Random Assignment and Identification of Mean Effects

Why Random Assignment Matters

A treatment-group mean and a control-group mean are measured in different people. Their difference is not automatically a causal effect: the groups might differ in ways that also affect outcomes. Random assignment addresses this problem by making treatment assignment unrelated to participants’ potential outcomes in the assignment mechanism. Consequently, the groups are comparable in expectation, even though they need not be exactly alike in any one trial.

The distinction between expectation and a realized trial is important. Randomization does not guarantee that each arm has the same average age, baseline health, or potential outcomes. It does guarantee, under the design, that systematic differences between arms are not built into assignment. This gives a basis for identifying a mean causal effect from observed outcomes, subject to additional assumptions.

Definition: Let \(A_i\) indicate whether unit \(i\) is assigned treatment (\(A_i=1\)) or control (\(A_i=0\)). Let \(Y_i(1)\) and \(Y_i(0)\) be that unit’s potential outcomes under treatment and control, respectively. The average treatment effect (ATE) in a target population is \(E[Y(1)-Y(0)]\). The effect is identified if the assumptions and observed-data distribution determine its value uniquely.

For this tutorial, random assignment means that the treatment assignment mechanism does not use the units’ potential outcomes. In a population formulation, this is written \(A \mathrel{\perp\!\!\!\perp} (Y(0),Y(1))\). In a trial that assigns a fixed number of participants to each arm, assignments across participants are dependent, but the random allocation is still made without using their potential outcomes. The key implication for each arm is the same: assignment does not systematically select units with different potential outcomes.

From Random Assignment to Identification

Random assignment alone is not enough to connect the outcome recorded in the data to the potential outcome named in the causal question. The following conditions make that connection precise.

Random assignment: the assignment mechanism is independent of the potential outcomes, as described above.
Positive assignment probabilities: each unit in the target population has a nonzero chance of assignment to each arm. In a population formulation, \(0<P(A=1)<1\).
Consistency: the observed outcome equals the potential outcome corresponding to the assignment received: \(Y=Y(A)\). The treatment and control conditions must be sufficiently well defined for these potential outcomes to refer to the interventions of interest.
No interference: one unit’s outcome is not changed by other units’ assignments. If outcomes can depend on other people’s treatment, the simple two-potential-outcome notation needs to be expanded.
Outcome observation: the outcomes used in the comparison are observed as intended. If outcomes are missing selectively, the observed participants in each arm may no longer preserve the comparison created by randomization.

Under these conditions, random assignment makes each arm’s mean potential outcome match the corresponding population mean potential outcome in expectation. Consistency then lets the observed arm outcome stand in for that potential outcome. In population notation, the identification argument is:

E[Y\mid A=1] =E[Y(1)\mid A=1] =E[Y(1)]

The first equality follows from consistency among units assigned treatment; the second follows from random assignment. Similarly, \(E[Y\mid A=0]=E[Y(0)]\). Therefore, the population difference in observed conditional means identifies the ATE:

E[Y\mid A=1]-E[Y\mid A=0] =E[Y(1)]-E[Y(0)] =E[Y(1)-Y(0)]

The last equality uses linearity of expectation. This is an identification result: it says what population quantity the arm contrast represents under the assumptions. It does not say that the two sample means in a particular trial will equal the population means, nor does it by itself quantify sampling uncertainty.

Worked Example: Expected Potential-Outcome Balance in a Small Trial

Suppose four hypothetical participants have potential outcomes measured in symptom points, where lower scores are better. Their outcomes are shown below. A trial randomly assigns exactly two participants to treatment and two to control, with each allocation of two participants equally likely.

Participant	\(Y(0)\), control	\(Y(1)\), treatment
1	2	3
2	4	5
3	6	7
4	8	9

The population mean under control is \((2+4+6+8)/4=5\) symptom points. Under treatment it is \((3+5+7+9)/4=6\) points. The finite-population ATE is therefore \(6-5=1\) point, indicating one point higher average symptoms under treatment in this hypothetical population.

Each participant has probability \(2/4=1/2\) of entering the treatment arm. The expected treatment-arm mean of \(Y(1)\) is \((1/2)(3+5+7+9)/2=6\) points: each participant’s treatment potential outcome receives weight \(1/2\) for inclusion, and the arm mean divides the total by two. Likewise, the expected control-arm mean of \(Y(0)\) is \((1/2)(2+4+6+8)/2=5\) points. Thus, over the possible random allocations, the arm means recover the respective potential-outcome means in expectation. The actual assignment can produce different means; the claim is about the allocation mechanism, not guaranteed equality in this particular trial.

Interpreting an Observed Trial Contrast

In practice, all that is observed for each participant is the outcome under the assigned condition. The sample mean among those assigned treatment estimates the treatment potential-outcome mean, and the corresponding control sample mean estimates the control potential-outcome mean. Their difference is therefore a natural estimator of the ATE when assignment is random and the outcome and treatment definitions satisfy the conditions above. A separate result establishes its sampling properties; here the central point is why the causal interpretation is justified.

Worked Example: Interpreting a Randomized Outcome Difference

In a hypothetical randomized trial, 50 patients are assigned to a new programme and 50 to usual care. The outcome is the number of symptom days during a fixed follow-up period. The treatment-arm mean is 8.2 days, and the control-arm mean is 10.1 days. Assume all assigned patients receive the condition they were assigned, all outcomes are observed, the programme and usual care are well defined, and one patient’s assignment does not affect another patient’s outcome.

