Defining The Hypothetical Assignment Mechanism

Why Specify the Assignment Mechanism?

A trial protocol needs more than a list of treatment strategies. It must say how eligible participants would be assigned to those strategies. “Participants are randomized” is incomplete: randomization could occur person by person, by clinic, or within prespecified strata; it could use a fixed allocation ratio or another probability; and it must occur at a stated point relative to the start of the assigned strategies.

The assignment mechanism is part of the hypothetical experiment that gives a causal question its meaning. It identifies the unit that receives an assignment and the rule governing that assignment. A hypothetical assignment rule should be operational: a reader should be able to determine who or what is randomized, when that happens, and the probability of each possible assignment.

Let \(A_i=1\) indicate that participant \(i\) is assigned to strategy 1, and \(A_i=0\) indicate assignment to strategy 0. Let \(Y_i(1)\) and \(Y_i(0)\) denote that participant’s potential outcomes under the two strategies. In a randomized trial, the allocation rule determines the distribution of \(A_i\); it does not determine which strategy a participant actually follows after assignment. Assignment and adherence are distinct.

What the Protocol Must State

A useful specification identifies at least four features. First, state the allocation unit: an individual participant or a cluster such as a clinic. Second, state when allocation occurs, in relation to eligibility and the start of the assigned strategy. Third, state the allocation ratio, such as 1:1 or 2:1. Fourth, describe any restrictions on the randomization, such as fixed totals, blocks, or strata.

A timing statement should make clear that the allocation is made before the participant begins the strategy to which they are assigned. If an assignment is made only after observing a post-assignment event, the protocol no longer describes the same allocation process for all eligible participants. The exact ordering of eligibility assessment, assignment, and follow-up is an important design issue, but the assignment specification should at minimum identify when the allocation rule is applied.

When each person is randomized independently with probability \(q\) to strategy 1, the number assigned to strategy 1 is random. Under complete randomization, by contrast, a fixed number \(n_1\) of \(N\) participants is selected for strategy 1. Each participant still has marginal assignment probability \(n_1/N\), but assignments are dependent because the total is fixed.

A stratified or blocked scheme adds restrictions to the allocation. For example, the protocol might randomize separately within each prespecified site or within groups defined by baseline risk. This can help ensure representation across those categories, but the protocol must say how the restrictions operate and what allocation probabilities result. It is not enough to name “stratified randomization” without defining the strata and the allocation rule within them.

Cluster-Based Assignment

In cluster randomization, a group rather than an individual receives the assignment. For example, clinics might be assigned to different care strategies, with participants at a clinic receiving the strategy assigned to that clinic. If \(A_j\) is the assignment of clinic \(j\), then each participant \(i\) in clinic \(j\) has \(A_i=A_j\) under this design.

The protocol should name the clusters, specify how many clusters are assigned to each strategy or the probability of each cluster’s assignment, and explain how participants are associated with clusters for purposes of the assignment rule. If clusters vary in size, a 1:1 allocation of clusters does not imply that exactly half of participants are assigned to each strategy. Nor should individual assignment probabilities be treated as independent when all participants in a cluster share the same assignment.

Why the Randomization Rule Matters

Under a correctly implemented randomized assignment, the assignment is generated independently of participants’ potential outcomes, in the sense specified by the randomization design. This is the central reason random allocation can support causal comparisons. The design does not guarantee that the realized groups will be identical in every baseline characteristic; randomization balances characteristics in expectation under the allocation procedure, not necessarily in a particular sample.

The level of randomization also matters for the analysis and interpretation. In an individually randomized trial, the allocation varies person by person. In a cluster-randomized trial, assignment is shared within clusters, so the number of independent assignment units is the number of clusters, not the number of people. Treating all people as if they had been independently randomized would misrepresent the assignment process and can produce uncertainty estimates that are too small.

For an observational study that emulates a hypothetical trial, the protocol still states the trial’s intended assignment mechanism. The observed treatment decisions in the data may not follow that random mechanism. A description of the hypothetical randomization must not be mistaken for evidence that treatment was actually randomized; the emulation must address the differences between the target trial’s assignment and treatment received in routine care.

Diagnostics, Alternatives, and Failure Modes

Before treating the assignment specification as coherent, check whether the rule could actually be carried out. Confirm that allocation occurs at the stated unit, that the allocation ratio and restrictions are explicit, and that the participant or cluster counts make the stated fixed totals possible. For a stratified design, verify that the strata are defined before allocation and that the within-stratum rules are stated. For cluster assignment, check the number and sizes of clusters and whether the intended assignment is truly shared within each cluster.

In a completed randomized trial, allocation records can be checked against the stated rule: group counts, stratum-specific totals, and cluster assignments should be consistent with it. A baseline table can describe observed imbalances, but a significance test of baseline differences is not a test that randomization was properly performed. Chance imbalances can occur under valid randomization; apparent balance does not prove that assignment was random.

The main alternatives differ in what is randomized and what is held fixed. Independent individual assignment makes each person’s assignment a separate random draw, so group sizes can vary. Complete individual randomization fixes group totals, creating dependence among assignments. Cluster randomization assigns groups and can be appropriate when delivery is inherently group-based, but it may yield unequal participant totals and fewer independent assignment units. Stratified or blocked randomization imposes additional balance restrictions across specified categories. A nonrandom usual-care allocation rule may describe actual practice, but it is not random assignment.

A frequent pitfall is to write “participants are randomized 1:1” when clinics, not participants, are actually allocated. The reasoning may be that every participant is in one of two arms, so the allocation is individual. The correction is to identify the unit at which the random draw occurs: if one clinic-wide draw determines the assignments of all its patients, assignment is cluster-based. Another pitfall is to infer a fixed half-and-half patient count from a 1:1 cluster allocation. With unequal cluster sizes, that conclusion does not follow, as the 690-patient example shows.