Statistics › Section J · Hypothesis testing with two samples › Topic J.5
Matched or paired samples
Measure the same students before and after a program, the same vials in two labs, the same twins under two conditions - and the design, not the data, changes the test. Paired data collapses into differences, and a one-sample test on those differences is both simpler and far more sensitive.
When the design pairs
Some comparisons pair their units by design: before-and-after measurements on the same people, two labs testing split samples from the same vials, matched volunteers given different treatments, both eyes or both arms of one patient. In every case each unit appears in both groups, and the pair - not the person - becomes the unit of analysis. The question "do the two conditions differ?" becomes "do the differences differ from zero?", and an ordinary one-sample test takes over.
The method: test the differences
For each pair, compute the difference \(d\) - after minus before, treatment minus control, lab A minus lab B, consistently signed. Then run a one-sample t test on the differences:
with \(n - 1\) degrees of freedom, \(n\) being the number of PAIRS. The null \(H_0: \mu_d = 0\) says the conditions do not differ on average; the alternative carries the question's direction. Every piece of the machinery is the one-sample chapter's - the design just hands it better data.
Why pairing is powerful
Pairing removes person-to-person variation before it can enter the analysis. Students arrive reading at wildly different levels - some at grade 3, some at grade 9 - and that spread swamps a two-group comparison. But each student's improvement measures their own change, and the strong readers' and weak readers' baselines cancel inside their own differences. The differences carry only what the comparison needs: the within-pair change. This is the logic of blocking - compare like with like, so differences between units cannot pose as an effect - and it is why paired designs are worth the logistical trouble whenever they are possible.
Conditions: the differences must behave
The conditions are the one-sample t's, applied to the differences. Pairs are randomly sampled, or treatments randomly assigned within pairs (which side of the split gets which lab, which order a student receives). Pairs are independent of one another - one student's change cannot influence another's. And the differences look roughly symmetric for small \(n\); by \(n \ge 30\) the CLT relaxes the requirement. Note what is NOT required: the raw before and after scores need not be symmetric or normal - only the differences do, and differences are often far better behaved than their ingredients.
Worked example: a reading program
Twelve students take a reading test before and after a month-long program. The improvements \(d\) average \(\bar{d} = 6.5\) points with \(s_d = 9.0\). Did the program help? 1. \(H_0: \mu_d = 0\); \(H_a: \mu_d > 0\) - one-sided, improvement is the claim. 2. Conditions: students sampled at random; the differences show no strong skew (reported). 3. \(\text{SE} = 9.0/\sqrt{12} \approx 2.6\); \(t = 6.5/2.6 = 2.5\) with 11 df; one-sided \(p \approx 0.015\). 4. \(p \le 0.05\): reject - convincing evidence of improvement, averaging 6.5 points per student.
Worked example: the interval, and the pairing dividend
The 95% interval for the mean improvement: \(6.5 \pm 2.201 \times 2.6 \approx 6.5 \pm 5.7\), or \((0.8, 12.2)\) points - excluding zero, consistent with the test (even the two-sided p, about 0.029, is below 0.05), and sizing the gain: plausibly between about 1 and 12 points. Now the contrast that shows what pairing bought: the raw test scores vary enormously between students (SDs around 15), and had the same data been analysed as two independent groups, that between-student spread would drive the SE to about \(15\sqrt{2/12} \approx 6.1\), giving \(t \approx 1.1\) and no significance at all. The paired analysis finds what the unpaired one cannot, because the differences never carried the baseline variation.
Paired or independent? The design decides
Choosing between this test and the two-sample t is not a statistical judgement but a design reading. Same units measured twice, or matched into pairs: paired test on differences. Different, unrelated people in each group: two-sample test - and forcing a paired analysis onto unpaired data is impossible, because there are no pairs to difference. The mistake runs the other way: analysing genuinely paired data as if independent, discarding the pairing and often the significance with it. Read the design before choosing the test, exactly as the hypotheses topic read the wording before writing \(H_a\).
Common mistakes to avoid
First - the classic - running a two-sample test on paired data: it dilutes the evidence with between-person variation that the design worked to remove. Second, losing the signs: differences must carry direction; |d| destroys the hypothesis. Third, averaging the before and after columns separately and comparing the averages - the pairing information lives in the pairs, not the means. Fourth, wrong \(n\): the degrees of freedom count PAIRS minus one - 12 students give 11, not 23. Fifth, checking normality on the raw scores instead of the differences. Sixth, claiming causation from a before-after study without randomisation: pairing removes person-to-person noise, but other time trends (maturation, season, practice) still confound unrandomised before-after designs.
Key takeaways
Paired designs - repeated measurements, split samples, matched units - make the pair the unit: compute each pair's difference and run a one-sample \(t\) on the differences, \(t = \bar{d}/(s_d/\sqrt{n})\) with \(n - 1\) degrees of freedom. Pairing removes person-to-person variation, which is why it is worth logistical trouble whenever possible - the reading program's improvement (\(t = 2.5\), 11 df, \(p \approx 0.015\), interval 0.8 to 12.2 points) would drown in a two-group analysis. Conditions: random pairs, independence between pairs, roughly symmetric differences at small \(n\). The design, read honestly, chooses the test; and before-after without randomisation still cannot claim causation on its own.
Flashcards
Flashcard
Each unit appears in both conditions
Test your understanding
Question 1 of 5
Blood pressure is measured on 25 patients before and after a drug. Which test fits?
Question 2 of 5
With d̄ = 6.5, s_d = 9.0 and n = 12 pairs, the paired test statistic is about:
Question 3 of 5
For 12 paired differences, the degrees of freedom are:
Question 4 of 5
The main statistical advantage of a paired design is that it:
Question 5 of 5
The 95% interval for the mean improvement is (0.8, 12.2) points. At α = 0.05, the two-sided test of no improvement will:
