Single-Arm Phase II Trials Explained: How They Work and Why They Matter

What Is a Single-Arm Phase II Trial?

Phase II clinical trials sit at an important point in drug development. A treatment has usually passed initial safety testing, but a much bigger question remains: does the treatment show enough activity to justify further investigation?

One common way to answer this question—particularly in oncology and other settings where the outcome can be assessed relatively quickly—is a single-arm Phase II trial.

Unlike a randomized controlled trial, a single-arm trial does not assign participants to different treatment groups. Everyone enrolled receives the investigational treatment.

Imagine that researchers are evaluating a new cancer treatment in 40 patients. All 40 receive the new treatment, and 12 experience a predefined response. The observed response rate is 12 / 40 = 30%.

But 30% by itself does not tell us whether the treatment is promising. Researchers need a reference point. If existing evidence suggests that a response rate of 10% or lower would not be clinically interesting, the real question becomes whether the observed activity is sufficiently greater than 10% to justify further study.

Where Does the Comparison Come From?

Because there is no concurrent control arm, the benchmark usually comes from external information. This may include previous clinical trials, historical control data, standard-of-care outcomes, published literature, clinical experience, or earlier studies in a similar patient population.

Suppose researchers believe that a response rate of 10% would mean the treatment is not sufficiently active, while a response rate of 30% would be promising. These values can be formalized as p₀ = 0.10, the response rate considered uninteresting, and p₁ = 0.30, the response rate considered promising.

The trial is then designed to distinguish between these two possibilities while controlling the probability of making incorrect decisions. The quality of the external comparison is therefore central to the credibility of the design.

The Statistical Question

A common formulation is H₀: p ≤ p₀ versus H₁: p ≥ p₁, where p is the true response probability of the investigational treatment.

The null hypothesis represents a treatment that is not sufficiently active to warrant further investigation. The alternative represents a level of activity that would make the treatment worth pursuing.

The trial therefore does not simply ask whether some patients responded. It asks whether there is enough evidence that the treatment’s true response rate exceeds a predefined uninteresting level.

That distinction is fundamental. The benchmark and the decision rule must be specified before the results are examined so that the evidence is interpreted against a prospectively defined clinical standard.

Why Not Simply Enroll a Fixed Number of Patients?

The simplest approach would be to choose a sample size, enroll every patient, and evaluate the results at the end. This is called a single-stage design.

For example, a study might enroll 40 patients and declare the treatment promising if at least a predefined number respond. But consider what happens if the treatment performs very poorly. If the first 15 patients are treated and none responds, should researchers automatically continue enrolling all 40 patients?

Possibly—but often there is little value in doing so. Continuing may expose additional patients to an ineffective treatment and consume clinical resources without much chance of changing the conclusion.

This is not only a question of efficiency. Early stopping also has an ethical rationale: if accumulating evidence suggests that a treatment is unlikely to provide sufficient benefit, a design that allows the study to stop can reduce the number of additional patients exposed to an ineffective therapy.

This motivates multi-stage designs.

The Idea Behind a Two-Stage Phase II Trial

A two-stage design introduces an early evaluation. Instead of enrolling every participant before making a decision, the study proceeds in two stages.

In Stage 1, a smaller group of patients is enrolled first and researchers count how many meet the predefined response criterion. If the number of responses is too low, the trial stops early for futility. If enough responses are observed, the study continues.

In Stage 2, additional patients are enrolled. At the end of the study, the total number of responses across both stages is evaluated. If the total exceeds a predefined threshold, the treatment is considered sufficiently promising for further investigation.

The important idea is simple: investigators do not necessarily need to enroll the maximum planned sample to learn that a treatment is unlikely to be worth pursuing.

A Simple Example

Suppose a Phase II trial is designed around p₀ = 0.10 and p₁ = 0.30. The study first enrolls a smaller group of patients. If very few responses are observed, the trial stops. If enough responses are observed, additional patients are enrolled.

At the end, the total number of responses is compared with a final decision threshold. The exact numbers of patients and response thresholds are not chosen arbitrarily. They are calculated so that the design satisfies predefined statistical error requirements.

This is where alpha and beta become important.

Alpha and Beta: Controlling Wrong Decisions

Any Phase II decision can be wrong. A treatment that is actually insufficiently active may appear promising, or a genuinely promising treatment may be rejected. These risks are controlled through two design parameters.

