Why the Assignment Process Matters
A comparison of treated and untreated people is easier to interpret when we know how they came to be in those groups. In one study, a chance mechanism may assign participants to treatment or control. In another, participants or providers may select treatment based on health, preferences, resources, or other characteristics. The groups can look similar in some respects in either study, but the process that formed them is different.
That distinction matters because pretreatment characteristics can be related both to treatment selection and to outcomes. If people with more severe illness are more likely to receive a new treatment, for example, their outcomes may differ from those of untreated people even apart from any effect of the treatment. A chance-based assignment process is designed not to select participants according to such characteristics. It does not, however, promise that every realized trial group will have identical baseline characteristics.
Let \(D_i=1\) indicate that participant \(i\) is assigned to treatment and \(D_i=0\) indicate assignment to control. Let \(X_i\) denote pretreatment characteristics, such as baseline disease severity or age, and let \(Y_i(1)\) and \(Y_i(0)\) denote the participant's potential outcomes under treatment and control. The causal question concerns outcomes under these treatment conditions; the design question is how \(D_i\) was determined. The earlier material on exchangeability explains why the relationship between treatment assignment and potential outcomes matters for identification.
Chance Assignment and Selection
In a randomized design, the protocol specifies how chance determines assignment. For example, a computer may assign each participant with a known probability, or a lottery may allocate a fixed number of treatment slots. The key feature is not that the allocation looks unpredictable after the fact. It is that the assignment procedure uses chance rather than participant or provider characteristics to decide who receives which assignment.
A simple expression of this feature is that, under a design that assigns treatment with probability \(p\) independently of baseline characteristics, the probability of treatment is the same at every value of \(X\):
For such a mechanism, the distribution of \(X_i\) among those assigned to treatment is the same as its distribution among those assigned to control across the repeated assignments represented by the design. This is a statement about how the mechanism operates, not a guarantee that two particular groups in one realized sample will match exactly.
In observational data, the assignment probability may vary with baseline characteristics. A physician might be more likely to prescribe a treatment to patients with severe symptoms; a patient might select a treatment because of insurance coverage or personal preference. In those settings, the treatment groups can differ before treatment begins for reasons tied to the selection process. Measuring some of those characteristics can support an adjusted comparison, but the existence of measured data does not itself show that all important selection factors have been captured.
Worked Example: A Lottery for Treatment Slots
Suppose 40 eligible participants are entered into a computer lottery. The protocol assigns 20 to a new program and 20 to usual care, with every possible set of 20 participants equally likely to receive the program. Before the lottery, 12 participants have high baseline severity and 28 have low severity. What does the design say about assignment, and what can be concluded about the realized groups?
Assignment mechanism. The lottery—not severity, participant preference, or a provider decision—selects the 20 program assignments. For any particular participant, the probability of program assignment is \(20/40=0.5\). The lottery is therefore a randomized assignment procedure.
Baseline characteristics. The design does not require the 12 high-severity participants to split exactly six and six between the two groups. Such a split is possible, but other splits are possible too. For instance, a realized allocation might put eight high-severity participants in the program group and four in usual care. That difference would not, by itself, show that the lottery used severity to assign treatment.
Conclusion. We can classify the assignment as randomized because the specified lottery determines treatment. We cannot infer from that classification that the observed groups have identical baseline severity. The actual allocation and its baseline characteristics should still be reported.
How Observational Selection Can Shape a Comparison
In an observational setting, assignment probabilities can depend directly on \(X\). The following example makes that dependence visible. The calculation illustrates selection into treatment; it does not establish that severity is the only relevant characteristic or that adjustment for severity alone would remove all bias.
Worked Example: Treatment Selection by Baseline Severity
Imagine a hypothetical clinic with 100 patients: 40 have high baseline severity and 60 have low severity. Clinicians treat 30 of the high-severity patients and 12 of the low-severity patients. Find the treatment probability within each severity group and compare the treated group's composition with the clinic's overall composition.
Among high-severity patients, the treatment probability is \(30/40=0.75\), or 75%. Among low-severity patients, it is \(12/60=0.20\), or 20%. Thus assignment depends on severity: \(P(D=1\mid X=\text{high})=0.75\) differs from \(P(D=1\mid X=\text{low})=0.20\).
There are \(30+12=42\) treated patients, of whom 30 have high severity. The high-severity share among treated patients is therefore \(30/42\approx0.714\), or 71.4%. In the whole clinic, the high-severity share is \(40/100=0.40\), or 40%.
Conclusion. The treated group is much more concentrated in high-severity patients than the source clinic. A raw outcome comparison could therefore reflect both treatment and pre-existing severity differences. This arithmetic identifies a baseline composition difference; it does not determine the direction or size of any resulting outcome bias.
An important practical question is whether the study's description matches the procedure that was actually followed. Calling a study a “trial” is not enough to establish that assignment was randomized. Investigators need records that document how the allocation was generated, protected, and implemented. Conversely, observational treatment selection may be complex rather than an explicit decision by one person: eligibility rules, availability, institutional policies, and timing can all affect who receives treatment.
