What Concurrent Controls Do and Do Not Guarantee

What Concurrent Controls Protect Against

In a trial with calendar-time drift, prognosis may change as participants enroll. Usual care, clinical practice, eligibility patterns, or the underlying condition may change over the course of the trial. A control group enrolled at the same time as a treatment group can provide a contemporaneous comparison: both groups are exposed to shared changes in the calendar environment. But “concurrent” does not by itself guarantee that a crude comparison is free of time effects. The protection depends on how allocation distributes participants over time and on how the comparison is constructed.

The key distinction is between a shared calendar-time change and a difference in calendar-time composition between arms. Shared changes cancel in expectation when randomization makes the treatment and control groups comparable within time. If the two arms contain different proportions of early and late enrollees, however, their pooled means can reflect different mixtures of prognosis. That can happen even when randomization is used and even when every participant has a positive probability of assignment to either arm.

A Finite-Population Setup

Suppose the trial has \(K\) calendar periods. In period \(k\), let \(N_k\) participants enroll, and let \(Y_i(1)\) and \(Y_i(0)\) denote participant \(i\)'s potential outcomes under treatment and control. Only the outcome corresponding to the assigned arm is observed. Assume consistency (the observed outcome equals the potential outcome under the received assignment), no interference between participants, and a well-defined treatment strategy. These assumptions let us describe the treatment effect for each participant as \(Y_i(1)-Y_i(0)\).

For the shared-drift argument, write the control potential outcome as \(Y_i(0)=b_k+u_i\), where \(b_k\) is the period-specific shared calendar component and \(u_i\) is the participant’s remaining prognostic contribution. For illustration, a constant treatment effect \(\tau\) gives \(Y_i(1)=b_k+u_i+\tau\). More generally, treatment effects may vary across participants or periods; define \(\bar Y_{ak}\) as the finite-population mean of potential outcome \(Y_i(a)\) among the participants enrolled in period \(k\).

A useful condition for a simple within-period comparison is uniform fixed-count randomization: in period \(k\), exactly \(m_{1k}\) of the \(N_k\) participants are assigned to treatment, and every subset of \(m_{1k}\) participants is equally likely to receive it. The remaining \(m_{0k}=N_k-m_{1k}\) receive control. Under this design, each participant has the same probability \(m_{1k}/N_k\) of treatment within the period, and both arm means are based on random subsets of that period’s participants. The condition is stronger than merely saying that assignment is randomized.

Why the Within-Period Result Holds

Under uniform fixed-count randomization, each participant has treatment inclusion probability \(m_{1k}/N_k\). The treatment mean is the average of the \(m_{1k}\) observed treatment outcomes. Its expectation is the average of all \(N_k\) treatment potential outcomes: each potential outcome is included with the same probability, and the fixed denominator is \(m_{1k}\). The same argument applies to control, for which each participant has inclusion probability \(m_{0k}/N_k\). Subtracting the expected means gives the finite-population average treatment effect in that period.

This is an expectation over the random assignment, conditional on the enrolled participants and their potential outcomes. It does not say that the realized arm means must be equal on prognostic characteristics or that a realized difference must equal the period’s treatment effect. Chance imbalances remain possible. Nor does it apply automatically when participants have different assignment probabilities: then the unweighted arm means need not average the same participant population.

Why Unequal Allocation Across Periods Can Distort a Pooled Contrast

A pooled crude contrast subtracts the mean among all treated participants from the mean among all controls, without first ensuring that the arms represent the same mix of calendar periods. Even if randomization is uniform within each period, treatment and control can have different period weights when their allocation counts differ across periods. Consequently, the two pooled means can average different period-specific prognoses.

To see this precisely, suppose fixed arm counts are used within every period and randomization is uniform within each period. Let \(M_1=\sum_k m_{1k}\), \(M_0=\sum_k m_{0k}\), \(w_{1k}=m_{1k}/M_1\), and \(w_{0k}=m_{0k}/M_0\). Then the expected pooled crude contrast is

Writing \(\bar Y_{1k}=\bar Y_{0k}+\tau_k\), where \(\tau_k=\bar Y_{1k}-\bar Y_{0k}\), separates the effect from the composition difference:

The first term is a calendar-composition component. It vanishes if the arm period weights match, or if the control potential-outcome means are constant across periods. Otherwise, even a constant treatment effect does not prevent the pooled contrast from including a drift contribution. The second term is an effect average weighted by the treatment arm’s period composition; if treatment effects vary over time, this weighting also matters to the estimand.

Why Randomization Alone Is Not Enough for Unweighted Means

The fixed-count uniform condition matters. In response-adaptive trials, assignment probabilities can differ across participants as the trial progresses. Randomization still provides a known chance of assignment, but if those chances vary, an unweighted arm mean can give disproportionate representation to participants with certain prognoses. Positive assignment probability for everyone prevents deterministic exclusion; it does not, by itself, make the unweighted difference unbiased.

Diagnostics, Alternatives, and Limits

A practical diagnostic is to inspect the arm-by-time enrollment table and the allocation-probability timeline together. Look for periods with very different arm counts or allocation probabilities, especially when prognosis is changing. These summaries identify possible composition problems; they do not prove that a crude contrast is biased or establish the direction or size of any distortion. The relevant concern is whether the periods that differ in arm representation also differ in expected outcomes.

Concurrent controls are generally preferable to relying only on historical controls for a time-varying setting: historical participants cannot share the trial’s actual calendar conditions. But a crude pooled comparison is not automatically preferable to a time-aware analysis. When period mixes differ, an analysis must be aligned with the intended estimand and the design; period-specific comparisons or an appropriately adjusted estimator may be needed. The choice depends on the randomization scheme and the target population, not merely on whether controls are concurrent.

The protection also has boundaries. It does not remove random realized imbalance, guarantee adequate representation of both arms in every period, or resolve treatment effects that change over time. It does not account for a trend that differs between arms for reasons other than the treatment, nor for a misspecified description of time. Very broad periods can hide meaningful changes within a period. Finally, concurrency cannot compensate for an estimator that ignores unequal assignment probabilities.