Why Pairing Depends on the Design
Two groups can look alike without being paired. They might have the same number of observations, similar ages, or similar average outcomes. None of those facts, on its own, creates a link between a particular observation in one group and a particular observation in the other. Pairing is a feature of how the observations were collected and connected, not a label added because the data look comparable.
This distinction matters because a paired analysis uses the link: it studies the difference within each pair. When there is no defensible link, those within-pair differences do not represent the design. A procedure chosen from the appearance of a data table, rather than from how the data were generated, can give an inappropriate account of the evidence.
In a before-and-after design, a person's two measurements form a pair because both are from that person. In a matched design, two different people or other units form a pair because the study deliberately linked them, for example by matching on specified baseline characteristics before outcomes were observed. The units in a pair need not be identical; what matters is that the relationship was part of the design and is recorded.
The matching rule should be meaningful for the study question and applied without using the outcome being compared to manufacture a favorable pairing. A study should retain a way to identify the two members of every pair, such as participant IDs and a pair ID. Without that link, a later analyst may be unable to distinguish genuine pairs from an arbitrary arrangement.
What Does Not Create a Pair?
Similar characteristics can be useful for describing or designing a comparison, but similarity by itself is not a one-to-one link. For example, two groups might contain similar proportions of people in each age category. That is a group-level resemblance: it does not say which person in the first group is paired with which person in the second.
Nor does equal sample size establish pairing. If two samples each contain 30 people, the observations might still be unrelated individuals with no planned correspondence. Writing the first observation from one sample beside the first observation from the other is only a formatting choice unless the study gives those rows a real meaning.
A useful first question is: Could the study explain, using its protocol or records, why these two particular observations belong together? If the answer is yes, and the link was created independently of the observed outcomes, pairing may be appropriate. If the only explanation is that the observations have similar values, appear in the same row, or were sorted into a convenient order, the link is not a design feature.
Worked Example: The Same People Measured Twice
In a hypothetical pilot, five volunteers have their resting pulse measured before and after a brief training program. The recorded values, in beats per minute, are:
| Volunteer | Before | After |
|---|---|---|
| A | 72 | 70 |
| B | 80 | 81 |
| C | 76 | 73 |
| D | 68 | 68 |
| E | 84 | 80 |
Is this paired? Yes. Each before-and-after combination belongs to the same volunteer. The volunteer ID supplies the link, so the pairs are not chosen by comparing pulse values or by putting similar values together.
If the difference is defined as after minus before, the five pair differences are \(-2, 1, -3, 0,\) and \(-4\) beats per minute. Their average is
This arithmetic describes the average within-volunteer change in this hypothetical sample. Its sign follows the stated direction: a negative difference means the after measurement is lower on average than the before measurement. The important point for identifying the design is that each difference uses two measurements on one identified person.
Worked Example: Similar Groups Without Pairing
In another hypothetical study, separate volunteers are recruited for two conditions. The study did not match participants or assign pair IDs. The ages and outcome scores are shown below; the scores are in points.
| Condition A age | Condition A score | Condition B age | Condition B score |
|---|---|---|---|
| 20 | 8 | 21 | 7 |
| 22 | 10 | 23 | 9 |
| 24 | 9 | 25 | 8 |
| 26 | 11 | 27 | 10 |
Are these genuine pairs? No. The groups have four people each, and their age ranges are similar, but the study gives no reason to link the first person in A to the first person in B, or any other specific combination. The table’s row arrangement is not evidence of matching.
The group means are \(\bar{x}_A=(8+10+9+11)/4=9.5\) points and \(\bar{x}_B=(7+9+8+10)/4=8.5\) points. Their observed difference, A minus B, is \(9.5-8.5=1.0\) point. That difference of group means is meaningful as a comparison of these samples, but it does not make the individual observations paired.
If someone treats the displayed rows as pairs, the apparent differences are \(8-7=1\), \(10-9=1\), \(9-8=1\), and \(11-10=1\) point. Those calculations depend entirely on an unsupported row-by-row assignment. A different arbitrary ordering would produce different individual differences, even though the two groups have not changed. The group means do not change with row order; the supposed pair differences do. Without a design-based link, the latter are not justified.
Matching Distinct Units Can Be Genuine
Pairs do not have to consist of repeated measurements on the same person. A study can deliberately match two distinct units, then retain the pairing throughout data collection and analysis. For instance, researchers might form pairs of participants with similar baseline characteristics before assigning one member of each pair to each condition. In that case, the pair is a real design unit even though its members are different people.
