Matched pairs vs completely randomized design

By Jude Wallis · Published

A completely randomized design assigns every experimental unit to a treatment by chance in a single step. A matched pairs design first forms pairs of similar units, or uses each unit twice, then randomizes two treatments within each pair. Pairing removes unit-to-unit variation from the comparison.

AP Statistics: Unit 1 (topics 1.13 Experimental Design). The completely randomized design and the matched pairs design are both part of topic 1.13, Experimental Design, in Unit 1 of the Fall 2026 AP Statistics course. The analysis of paired data, as a mean difference, comes later in Unit 4.

Matched pairs vs completely randomized: the short answer

Both are designs for experiments, and both use random assignment. They differ in where the randomization happens.

In a completely randomized design, all the experimental units go into one pool and chance alone decides which treatment each one gets. Nothing is sorted or grouped first.

In a matched pairs design, there are exactly two treatments and the units are arranged in pairs before any randomization. Either two similar units form a pair and chance decides which member gets which treatment, or one unit receives both treatments and chance decides the order. The comparison is then made inside each pair.

The completely randomized design is simpler and works with any number of treatments. The matched pairs design gives a more precise comparison when the units differ a lot from one another on something that affects the response.

What a completely randomized design is

A completely randomized design has one randomization applied to all the units at once. To compare two shoe models with 12 runners, you could write the 12 names on slips, mix them, and draw 6 for model A, with the other 6 getting model B.

Random assignment spreads every other variable, such as running experience, weight, and age, across the two groups by the same chance process. The groups will not match exactly on those variables, but no variable is steered toward one treatment. A difference in the response then has two possible explanations: the treatments, or the chance in the assignment.

The cost is that all the natural variation between runners stays in the data. If runners differ from one another by far more than the shoes differ, the shoe effect is hard to see. How to write this design up for credit is covered in how to describe a completely randomized design.

What a matched pairs design is

A matched pairs design compares two treatments within pairs. It comes in two forms.

  • Two units per pair. Pair units that are alike on a variable expected to affect the response, such as two runners with nearly equal race times. Within each pair, flip a coin to decide who gets treatment A. The other gets treatment B.
  • One unit, both treatments. Each unit receives both treatments, and a coin flip decides which comes first. Randomizing the order keeps effects such as practice or fatigue from always favoring the same treatment.

Either way, the data become one difference per pair, and the analysis is run on those differences. The sample size is the number of pairs. A matched pairs design is a randomized block design in which each block holds two units and there are two treatments.

The differences side by side

FeatureCompletely randomizedMatched pairs
Grouping before randomizationNoneUnits paired, or each unit used twice
Where randomization happensAcross all units at onceWithin each pair (which member, or which order)
Number of treatmentsTwo or moreExactly two
What is comparedTwo or more group meansThe difference within each pair
Unit-to-unit variationStays in the comparisonLargely removed from the comparison
Data analysis for meansTwo-sample proceduresOne-sample procedures on the differences
Works best whenUnits are similar, or no good pairing variable existsUnits vary a lot on something related to the response

The design decides the analysis. Paired data are analyzed as one sample of differences, and data from two independently assigned groups are analyzed as two samples. The paired vs two-sample t-test page covers that step.

When to use which

Choose matched pairs when three things hold: there are two treatments, the units differ noticeably on a variable that affects the response, and you can either measure that variable beforehand or give both treatments to each unit. Each subject trying both treatments in random order, left hand versus right hand with the side chosen by chance, and twin studies are standard cases. A before-and-after measurement gives paired data too, but the order is fixed, so it is not a randomized matched pairs experiment.

Choose a completely randomized design when there are more than two treatments, when the units are already similar, or when no sensible pairing exists. It is also the safer choice when giving both treatments to one unit would let the first treatment carry over into the second, as with a drug that stays in the body.

Pairing has a price. With 6 pairs, the analysis of differences has 6−1=56 - 1 = 5 degrees of freedom, fewer than a two-sample analysis of the same 12 measurements would have. If the pairing variable has little to do with the response, you pay that price and gain nothing.

The classic mix-up and how to avoid it

The usual error is calling any study with two equal-sized groups a matched pairs design. Equal group sizes say nothing about pairing. A completely randomized design with 6 runners in each group has equal sizes and no pairs. A design is matched pairs only when each observation under one treatment is linked to one specific observation under the other, by a shared unit or by deliberate matching.

