, may require another level of adjustment.
Incomplete pairs need explicit handling. Restricting analysis to complete pairs can alter the target population and create bias if missingness is related to outcomes. Preserve the reasons for missing observations and use an appropriate analysis plan.
Pairing improves comparability but does not by itself establish causality. Before-and-after changes can reflect time trends, learning, regression to the mean, or other interventions.

Three practical examples

AI benchmarking: two models answer the same set of questions, allowing item-level comparison while retaining question identifiers.
Workflow evaluation: the same employees complete comparable tasks under two procedures, with order and learning effects considered.
Measurement comparison: two instruments measure each reference object, producing paired readings for studying differences and agreement.

Think of it this way

Think of comparing two shoes worn by the same walker. Keeping each walker's observations together can separate the shoe comparison from differences between walkers. Mixing one person's first shoe with another person's second shoe destroys that useful connection.

Worked example

Four fictional workers have task times before an update of 12, 20, 15, and 25 minutes. Their corresponding after-update times are 10, 17, 14, and 21 minutes.
Define each reduction as before minus after. The paired reductions are 2, 3, 1, and 4 minutes, with mean (2 + 3 + 1 + 4)/4 = 2.5 minutes.
Their deviations from 2.5 are -0.5, 0.5, -1.5, and 1.5. Squared deviations sum to 5, so the sample variance of reductions is 5/(4 - 1) = 5/3. The standard error of their mean, assuming independent workers and an appropriate common sampling model, is sqrt[(5/3)/4], approximately 0.6455 minutes.
The before mean is 18 and the after mean is 15.5, giving the same point difference of 2.5. However, the paired standard error comes from the reductions, not from treating all eight measurements as independent.
The four workers are illustrative, not a sufficient basis for a broad rollout claim. A causal interpretation would additionally need a design addressing changes other than the update.

Illustrative case study

Illustrative fictional case: Elm Documents compares two summarization systems on the same evaluation set. The analysts preserve a document identifier and calculate each system's score under a common rubric.
They discover that some documents failed in one system and were silently removed from an early analysis. The team restores the failures to the evaluation record, defines how failures count, and distinguishes matched score comparisons from overall completion rates. This fictional case shows why preserving pairs includes tracking missing or failed outcomes, not simply analyzing the rows that remain.

Common misconceptions and mistakes

Do not infer pairing from equal group sizes. Do not independently sort both columns before subtracting, because this can create artificial matches. Do not count two observations per person as two independent people.
Avoid assuming pairing removes all confounding. A high correlation between paired measurements also does not establish close agreement; one method may consistently read higher. Finally, a test of mean differences is not automatically a test of equivalent performance.

Business use

Paired data can make business comparisons more informative when items or people vary widely. They support transparent item-level evaluation and can improve precision when the pairing is useful. Reports should identify the pairing unit, explain outcome handling, show the distribution of differences, and distinguish average improvement from reliability across individual cases.

Check your understanding

What is the independent sample size in the example?

Answer: Four workers, assuming workers are independent.

Why keep the original row links?

Answer: They define each person's change and determine the variance of the paired differences.

Related terms

Covariance
Missing data
Matched pairs
repeated measures
paired t-test
within-subject design
agreement analysis
crossover design

Learn next

Next study inference for paired differences and the distinction between agreement and association. Then examine order effects, missing pairs, and clustered evaluations with multiple observations per participant or customer.

Sources

Frank Harrell, Glossary of Statistical Terms, paired data: https://hbiostat.org/glossary/ — primary terminology and within-subject comparison context. SciPy, scipy.stats.ttest_rel: https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.ttest_rel.html — official documentation for a related-samples mean comparison. Both pages were consulted; the four-worker calculations and fictional case are original.

Previous: Pair Confusion Matrix Next: Paired difference test