The Central Idea
An experiment compares outcomes under different treatments. For that comparison to support a causal conclusion, the treatment groups should not differ systematically before treatment begins.
Random assignment gives each eligible experimental unit a known chance of receiving each treatment. The assignments are produced by a chance mechanism rather than by preference, judgment, convenience, prognosis, or the order in which units arrive.
\[ \boxed{ \text{chance assignment} \;\longrightarrow\; \text{comparable groups on average} \;\longrightarrow\; \text{credible causal comparison} } \]Random assignment does not make the groups identical. It makes systematic selection into treatments less plausible and turns remaining group differences into chance differences that statistical methods can evaluate.
Why Voluntary or Judgment-Based Assignment Can Fail
Suppose students may choose between a new tutoring program and the usual support program. Students who choose the new program may be more motivated, have more available time, or have greater family support. If they later earn higher scores, the difference might reflect the tutoring, the preexisting differences, or both.
Let (T) denote treatment choice, (M) motivation, and (Y) the outcome. A possible structure is
\[ M\longrightarrow T \qquad\text{and}\qquad M\longrightarrow Y. \]Motivation confounds the treatment comparison. Random assignment breaks the deliberate connection between measured or unmeasured characteristics and the treatment received.
The Randomization Mechanism
A valid randomization mechanism has a probability structure determined in advance. Appropriate methods include:
- drawing thoroughly mixed labels from a container;
- using a table of random digits;
- using a verified random-number generator;
- randomly shuffling a prepared list of treatment labels; or
- using specialized experimental-assignment software.
The physical or digital tool is not the essential feature. The essential feature is that the assignment follows the stated chance rule and cannot be altered after a unit's characteristics are known.
Example: Assigning Ten Units to Two Treatments
Suppose ten experimental units, labeled (1) through (10), must be divided equally between treatment (A) and treatment (B).
- Write five (A) labels and five (B) labels.
- Thoroughly shuffle the ten labels.
- Assign one label to each numbered unit without looking ahead.
- Record the complete assignment before treatment begins.
One possible assignment is
\[ \begin{array}{c|cccccccccc} \text{unit} & 1&2&3&4&5&6&7&8&9&10\\ \hline \text{treatment} & A&B&A&B&B&A&A&B&A&B \end{array} \]Many other assignments are possible. What matters is that the observed unit characteristics did not determine which valid arrangement occurred.
Complete Random Assignment
In complete random assignment, the numbers of units assigned to the treatment groups are fixed in advance, and every allocation with those group sizes is equally likely.
If (20) units are divided into two groups of (10), the number of possible ways to select the units for treatment (A) is
\[ \binom{20}{10}=184{,}756. \]Once the ten units for (A) are chosen, the remaining ten receive (B). Under complete random assignment, each of these (184{,}756) allocations has the same probability.
Fixed group sizes are often useful because they ensure that one treatment group does not become much smaller than another merely by chance.
Independent Assignment
Another design assigns each unit independently, such as by flipping a fair coin:
\[ P(A)=\frac12, \qquad P(B)=\frac12. \]With (n) units, the number assigned to (A) is then random. Its expected value is (n/2), but the realized groups need not be equal. For (20) units, assignments such as (9) and (11), or even (7) and (13), are possible.
Independent assignment is genuinely random, but complete random assignment is often preferred when balanced group sizes improve precision or simplify operations.
Unequal Allocation
Random assignment does not require equal group sizes. A study might assign participants in a (2:1) ratio, giving two participants the new treatment for every one assigned to control.
Unequal allocation may be used to gather more safety information about a new treatment, reduce costs, increase willingness to participate, or support several treatment groups sharing one control group. The ratio should be chosen before assignment and justified by the study's goals.
Equal allocation often provides the most statistical information for a fixed total sample when treatment costs and outcome variability are similar, but it is not a defining requirement of randomization.
Balance in Expectation, Not Perfect Balance
Random assignment balances characteristics in expectation. Across many repetitions of the assignment process, no treatment group is systematically favored. In one realized experiment, however, the groups may differ.
For example, a randomized study may happen to place more older participants in one group. That does not prove that randomization failed. Chance processes naturally produce variation.
Larger groups tend to reduce severe imbalances, but even large randomized experiments do not guarantee equality on every characteristic. Randomization provides a probability-based design, not a promise of identical groups.
Measured and Unmeasured Characteristics
Statistical adjustment can account for characteristics that researchers have measured well. Random assignment has a broader purpose: it gives both measured and unmeasured pre-treatment characteristics an opportunity to distribute without systematic connection to treatment.
Researchers may not know every variable that affects an outcome. A chance mechanism does not need to identify those variables one by one before it can protect the treatment comparison from deliberate selection on them.
Randomization therefore addresses hidden confounding more directly than merely comparing groups on a short list of recorded characteristics.
