AI Glossary / Statistics and probability / Intermediate

Randomization, scientific

Definition

Scientific randomization uses an explicit chance mechanism to assign experimental units to treatments, determine run order or select units for observation. The mechanism supplies known selection or assignment probabilities rather than leaving choices to convenience or personal judgment. NIST's glossary describes this using random-number tables; properly implemented computer-generated randomization serves the same design purpose. The statistical meaning depends on what is randomized and on any restrictions imposed by the study design.

What it means

Randomization helps prevent assignment from systematically following characteristics that could distort the comparison. It does not guarantee perfectly balanced groups in one experiment, eliminate measurement error or repair missing data. Random assignment and random sampling also have different purposes. Assignment supports treatment comparisons within the study under relevant assumptions, while sampling concerns how observed units relate to a population. A study can have one form of randomization without the other.

How it works

Define the eligible units, treatments or sampling frame before drawing assignments. Select a design, such as unrestricted allocation, fixed group sizes or randomization within blocks. Use a suitable random mechanism to select among permitted outcomes according to the design's probabilities. Preserve the generated allocation and document departures. For experiments where advance knowledge could influence enrollment, conceal upcoming assignments appropriately. Analyze the data using methods that respect the actual randomization, including grouping or clustering restrictions.

Three practical examples

Example 1: A production experiment randomly assigns comparable pieces to two processing methods. Example 2: A laboratory randomizes run order to reduce systematic alignment between treatment and instrument drift. Example 3: A survey selects records using a specified probability-sampling design. In each example, the chance mechanism acts on a particular object. Randomly ordering already biased observations does not transform their original selection into a representative population sample.

Think of it this way

Think of randomization as an impartial allocation rule applied before outcomes are known. The rule can be constrained to keep group sizes fixed or to respect blocks, while still choosing unpredictably among allowed assignments. Its value comes from the specified chance process and the reasoning that process permits. It is not equivalent to an investigator making choices that merely look mixed or balanced.

Worked example

Suppose six experimental units are assigned to two treatments with exactly three units per group. There are 6 choose 3 = 20 possible treatment-A groups. Selecting uniformly among them gives every unit probability 3/6 = 0.50 of receiving A. To illustrate the randomization logic, assume the sharp null of no treatment effect and fixed outcomes 10, 11, 12, 13, 14 and 15. Assigning the first three values to A yields a mean difference A minus B of 11 - 14 = -3; reversing the groups yields +3. Across all 20 assignments, the mean difference averages zero because every assignment has an equally probable complementary assignment. This expectation does not imply that any one realized assignment produces zero difference.

Illustrative case study

Fictional case study: Cedar Materials tests two coatings. Its initial plan assigns the first production batch to one coating and the next batch to the other, which would confound coating with production order. The team instead defines a randomized allocation within suitable batches and records the treatment sequence. It separately checks measurement consistency. This fictional case illustrates design improvement, not a claim that randomization guarantees a successful experiment or removes every possible source of bias.

Common misconceptions and mistakes

Common mistakes include calling alternating assignments random, assuming visibly balanced groups prove a random process, and changing assignments after seeing inconvenient characteristics. A reproducible computer sequence can support randomization when generated and used appropriately; reproducibility does not make it a convenience allocation. Randomization can balance unknown factors in expectation but cannot ensure exact balance in a small study. Ignoring restricted allocation in the analysis can also give incorrect uncertainty estimates.

Business use

Business use: Scientific randomization strengthens experiments comparing products, workflows and process settings. Specify the unit of assignment carefully: customers, sites, machines and individual observations are not interchangeable. Decide whether blocking or clustering is needed before allocation. Keep the procedure auditable and distinguish assignment failures from outcome exclusions. When presenting results, explain the population studied and the assumptions required for causal or population-level conclusions rather than attributing every desirable property to randomization alone.

Check your understanding

In the six-unit example, is every unit independently assigned to A?

Answer: No. Fixed group sizes make assignments dependent even though each unit has probability one-half of receiving A.

Why does the average null difference equal zero?

Answer: Complementary assignments produce opposite differences and are equally probable under the stated mechanism.

Related terms

Random assignment allocates treatments

Learn next

Enumerate the allowed assignments in a small fixed-size experiment and calculate the distribution of a treatment contrast under a sharp null. Compare it with unrestricted coin-flip allocation. Then add blocks and examine how the permitted assignments change. Document the experimental unit and distinguish the roles of randomization, concealment, blinding and population sampling.

Sources

Sources: NIST/SEMATECH e-Handbook glossary, randomization, scientific: https://www.itl.nist.gov/div898/handbook/glossary.htm ; NIST, The data are randomly sampled from the process: https://itl.nist.gov/div898/handbook/pmd/section2/pmd215.htm . These primary institutional sources explain randomisation and its design role. The exact 20-assignment calculation and fictional coating experiment are original illustrations.

Previous: Random-walk Metropolis Next: Randomized blocks design