Alpha, written α, controls the probability of incorrectly declaring a treatment promising when its true response rate is at the null level. For example, α = 0.05 means the design limits this false-positive probability to approximately 5% at the specified null response rate.

Beta, written β, controls the probability of failing to declare the treatment promising when its true response rate is at the alternative level. For example, β = 0.20 corresponds to a target power of 1 − β = 80%.

Together, p₀, p₁, alpha, and beta define the statistical requirements from which the sample size and decision boundaries can be derived.

Where Simon’s Two-Stage Design Fits

One of the best-known approaches for this problem is Simon’s two-stage design, introduced by Richard Simon in 1989. It was developed for Phase II studies with a binary endpoint, such as response versus no response, success versus failure, or disease control versus no disease control.

A Simon design determines four important quantities: n₁, the number of patients enrolled in Stage 1; r₁, the Stage 1 futility boundary; n, the maximum total sample size; and r, the final decision boundary.

These numbers define exactly when the trial stops and when it continues. Conceptually, the rule is to enroll n₁ patients first, stop for futility if the number of responses is at or below r₁, continue enrollment to n patients otherwise, and use the final boundary r to decide whether the treatment is sufficiently promising.

The actual values depend on the chosen p₀, p₁, alpha, and beta. But Simon’s method does more than produce a statistically valid design. There are usually multiple two-stage designs capable of satisfying the same error requirements, so the next question becomes which valid design to choose.

Optimal vs Minimax Designs

Two well-known Simon design criteria answer that question differently.

The optimal design minimizes the expected sample size under the null hypothesis. In practical terms, it tries to use fewer patients on average when the treatment is ineffective. This can be attractive when early termination of inactive treatments is particularly important.

The minimax design minimizes the maximum total sample size. This can be useful when the total number of patients that can be recruited is the primary constraint.

The trade-off is conceptual. One design may have a smaller expected enrollment when the treatment is ineffective but require a somewhat larger maximum sample if the trial continues. Another may guarantee a smaller maximum sample size while being less efficient at stopping inactive treatments early.

Neither is universally better. They optimize different aspects of the trial. Other admissible designs can also provide useful compromises between expected and maximum sample size.

Probability of Early Termination

One of the most useful characteristics of a two-stage design is the probability of early termination, often abbreviated as PET.

PET answers the question: what is the probability that the trial stops after Stage 1? Suppose a design has PET under p₀ = 0.70. This means that if the treatment’s true response probability is p₀, there is a 70% probability that the trial will stop after Stage 1. Only about 30% of such trials would proceed to Stage 2.

When the treatment is ineffective, a high PET can therefore translate into fewer patients being enrolled on average. This is also why looking only at the maximum sample size can be misleading.

A design might allow up to 40 patients but enroll substantially fewer patients on average when the treatment is inactive.

Expected Sample Size

Because a two-stage trial can stop early, the number of patients actually enrolled is not always equal to the maximum sample size. The expected sample size accounts for this.

In simplified form: expected sample size = Stage 1 sample size + probability of continuing × additional Stage 2 sample size.

If a trial frequently stops after Stage 1 when the treatment is ineffective, its expected enrollment under p₀ can be substantially smaller than its maximum sample size.

This gives a more realistic picture of the design’s efficiency and explains the central trade-off between many Simon designs: maximum enrollment tells you the worst-case requirement; expected enrollment tells you what you expect to use under a specified response probability.

Why Single-Arm Trials Are Attractive

In appropriate settings, single-arm Phase II trials can allow an efficacy signal to be assessed with fewer patients than a randomized comparison, while remaining relatively straightforward to conduct. They can also provide an efficient way to screen whether a new treatment has enough activity to justify larger studies.

Two-stage designs add another important advantage: the possibility of stopping early when results are clearly disappointing. This can reduce unnecessary patient exposure, conserve limited clinical resources, shorten unsuccessful development paths, and allow investigators to redirect effort toward more promising treatments.

These characteristics help explain why single-arm designs have historically played an important role in early oncology drug development.

In some clinical and regulatory settings—particularly where there is substantial unmet need, the treatment effect is expected to be large, the endpoint is objective, and the disease’s natural history is sufficiently well understood—single-arm evidence may also play a larger role in development decisions.