Random Assignment Is Not Random Sampling
The two ideas both use chance, but they solve different problems.
| Feature | Random Sampling | Random Assignment |
|---|---|---|
| Main question | Who enters the sample? | Which treatment does each unit receive? |
| Chance operates between | Population and sample | Experimental units and treatment groups |
| Primary benefit | Supports population generalization | Supports causal comparison |
| Does it make a study experimental? | No | It is a defining feature of a randomized experiment |
A randomized experiment can use volunteers rather than a random sample. Its treatment comparison may have strong internal causal validity even though generalization beyond the volunteers requires caution.
Two Chance Steps in One Study
A study can use both random sampling and random assignment:
\[ \text{population} \xrightarrow{\text{random sample}} \text{study units} \xrightarrow{\text{random assignment}} \text{treatment groups}. \]The first step supports claims about the population. The second supports claims about treatment effects. Strong execution of one step does not replace the other.
Restricted Forms of Random Assignment
Simple randomization is not always the most informative design. Researchers may first organize units into meaningful groups and then randomize within those groups.
- Blocked assignment randomizes separately within groups sharing an important characteristic.
- Matched-pair assignment forms pairs of similar units and randomly assigns the two treatments within each pair.
- Cluster assignment randomly assigns intact groups, such as classrooms or clinics, rather than individuals.
These designs still use random assignment. The randomization occurs within a structure chosen to improve balance, respect logistics, or match the true unit at which treatment can be delivered.
The Unit of Assignment Must Be Clear
Suppose twelve schools are randomly assigned to two curricula, with six schools per curriculum. Students are measured, but schools receive the assignment. This is a cluster-randomized experiment.
Students within the same school share teachers, schedules, policies, and local conditions. Their outcomes are therefore not independent in the same way as outcomes from students assigned individually.
The analysis and claimed amount of independent experimental evidence must reflect the school-level assignment. A large number of student measurements cannot replace a small number of independently randomized schools.
Allocation Concealment
Allocation concealment prevents the next treatment assignment from being known before an experimental unit is enrolled. It protects the randomization process from conscious or unconscious manipulation.
For example, if a recruiter can predict that the next participant will receive the new treatment, the recruiter might delay one participant and enroll another. The assignment sequence may have been random originally, but predictable or exposed assignments allow selection bias to enter.
Allocation concealment occurs before and during assignment. Blinding usually concerns whether treatment identity is known after assignment. They protect different stages of the experiment.
Documenting Digital Randomization
When software performs the assignment, the procedure should be reproducible and auditable. Researchers should document:
- the list of eligible units before assignment;
- the treatment ratio and any fixed group sizes;
- blocking, matching, or clustering rules;
- the software or algorithm used;
- the random seed when appropriate;
- the generated assignment record; and
- who could access or alter the sequence.
A recorded seed allows the same pseudo-random sequence to be recreated. The seed should be fixed or generated according to the study protocol before the assignments are examined, not searched repeatedly until a favorable balance appears.
Methods That Are Not Reliably Random
The following methods may look neutral but can create predictable or biased groups:
- letting participants choose their treatments;
- letting a researcher decide case by case;
- assigning the first half to one treatment and the second half to another;
- alternating treatments when the next assignment is visible;
- using day of the week, room number, or record number without a justified chance process;
- assigning by convenience or availability; or
- rerandomizing until the observed arrangement looks favorable without a prespecified rule.
Predictability is dangerous because anyone controlling enrollment can influence which units reach which treatment. Random assignment requires more than an orderly rule; it requires a genuine chance mechanism.
Preserving the Assigned Comparison
Random assignment occurs at the beginning, but events after assignment can weaken the comparison. Participants may not follow the assigned treatment, may switch treatments, or may leave the study.
One important principle is to preserve the original assignment groups in the main analysis. Comparing only those who completed the treatment they preferred can recreate selection bias. The original assignment remains the chance-created basis for comparison.
Researchers may also study actual treatment received, but that analysis answers a different question and requires additional assumptions.
Randomization Creates a Reference for Chance
Even if treatments have no effect, random assignment will usually produce some difference between group outcomes. The assignment mechanism tells us which other groupings could have occurred by chance.
Statistical inference can compare the observed treatment difference with the differences that would arise under alternative valid assignments. If the observed result is unusually large relative to chance assignment variation, it provides evidence against a no-effect explanation.
Thus random assignment serves two connected roles: it protects the comparison from systematic selection and supplies a probability model for evaluating chance variation.
Example: Evaluate an Assignment Procedure
A clinic wants to compare two appointment-reminder systems. Patients arriving on Monday and Tuesday are assigned to system (A); patients arriving on Wednesday through Friday are assigned to system (B).
Is the Assignment Random?