The second error is leaving out the randomization. "Each runner tries shoe A, then shoe B" is incomplete, because every runner would be more tired for shoe B. The order has to be chosen by chance for each runner.

To identify the design in a problem, look at how the treatments were assigned. If chance split one pool of units, it is completely randomized. If chance acted inside pairs, it is matched pairs.

Why pairing makes a small effect visible

Six people take a reaction-time test twice, once after caffeine and once without, with the order chosen by coin flip for each person. Times in milliseconds without caffeine: 310, 265, 342, 288, 251, 330. With caffeine, in the same person order: 298, 256, 327, 280, 240, 317. Find the mean and standard deviation of the differences, and compare that spread to the spread of the times themselves.

  1. Compute each difference, without minus with: 310−298=12310 - 298 = 12, 265−256=9265 - 256 = 9, 342−327=15342 - 327 = 15, 288−280=8288 - 280 = 8, 251−240=11251 - 240 = 11, 330−317=13330 - 317 = 13.

  2. Find the mean difference: xˉd=12+9+15+8+11+136=686≈11.33\bar{x}_d = \frac{12 + 9 + 15 + 8 + 11 + 13}{6} = \frac{68}{6} \approx 11.33 milliseconds.

  3. Find the standard deviation of the differences. The squared deviations from 11.33 sum to about 33.33, so sd=33.335≈2.58s_d = \sqrt{\frac{33.33}{5}} \approx 2.58 milliseconds.

  4. Find the standard deviation of each column of times. Without caffeine the mean is about 297.67 and s≈36.06s \approx 36.06. With caffeine the mean is about 286.33 and s≈34.17s \approx 34.17.

  5. Compare. People differ from one another by about 35 milliseconds, roughly three times the 11.33 millisecond caffeine effect. The differences vary by only 2.58 milliseconds, because each person serves as their own comparison.

  6. Put it in standard-error terms. For the paired differences, sd6=2.586≈1.05\frac{s_d}{\sqrt{6}} = \frac{2.58}{\sqrt{6}} \approx 1.05. If the same twelve numbers had come from two separate groups of six people, the standard error would be 36.0626+34.1726≈20.28\sqrt{\frac{36.06^2}{6} + \frac{34.17^2}{6}} \approx 20.28.

The mean difference is about 11.33 milliseconds with a standard deviation of 2.58. The times themselves have standard deviations near 35. Pairing gives a standard error of about 1.05, against about 20.28 if the same numbers were treated as two independent groups, so the caffeine effect that pairing shows clearly would be buried in person-to-person variation under a completely randomized design.

Counting the possible assignments

Twelve runners will compare two shoe models. Count the possible treatment assignments under (a) a completely randomized design with 6 runners per shoe, and (b) a matched pairs design in which the runners are first paired by race time into 6 pairs.

  1. For the completely randomized design, choose which 6 of the 12 runners get shoe A. The rest get shoe B. The count is (126)=924\binom{12}{6} = 924.

  2. For matched pairs, each pair has 2 possible assignments: the faster runner gets A, or the faster runner gets B.

  3. With 6 pairs assigned independently, the count is 26=642^6 = 64.

  4. Compare: 64924≈0.069\frac{64}{924} \approx 0.069. Every matched pairs assignment is also one of the 924, so pairing keeps about 7% of the completely randomized assignments.

  5. Describe the ones it keeps. In all 64, each shoe gets exactly one runner from every pair, so the two shoe groups are balanced on race time by construction.

There are 924 assignments under the completely randomized design and 64 under matched pairs. Pairing rules out every assignment where a shoe gets both runners from the same pair, which is how it keeps race time from favoring either shoe.

Frequently asked questions

Is matched pairs a type of randomized block design?

Yes. A matched pairs design is a randomized block design with two treatments in which each block is a pair of similar units, or a single unit measured under both treatments. The randomization happens inside each block, the same as in any block design.

Does a matched pairs design still need random assignment?

Yes. With two units per pair, chance decides which member gets which treatment. With one unit receiving both treatments, chance decides the order. Without that step, something like fatigue or the researcher's choices could consistently favor one treatment.

Can a matched pairs design compare three treatments?

No. A pair holds two observations, so matched pairs is limited to two treatments. To compare three or more treatments while controlling for a variable, use a randomized block design with blocks large enough to hold every treatment.