But the lack of randomization makes the quality of the external comparison especially important.

The Biggest Limitation: Historical Comparisons

The absence of a concurrent control group is both the defining feature and the major limitation of a single-arm trial.

Suppose a study observes a 30% response rate and compares it with a historical benchmark of 10%. That comparison is meaningful only if the historical population and the current study population are sufficiently comparable.

Differences may exist in patient selection, disease severity, diagnostic criteria, prior treatments, supportive care, endpoint definitions, assessment schedules, follow-up, and changes in clinical practice.

For example, imagine that the 10% historical response rate came from a substantially different patient population or was measured using different response criteria. A 30% response rate in the new trial may still look impressive, but the apparent improvement cannot automatically be attributed to the treatment.

This is one reason the choice of p₀ deserves careful clinical justification.

The Endpoint Matters Too

Single-arm designs are most natural when the primary endpoint can be measured meaningfully without a concurrent control group. Objective tumor response is a classic example.

Other binary endpoints may also be suitable depending on the disease and treatment. Time-to-event endpoints such as progression-free survival can sometimes be evaluated against historical benchmarks, but their interpretation in a single-arm setting is generally more challenging.

Outcomes may be influenced by patient prognosis, assessment practices, patient selection, follow-up, and other factors that differ between the contemporary study and the historical data.

The statistical design should therefore follow the clinical question—not the other way around.

What a Single-Arm Phase II Trial Can—and Cannot—Tell Us

A successful single-arm Phase II trial can provide evidence that a treatment shows sufficient activity relative to a predefined benchmark. It can support a decision to continue development.

But it usually does not provide the same strength of comparative evidence as a well-designed randomized trial. Without concurrent randomization, differences between the observed results and historical outcomes may reflect more than the treatment itself.

For this reason, a useful way to think about many single-arm Phase II trials is as decision-making studies. Their purpose is often not to prove definitively that a treatment is better than another treatment. Instead, they help answer whether the treatment is promising enough to justify the next stage of development.

That is a narrower question—but an extremely important one.

From the Clinical Question to the Design

Designing a single-arm Phase II trial involves a chain of decisions: define the clinical question, choose an appropriate endpoint, define the uninteresting response rate p₀, define the promising response rate p₁, choose acceptable alpha and beta, calculate candidate designs, compare sample size and operating characteristics, and select the design that best fits the study.

The mathematics comes near the end of this process. The most consequential decisions often come earlier—especially the choice of endpoint, p₀, and p₁.

Once those clinical assumptions are clear, tools can help enumerate valid candidate designs and make their trade-offs easier to review.

Key Takeaways

A single-arm Phase II trial has no concurrent randomized control group. Results are evaluated against a predefined benchmark, often informed by historical evidence.

p₀ defines an uninteresting level of activity, while p₁ defines a promising level. These assumptions drive the design.

Alpha and beta control the risks of making incorrect development decisions.

Two-stage designs allow early stopping for futility. This can reduce both unnecessary patient exposure and resource use when a treatment is ineffective.

Simon’s two-stage design provides a systematic framework for choosing Stage 1 and final sample sizes and decision boundaries.

Optimal and minimax designs optimize different objectives. One emphasizes expected enrollment under the null; the other emphasizes maximum enrollment.

Historical comparability is critical. Statistical rigor cannot rescue a poorly chosen benchmark.

A positive single-arm Phase II trial usually means “promising enough to pursue,” not “definitively proven better.”

The Bigger Picture

A single-arm Phase II trial may look simple because every patient receives the same treatment. Statistically, however, the design represents a carefully structured decision under uncertainty.

Researchers must define what level of activity would be disappointing, what level would be genuinely promising, how much risk of an incorrect decision they are willing to accept, and how many patients should be exposed before that decision is made.

Two-stage designs make that process more efficient by allowing an ineffective treatment to be stopped early. Simon’s two-stage design provides a systematic framework for doing exactly that.

A Phase II design is not just a sample-size calculation. It is a formal definition of the evidence required to decide whether a treatment deserves to move forward.

Understanding that principle makes the individual parameters—p₀, p₁, alpha, beta, stopping boundaries, PET, and expected sample size—much easier to understand.

Frequently Asked Questions