No. Day of arrival determines treatment, and the day is not generated as a chance assignment for each patient.
What Could Go Wrong?
Patients attending early and late in the week may differ in work schedules, appointment types, urgency, staffing, or access to transportation. Those differences could affect whether patients respond to reminders.
How Could the Design Be Improved?
After each patient is enrolled, use a concealed randomization procedure to assign the reminder system. If day of week strongly affects the outcome, the clinic could randomize between systems separately within each day.
Ethics and Random Assignment
Random assignment is appropriate only when the treatment comparison is ethically permissible. Researchers should have genuine uncertainty about which treatment is preferable, disclose relevant risks, obtain required consent, and avoid withholding established effective care without justification.
Chance does not make an unethical treatment ethical. Randomization determines assignment among acceptable study conditions; ethical review determines whether those conditions may be studied at all.
A Random-Assignment Checklist
- Define eligibility. Which units enter the assignment pool?
- Identify the assignment unit. Individuals, pairs, or clusters?
- Specify treatments. Which complete conditions can be assigned?
- Choose allocation. Equal groups, unequal ratio, or fixed sizes?
- Choose restrictions. Is blocking or matching justified?
- Generate by chance. Which probability mechanism will be used?
- Conceal the sequence. Can anyone predict or alter the next assignment?
- Record the result. Is the assignment auditable and reproducible?
- Preserve the comparison. How will switching, noncompliance, and attrition be handled?
Common Errors
Assuming Random Means Haphazard
Random assignment follows a defined probability mechanism. Casual, improvised, or undocumented assignment is not equivalent to randomization.
Expecting Identical Groups
Randomization prevents systematic assignment bias, but realized groups can differ by chance. Exact balance on every characteristic is not guaranteed.
Confusing Assignment with Sampling
Random assignment supports causal inference. It does not automatically make the participants representative of a larger population.
Randomizing the Wrong Unit
If a treatment is delivered to whole classrooms, clinics, or communities, the unit of assignment is the group. The analysis must respect that structure.
Allowing the Sequence to Be Predicted
A mathematically random list can still be compromised if the next assignment is visible before enrollment. Allocation concealment protects the process.
Guided Practice
Problem 1
Twenty-four plants are numbered (1) through (24). A computer randomly selects twelve numbers for fertilizer (A); all remaining plants receive fertilizer (B). Identify the assignment design.
Show the solution
This is complete random assignment with fixed group sizes of twelve. Every set of twelve plants has the same chance to become the (A) group, and the remaining plants form the (B) group.
Problem 2
A researcher flips a fair coin independently for each of 30 participants, assigning heads to treatment and tails to control. Must the groups contain exactly 15 participants each?
Show the solution
No. Each participant has probability (1/2) of entering either group, but the final group sizes are random. Fifteen per group is one possible result, not a guarantee.
Problem 3
Participants are allowed to choose between two exercise programs. Explain why the resulting groups do not support a clean causal comparison.
Show the solution
Program choice may be related to fitness, motivation, schedule, health, or personal preference. Those characteristics may also affect the outcome, so treatment choice is confounded with preexisting differences.
Problem 4
A randomized study happens to place more experienced workers in treatment (A) than in treatment (B). Does this prove that the randomization was invalid?
Show the solution
No. Valid random assignment can produce imbalance by chance. The procedure should be checked, but the observed imbalance alone does not prove invalidity. If experience was known in advance to be highly important, blocking by experience level could have improved balance.
Problem 5
Forty classrooms are randomly assigned to two curricula. Each classroom has 25 students. How many independently assigned experimental units are there?
Show the solution
There are 40 independently assigned experimental units: the classrooms. The experiment measures outcomes for (40\times25=1000) students, but the treatment was assigned at the classroom level.
A Final Validity Check
Ask one decisive question about the assignment:
\[ \boxed{ \text{Could a unit's characteristics predict or influence its treatment?} } \]Under a properly implemented and concealed randomization procedure, the answer should be no beyond chance. That is the protection random assignment provides.
Summary
- Random assignment uses a specified chance mechanism to determine treatment membership.
- It prevents systematic selection into treatment groups and addresses both measured and unmeasured pre-treatment characteristics.
- Randomization balances groups in expectation, not necessarily exactly in one realized experiment.
- Complete random assignment fixes group sizes; independent assignment allows group sizes to vary.
- Random sampling supports generalization, while random assignment supports causal comparison.
- Blocking, matching, and cluster assignment are structured forms of randomized design.
- Allocation concealment, documentation, and preservation of the original groups protect the randomization after the sequence is generated.
Next Step
Random assignment creates a fair treatment comparison, but one or two observations cannot reveal stable patterns. The next lesson develops replication—the use of enough independent experimental units to measure variability and distinguish treatment effects from chance fluctuations.