Matching must be distinguished from simply recruiting two groups that happen to resemble one another. A protocol that says “each treated participant was matched with one comparison participant on specified baseline characteristics” establishes a design link only if the study actually records which two participants were matched. Similar group averages do not supply those individual links after the fact.
Worked Example: A Deliberately Matched Study
Suppose a hypothetical study forms four pairs of participants with similar baseline scores. Within each pair, one participant is assigned to a program and the other to usual practice. The outcome is improvement, measured in points. The pair IDs are set before the outcomes are measured.
| Pair | Program improvement | Usual-practice improvement |
|---|---|---|
| 1 | 6 | 4 |
| 2 | 3 | 1 |
| 3 | 5 | 4 |
| 4 | 2 | 3 |
Is this paired? Yes. Each row represents a pair created by the study design, not a row created by sorting outcome scores. The match is recorded, and the two assignments within each pair are defined by the protocol.
Taking program minus usual practice, the within-pair differences are \(6-4=2\), \(3-1=2\), \(5-4=1\), and \(2-3=-1\) points. Their mean is
This is the average program-minus-usual-practice difference across the four designed pairs. The calculation preserves the study’s intended correspondence. It does not, by itself, establish that the program caused the difference; a causal conclusion depends on the broader design and its assumptions.
Why the Link Can Matter for Variation
A genuine pairing can matter because two observations in the same pair may be related. For example, two measurements on the same person often share characteristics of that person. The difference within a pair can remove some variation associated with what the two observations share. This is one reason a paired design can be informative.
The variance identity makes the point precise. Let \(X\) and \(Y\) denote the two outcomes for a randomly selected pair, with finite variances. The variance of their difference is
To justify the identity, start from the definition of variance and expand:
When \(X\) and \(Y\) tend to be high or low together, their covariance is positive, and the subtraction of \(2\operatorname{Cov}(X,Y)\) reduces the variance of their difference compared with the sum of their separate variances. If their covariance is negative, the difference can instead have greater variance. Thus genuine pairing does not guarantee a more precise result; its effect depends on the within-pair relationship. The formula also explains why an invented pairing is not harmless: it substitutes an arbitrary relationship for the one, if any, created by the design.
How to Verify Pairing in Practice
When a study description is brief, look for evidence about the data-collection process rather than relying on the layout of the results. Ask how observations were recruited, measured, assigned, matched, and recorded. A genuine pair should have an identifiable basis that can be stated in the context of the study.
Determine whether the two measurements could come from the same unit at different times, or from two distinct units deliberately linked by the design.
Look for repeated measurement on the same identified unit, a matching rule, or a recorded pair ID. A similar age range or equal sample size is not a link.
The pairing should come from the study protocol or from information determined independently of the outcomes being compared, not from arranging outcome values to create appealing differences.
If the design establishes pairs, keep each pair’s observations together. If it does not, do not create pair differences just by lining up rows or sorting values.
Common Mistakes
- “The groups have the same sample size, so they are paired.” Equal counts make one-to-one row alignment possible, not meaningful. Pairing requires a design-based correspondence.
- “The groups have similar ages, so the first person in each group must be a pair.” Similar age distributions are group-level similarities. A specific person-to-person match must be established and recorded.
- “The data are already in two columns, so each row is a pair.” A table can display unrelated observations on the same row for convenience. Check what the rows mean in the study protocol.
- “We can sort both groups by outcome and pair matching ranks.” Sorting by the outcome creates a data-dependent link rather than recovering a design feature. The resulting within-pair differences do not represent the original study design.
- “Matched means identical.” Matched units are linked according to a design rule; they need not be identical on every characteristic. The link and its rationale, not perfect similarity, define the pair.
Check Your Understanding
For each question, decide what information in the study design supports or rules out pairing.
- A clinic measures each of 18 patients before and after a change in their daily routine. What makes these measurements paired?
- Two groups each contain 25 people, and their average ages are nearly equal. The participants were recruited separately, with no matching. Are the observations paired? Explain.
- A study matches each participant receiving a new program with a comparison participant using baseline score, assigns pair IDs, and records both outcomes. What makes these genuine pairs?
- Two unrelated samples are sorted by outcome and placed in adjacent rows. Why is calculating row-by-row differences not justified by that arrangement?
- For paired outcomes \(X\) and \(Y\), what role does \(\operatorname{Cov}(X,Y)\) play in \(\operatorname{Var}(X-Y)\), and why does that not imply pairing always lowers variation?