Is alternating A and B a randomized restriction?

Not by itself; a deterministic sequence is predictable rather than random.

Can block sizes vary?

Yes, if the scheme specifies how they are chosen and implemented.

Does restriction always improve power?

No; effects depend on the design, constraints, and available independent replication.

AI Glossary / Statistics and probability / Intermediate

Restricted randomization

Definition

Restricted randomization is random assignment carried out within predefined constraints rather than over every otherwise possible allocation. Restrictions may preserve balance within blocks, respect a hierarchy of experimental units, or accommodate factors that are difficult to change. The resulting design remains randomized, but the permitted assignments and their probabilities determine how the experiment should be analyzed.

What it means

The restriction says which assignments are allowed; randomization chooses among them. For example, a study may require equal numbers of two treatments within each site, or assign an oven temperature to a whole batch while randomizing smaller units within that batch.
This is different from choosing assignments by convenience or alternating treatments deterministically. A practical constraint can be handled by a valid randomized design, but it cannot be ignored when interpreting the results.

How it works

Identify the experimental units, the treatments, and the sources of variation that matter. Distinguish the unit receiving an independent assignment from the unit on which an outcome is measured; they may not be the same.
Specify the restriction before observing outcomes. In a randomized block design, similar units are grouped into blocks and treatments are randomized within each block. This can improve precision when the blocking factor explains outcome variation.
In a split-plot design, a hard-to-change factor is assigned to larger whole plots, while another factor is randomized to subplots inside them. The different assignment levels create different sources of variation and different information for estimating effects.
Other restrictions can control group sizes or covariate imbalance. Define the admissible allocation set or the sequential assignment algorithm, including any probabilities that depend on previous assignments. Not every restriction produces equally likely final allocations.
Generate and preserve the assignment using the specified random mechanism. Where foreknowledge could affect enrollment or behavior, protect allocation concealment. Small predictable blocks can create opportunities to anticipate later assignments if their structure is known.
Analyze according to the actual design. Blocking, clustering, and multiple randomization levels can change standard errors, degrees of freedom, and the relevant error terms. Treating every measured row as an independently assigned unit can overstate information.
A randomization or permutation analysis must reproduce the restrictions. Permuting labels freely across blocks or whole plots when the original assignment did not allow that can produce the wrong reference distribution.
Document deviations from the planned allocation and their causes. Restrictions should serve the study's practical and inferential objectives, not be added after outcomes are known to favor a desired comparison.

Three practical examples

AI experiments: a team balances two workflows within each site to avoid a site imbalance dominating the comparison.
Manufacturing: engineers randomize a hard-to-change temperature by batch and another setting within each batch.
Training evaluation: participants are randomized within predeclared experience groups to improve balance on an important baseline characteristic.

Think of it this way

Think of shuffling cards inside separate boxes instead of pouring all cards onto one table. The shuffle is still random, but a card cannot move to another box. Any analysis of possible deals must respect that rule.

Worked example

Suppose eight distinct units belong to two blocks of four. Exactly four units overall will receive treatment A and four treatment B.
Without a within-block restriction, choosing four of the eight for A gives 8 choose 4 = 70 allocations. Some allocations place all four A units in one block.
Now require two A and two B assignments within each block. Each block has 4 choose 2 = 6 allowed choices for A. Randomizing independently and uniformly within the two blocks yields 6 × 6 = 36 overall allocations, each with probability 1/36.
For any particular unit, three of its block's six allocations assign it A. Combining with all six allocations in the other block gives 18 of the 36 overall allocations, so its marginal probability of A remains one-half.
The restriction guarantees within-block counts, not equality of every other characteristic. A randomization test for this design should use the 36 allowed allocations, not the 70 allocations from the less restricted design.

Illustrative case study

Fictional case: Poplar Evaluation tests an AI workflow at two offices. The team randomizes equal numbers within each office, records the assignment procedure, and analyzes the comparison with office accounted for. It does not later shuffle users freely between offices in a permutation analysis, because those reallocations were never part of the study's actual design.

Common misconceptions and mistakes

Do not equate restricted randomization with nonrandom assignment. The selection can remain random within a constrained set.
Do not assume every measured observation supplies an independent treatment assignment. In a split-plot study, many measurements can share the same whole-plot assignment.
Restrictions do not guarantee balance on all measured or unmeasured characteristics. They guarantee only what the scheme enforces.
Avoid analyzing a constrained design as completely randomized merely because that analysis is simpler. The wrong error structure can produce misleading uncertainty estimates.

Business use

Restricted randomization can make experiments feasible while improving balance or respecting operational limits. Organizations should budget enough independent units at each assignment level and preserve a clear record of the scheme. The practical benefit comes from a design that operations can execute and analysts can evaluate correctly.

Check your understanding

Why does the worked design have 36 allocations?

Answer: Each of two blocks independently contributes six allowable assignments.

Why is each unit's probability still one-half?

Answer: It receives A in half the permitted allocations within its block.

Related terms

Randomized blocks design
Allocation concealment
Stratification
Permutation test
randomization
randomized block design
split-plot design
experimental unit
cluster randomization

Learn next

Learn next how blocking changes precision, how split-plot designs create distinct error terms, and how randomization-based inference uses the actual assignment mechanism.

Sources

Sources: NIST/SEMATECH, Randomized block designs, https://www.itl.nist.gov/div898/handbook/pri/section3/pri332.htm ; NIST/SEMATECH, Nested variation and restricted randomization, https://www.itl.nist.gov/div898/handbook/pri/section5/pri55.htm ; CONSORT 2025 explanation, restriction reporting item, https://pmc.ncbi.nlm.nih.gov/articles/PMC11995452 . Primary NIST pages and CONSORT search excerpts support design and reporting distinctions. The 36-allocation example was independently enumerated.

Previous: Restricted mean survival time Next: Rete