The observed contrast is \((8.2-10.1)\) days \(=-1.9\) days. With random assignment, consistency, and the stated conditions, this estimates an ATE of \(-1.9\) symptom days in the trial’s target population: assignment to the programme is associated with 1.9 fewer mean symptom days than assignment to usual care, as a causal effect of the assigned conditions.

This does not establish that every patient benefits by 1.9 days. The ATE averages individual effects, which can vary. Nor does the observed estimate have to equal the true ATE exactly: it is a result from one randomized allocation. The example provides no uncertainty interval, so it does not establish the precision of the estimate.

A useful diagnostic question is whether the actual design and data collection preserved the randomized comparison. Baseline differences between arms can occur by chance and do not, by themselves, show that randomization failed. More concerning are deviations that make assignment depend on prognosis, systematic loss of outcomes, or treatment versions that differ in consequential ways. A description of the assignment procedure, arm-specific outcome availability, and what participants actually received helps assess these risks.

Worked Example: What Changes With Nonadherence?

Imagine a hypothetical trial that randomly assigns participants to an offer of a coaching programme or to no offer. Some assigned an offer never attend, while some controls obtain similar coaching elsewhere. The outcome is stress score in points, and the mean score is 12 in the offer arm and 14 in the control arm.

Because assignment was randomized, the contrast \(12-14=-2\) points can, under consistency and the other stated design conditions, be interpreted as the causal effect of being assigned or offered the programme. It is not automatically the causal effect of actually receiving coaching. Receipt is no longer the randomized condition: people who attend may differ from people who do not, and controls may also receive coaching. The observed difference alone does not identify the effect of receipt.

To interpret the result, name the treatment precisely. If the intervention of interest is the offer, nonattendance is part of the offer’s real-world effect. If the question concerns receipt, additional assumptions and methods are needed; a simple comparison by receipt would discard the protection supplied by randomization.

When the Randomized Comparison Can Mislead

The assumptions are substantive, not just mathematical details. If assignment is predictable or altered according to participants’ expected outcomes, the arms may differ systematically in their potential outcomes. If the treatment is not well defined, “the treatment mean” may combine materially different interventions. If one participant’s assignment changes another’s outcome, the two-potential-outcome setup may not describe the relevant causal effect. And if outcome observation depends on both assignment and prognosis, a comparison restricted to observed outcomes can lose the original randomized balance.

Worked Example: Differential Outcome Missingness

In a hypothetical randomized study, 100 participants are assigned to each arm. Follow-up outcomes are recorded for all 100 controls but only 70 of the 100 treatment-assigned participants. The observed mean outcome is 20 units among observed treatment participants and 18 units among controls. A direct calculation gives \(20-18=2\) units.

The original assignment was randomized, but the available-outcome comparison is between 70 treatment participants and 100 controls. If outcome availability in the treatment arm is related to the potential outcome—for example, if participants with worsening outcomes are more likely to miss follow-up—the observed treatment mean need not equal the treatment-arm mean that would have been seen with complete follow-up. Thus, the 2-unit difference is not automatically an identified ATE. The missingness pattern and assumptions about missing outcomes must be addressed before making that causal claim. Reporting the 70% and 100% outcome-observation rates makes this potential problem visible; equal rates alone would not prove that missingness is harmless.

Random assignment is a design-based alternative to approaches such as propensity score matching or covariate adjustment in observational studies. Those approaches use measured covariates to seek comparability and depend on assumptions about whether relevant confounding has been measured and handled. In a properly implemented randomized trial, assignment itself supplies the key comparability argument. Neither approach removes the need to define the intervention and outcome carefully, and randomization cannot by itself solve interference, nonadherence questions, or outcome missingness.

Common Pitfalls and Practical Limits

“Randomized means perfectly balanced.” No. Randomization ensures comparability in expectation across its possible allocations. Chance imbalances remain possible in an individual trial.
“Any observed difference is the exact causal effect.” No. The arm contrast targets the ATE under the stated assumptions, but a realized estimate varies from allocation to allocation.
“Randomization guarantees that the effect of treatment received is identified.” No. It randomizes assignment, not necessarily receipt. The directly supported contrast concerns assignment to the trial conditions.
“The original randomization protects every later subset.” No. Restricting analysis to participants with observed outcomes can undermine comparability if inclusion depends on prognosis or treatment assignment.

Key takeaway: Random assignment makes potential outcomes comparable across arms in expectation. With positive assignment probabilities, consistency, no interference, and suitable outcome observation, the population difference in mean observed outcomes identifies the ATE. A single trial can still show chance imbalance, and failures in treatment definition or outcome follow-up can limit the causal interpretation.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions below.

In one sentence, distinguish “comparable in expectation” from “exactly balanced” in a realized randomized trial.
A trial reports mean outcomes of 5.4 points in the treatment arm and 6.1 points in control. Under the assumptions in this tutorial, what is the estimated mean effect, and what does its sign mean?
Why does random assignment to an offer not automatically identify the effect of receiving the offered treatment?
Why can a complete-case comparison lose the causal protection of randomization when follow-up is incomplete?

Solutions. (1) Over the allocations allowed by the randomization procedure, each arm represents the relevant potential outcomes without systematic selection; one particular allocation can still have chance differences. (2) The estimated contrast is \(5.4-6.1=-0.7\) points. A negative value means the treatment-arm outcome average is 0.7 points lower; if lower scores are favorable, this is an estimated favorable mean effect in that direction, not proof that every unit benefits. (3) Receipt may differ from assignment: attendance or uptake is not randomized, and people receiving treatment can differ systematically from those who do not. The randomized contrast directly concerns the offer. (4) If the chance of having an observed outcome is related to potential outcomes, the observed subset in an arm can be selectively composed, so its mean need not represent the corresponding randomized arm mean.