Assignment Is Not Always Treatment Received
A treatment assignment and the treatment a participant actually receives are distinct when there is nonadherence, crossover, or a failed delivery. A randomized protocol can assign participants by chance even if some assigned participants do not take the treatment and some assigned to control later obtain it. The assignment mechanism remains a feature of the design, while the received-treatment pattern describes what happened after assignment.
Worked Example: Random Assignment With Crossover
In a hypothetical randomized study, 50 participants are assigned to a new service and 50 to standard care by computer. Of those assigned to the new service, 45 receive it; of those assigned to standard care, 8 later obtain the new service elsewhere. Does crossover change how the original assignment was generated?
No. The computer assigned 50 participants to each group by the stated chance procedure, so the original assignment is randomized. The receipt counts are \(45/50=0.90\), or 90%, in the assigned new-service group and \(8/50=0.16\), or 16%, in the assigned standard-care group.
These counts show that assignment and receipt are not identical for every participant. They do not turn the original allocation into observational assignment. They do mean that a comparison classified by treatment actually received would no longer be the same as a comparison by randomized assignment: receipt may reflect decisions made after allocation. The study report should therefore distinguish assigned group from treatment received.
The reverse caution also matters. A system may assign participants according to a rule that appears impersonal—for example, a provider's judgment or a threshold in an administrative process—but that alone does not establish randomization. To classify a design, identify the actual assignment rule and ask whether chance, rather than characteristics or decisions, determined the assignment.
Auditing the Design and Recognizing Failure Modes
A careful design audit focuses first on the assignment procedure, not on whether the treatment groups happen to have similar baseline summaries. Useful questions include:
- Who or what determined treatment assignment: a chance device, the participant, a provider, or a service system?
- Was the random procedure specified before assignment, and can records verify that it was followed?
- Could staff or participants know or influence the next assignment before enrolling or assigning a participant?
- Were there overrides, exclusions after assignment, crossovers, or missing records that could alter the groups being compared?
- Is “assignment” being reported separately from treatment actually received?
These checks help distinguish an intended design from its implementation. For example, a computer-generated sequence may be random, but if staff can predict upcoming assignments and selectively enroll participants, the process that forms the study groups may no longer follow the intended sequence. A design label cannot repair a compromised procedure.
Baseline tables are useful for describing the groups, but a difference in a measured characteristic does not prove that a randomized mechanism failed. Nor does apparent similarity prove that an observational comparison is free of selection concerns. Baseline summaries describe observed data; design documentation explains how assignment occurred. Both kinds of evidence matter, but they answer different questions.
Comparing the Design Options
| Feature | Randomized assignment | Observational assignment |
|---|---|---|
| What determines assignment? | A specified chance mechanism | Participant, provider, or system decisions and characteristics |
| Can baseline characteristics guide assignment? | Not under a procedure that assigns independently of those characteristics | They may directly influence treatment selection |
| What do baseline summaries tell us? | They describe the groups actually produced; differences can occur | They describe the groups and may reveal measured selection differences |
| What must be checked? | Whether the random procedure was implemented as intended | How selection arose and whether important causes of treatment and outcome are measured |
Neither label removes the need to define the causal comparison and justify its assumptions. Randomized assignment is a design feature that can support comparability because assignment is not chosen according to baseline characteristics under the specified mechanism. Observational assignment requires careful attention to the factors that shaped treatment selection. Covariate adjustment and related methods can address some measured differences under additional assumptions, but they do not change how participants originally entered their treatment groups.
The central limitation is therefore one of inference, not vocabulary. In a randomized study, implementation records and the realized data must still be examined; chance does not force exact sample equality. In an observational study, measured baseline similarity is not sufficient evidence that treatment selection was unrelated to potential outcomes. Classifying the assignment process accurately is the first step toward deciding what claims a treatment comparison can support.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A study allocates treatment using a lottery, but the treated group has a higher mean baseline age. Does the age difference alone show that assignment was not randomized?
- In a hypothetical clinic, 18 of 30 high-severity patients and 10 of 50 low-severity patients receive a treatment. Calculate the treatment proportion in each severity group. Does this pattern indicate that assignment depends on severity?
- A provider chooses treatment based on a patient's symptoms. The treated and untreated groups have equal average age. Is this randomized assignment? Explain.
- In a randomized study, 6 participants assigned to treatment do not receive it. Does that fact alone change the original assignment mechanism?
- What evidence is more direct for classifying an assignment process: a baseline table or records describing the allocation procedure? Why?
Solutions. (1) No. A lottery can produce a realized age difference; the difference alone does not establish that the procedure was not random. Check the allocation procedure and report the observed difference. (2) The high-severity treatment proportion is \(18/30=0.60\), or 60%. The low-severity proportion is \(10/50=0.20\), or 20%. Since these probabilities differ across severity groups, the pattern is consistent with treatment selection depending on severity. (3) No. The provider's symptom-based choice determines treatment, so assignment is observational even if the groups have equal average age. (4) No. Nonreceipt describes what happened after assignment; it does not by itself alter how the original assignments were generated. (5) Allocation-procedure records are more direct because they document the mechanism that determined assignment. A baseline table describes the resulting groups but cannot establish that the mechanism used chance.