Does it guarantee every variable is balanced?

No. It targets specified strata and depends on the allocation rule.

Can strata be defined after outcomes arrive?

That is not baseline stratified randomization.

Is it necessary for every experiment?

No. Its usefulness depends on the design and the importance of the chosen baseline factors.

AI Glossary / Statistics and probability / Intermediate

Stratified randomization

Definition

Stratified randomization performs random assignment separately within predefined subgroups, or strata, so that important baseline characteristics can be balanced across comparison groups. In an experiment, participants are classified before assignment and then randomized within their strata. Related language is sometimes used for random sampling within strata; that is a selection procedure rather than treatment assignment. This entry focuses on experimental allocation and makes the distinction explicit.

What it means

Suppose a business compares two training methods and expects prior experience to affect the outcome. Randomizing separately among experienced and inexperienced staff can keep that characteristic more evenly distributed across methods.

Stratification identifies where randomization occurs; the within-stratum procedure determines the degree of numerical balance. Independent coin flips within each stratum do not guarantee equal counts. Restricted procedures, such as allocating a fixed number to each condition within a completed group, can enforce specified counts while preserving a random element. The design should state both the strata and the allocation rule rather than promising balance from the label alone.

How it works

Choose a small, justified set of baseline variables before outcomes are observed. Define categories and rules for missing values, then form non-overlapping strata. Avoid creating so many combinations that many strata contain too few participants for the intended allocation procedure.

Generate assignments separately within each stratum using the planned random mechanism. For sequential enrollment, a concealed allocation system helps prevent recruiters from predicting or influencing the next assignment. If blocks are used, their sizes and handling of incomplete blocks affect balance and predictability.

Record the stratum, assignment, and relevant deviations for every participant. Prespecify an analysis that accounts for the design and addresses the intended treatment comparison. If effects may differ by subgroup, plan how those questions will be evaluated rather than assuming stratification itself demonstrates different effects.

Assess the resulting groups without expecting all characteristics to match exactly. Balance is targeted on the chosen variables and remains subject to the allocation rules and available participants. Random assignment supports causal comparison under additional conditions; it does not automatically solve nonadherence, missing outcomes, interference, or measurement problems.

Three practical examples

1. A company randomizes two training methods separately within novice and experienced staff groups to prevent a large experience imbalance.

2. An AI interface experiment randomizes within device categories because screen size may strongly affect task performance.

3. A multi-site service evaluation uses separate allocation schedules by site so each location contributes participants to the comparison conditions.

Think of it this way

Think of dealing cards from two labeled packs instead of mixing everything first. If each pack represents an important background group, dealing a planned number from each into both hands controls that feature of the hands. Random dealing still determines which particular cards go where.

Worked example

Consider an original experiment with eight participants: four novices and four experienced users. Four participants must receive method A and four method B. Without stratification, uniformly choosing the four A participants gives C(8,4) = 70 possible assignments.

To obtain exactly two novices and two experienced users in A, choose two of the four novices and two of the four experienced users. There are C(4,2) × C(4,2) = 6 × 6 = 36 such assignments. Under the unstratified fixed-total design, the probability of that exact experience balance is 36 / 70, about 51.4%.

Now use a stratified design that uniformly assigns exactly two people to A within each experience stratum. It selects among those 36 balanced assignments, and both A and B necessarily contain two novices and two experienced users. Each individual still has assignment probability 2 / 4 = 0.5 within their stratum.

This toy enforces balance on experience and treatment counts, not on every other characteristic. It also uses a fixed enrolled group, not an uncontrolled sequence of independent coin flips. No outcome effect or statistical power follows from these allocation counts alone.

Illustrative case study

In a fictional company, Spruce Learning plans to compare an AI-supported training format with its existing format. Prior pilot results suggest experience level matters, so the evaluation team defines two experience strata before enrollment and randomizes within each. It preserves the assignment records and plans an analysis that includes the stratification factor. When one format performs better overall, the report separates that estimated comparison from exploratory differences between experience groups. The allocation design improves a planned balance without turning every subgroup pattern into a confirmed finding.

Common misconceptions and mistakes

Stratified randomization is not the same as stratified random sampling. Sampling determines who enters a study; experimental randomization determines which condition enrolled units receive.

Exact balance requires a suitable within-stratum restriction. Merely using independent random assignments in each subgroup does not force equal numbers.

More stratification factors are not always better. Many small cells can make implementation difficult and leave incomplete allocations. Stratification on a variable measured after treatment can also change the question and introduce bias. Finally, balanced baseline groups do not eliminate the need for appropriate outcome measurement, follow-up, and analysis.

Business use

Businesses can use stratified randomization in experiments where a few known baseline features strongly influence outcomes or need representation in both conditions. It helps make comparisons more interpretable and can improve precision when used appropriately.

Plan the variables, allocation mechanism, and analysis together. In online experiments, confirm that the unit of assignment matches the exposure process and avoid assigning the same person inconsistently across devices or accounts. Preserve the original design information so reviewers can distinguish planned allocation from later adjustments. Operational convenience should not silently replace the random mechanism.

Check your understanding

Answer: An experiment assigns each person by a fair coin within two strata.

Are equal treatment counts guaranteed in each?

Answer: No. The assignments are random, but exact count balance requires an additional restriction such as fixed allocation within a completed group.

Related terms

Stratified sampling
Blocking
Allocation concealment
Treatment effect
Subgroup analysis
Random assignment
baseline covariate
experimental design

Learn next

Study restricted randomization and analysis of stratified experiments. Then distinguish baseline balance from causal identification and subgroup-effect estimation.

Restricted randomization

Sources

ICH E9, Statistical Principles for Clinical Trials, section 2.3.2 on randomization: https://database.ich.org/sites/default/files/E9_Guideline.pdf . This primary methodological guidance supports baseline stratification and allocation principles; the entry is not clinical or regulatory advice. The supplied title has sampling and allocation usages, explicitly distinguished here. The combinatorial example and fictional business case are original.

Previous: Stratification Next: Stratified resampling