Patient Benefit When Treatments Are Equally Effective

Why Equal Effectiveness Changes the Patient-Benefit Question

Response-adaptive randomization changes how patients are assigned as trial outcomes accumulate. If one treatment truly produces more successes than another, assigning more patients to it could affect the total number of successes. But when treatments are equally effective, moving assignments between them cannot improve a patient's success probability under a simple homogeneous model. The expected total number of successes is then the same as under fixed randomization.

That statement concerns an expectation across repeated trials, not a promise about what will happen in any one trial. One adaptive trial can have more successes than a fixed-randomization trial purely by chance; another can have fewer. To make the comparison precise, we need to specify the outcome model and what “patient benefit” means.

Definition: Consider \(n\) patients assigned sequentially to treatments \(A\) or \(B\). Let \(A_i\) denote the treatment assigned to patient \(i\), and let \(Y_i=1\) if that patient has the defined successful outcome and \(Y_i=0\) otherwise. The total number of successes is \(S=\sum_{i=1}^{n}Y_i\). In this tutorial, patient benefit means the expected value of \(S\), the expected total number of successful outcomes.

Assumptions for the Equal-Effect Result

For the central result, assume a fixed number \(n\) of patients, a binary success outcome, and two treatments with the same success probability \(p\), where \(0\leq p\leq 1\). Outcomes are generated independently: after any assignment history, the next patient's success probability is \(p\), whichever treatment is assigned. The adaptive assignment rule may use previous assignments and observed outcomes, but it must choose the current assignment before observing that patient's outcome. There is no carryover between patients, and the outcome is fully observed.

These assumptions describe more than equal average success rates in some loosely defined population. They require that, at every step in the trial, the assigned treatment does not change the current patient's success probability, and that the common probability does not drift with history. Under them, \(Y_i\) has a Bernoulli distribution with success probability \(p\), regardless of how \(A_i\) was chosen.

Theorem: Under the assumptions above, response-adaptive randomization and fixed randomization both have \(E(S)=np\). In fact, under this homogeneous independent-outcome model, \(S\sim\operatorname{Binomial}(n,p)\) under either allocation rule. Thus response-adaptive allocation has no expected advantage in total successes when treatments are equally effective, although realized totals can differ.

Deriving the Expected Total

Let \(\mathcal{H}_{i-1}\) represent the assignment and outcome history available before patient \(i\) is assigned. The adaptive rule can make \(A_i\) depend on \(\mathcal{H}_{i-1}\), but the outcome assumption says that, conditional on that history and the treatment assigned, the success probability remains \(p\). Therefore the conditional expected outcome for patient \(i\) is \(p\). Summing these expectations gives the result.

E(Y_i\mid \mathcal{H}_{i-1}, A_i)=p,\qquad E(S)=E\left(\sum_{i=1}^{n}Y_i\right) =\sum_{i=1}^{n}E(Y_i) =\sum_{i=1}^{n}p=np.

The same calculation applies to fixed randomization: each patient's success probability is \(p\), so the expected total is \(np\). For the stronger distributional statement, condition successively on the available history. Each next outcome has probability \(p\) of success and \(1-p\) of failure, independent of the preceding outcomes. Thus the joint probability of any particular sequence with \(s\) successes and \(n-s\) failures is \(p^s(1-p)^{n-s}\). There are \(\binom{n}{s}\) such sequences, so \(P(S=s)=\binom{n}{s}p^s(1-p)^{n-s}\), the binomial probability.

This argument does not require equal numbers assigned to the treatments. It requires equal success probabilities for the patients under either assignment. The distinction matters: adaptive randomization may change who receives each treatment without changing the success probability for any patient.

Worked Example: Expected Successes With Changing Assignment Probabilities

Suppose a hypothetical trial enrolls \(n=5\) patients. Both treatments have success probability \(p=0.40\). An adaptive rule may raise or lower the probability of assigning treatment \(A\) after observing earlier outcomes. Assume each assignment is made before the patient's outcome, and each patient's outcome still has success probability \(0.40\) on either treatment.

For every patient \(i\), conditional on the history and current assignment, \(E(Y_i\mid \mathcal{H}_{i-1},A_i)=0.40\). Therefore:

E(S)=\sum_{i=1}^{5}E(Y_i)=5(0.40)=2.00\text{ successes}.

A fixed-randomization design enrolling the same five patients has the same expected total: \(5(0.40)=2.00\) successes. The adaptive rule can change the treatment counts and the path of assignments, but those changes do not raise the expected success probability in this example.

Expected Benefit Is Not a Guaranteed Trial Result

The result is about an average over repetitions of the design. The realized number \(S\) is random under either allocation rule. If \(n=6\) and \(p=0.50\), for example, a trial could observe five successes, two successes, or any other count from zero to six. Under the theorem's assumptions, each allocation policy gives \(S\sim\operatorname{Binomial}(6,0.50)\), but two separate trials need not have the same observed total.

For a binomial total, the probability of exactly \(s\) successes is \(\binom{n}{s}p^s(1-p)^{n-s}\). In the six-patient example, the probability of at least five successes is:

P(S\geq 5)=\binom{6}{5}(0.50)^5(0.50)+\binom{6}{6}(0.50)^6 =\frac{6+1}{64} =0.1094.

That probability is the same under adaptive and fixed randomization under the stated model. It does not mean the designs will produce identical outcomes in a particular pair of trials. Their realized totals can differ because the patients' outcomes are random; a difference by itself does not establish that one allocation policy increased expected benefit.

Worked Example: Interpreting Two Hypothetical Trial Results

Imagine two separate, hypothetical six-patient trials, each with equally effective treatments and \(p=0.50\). One trial uses response-adaptive allocation and observes five successes; the other uses fixed randomization and observes two.

The observed totals are \(S_{\text{adaptive}}=5\) and \(S_{\text{fixed}}=2\), a difference of \(5-2=3\) successes in these particular trials. But the model gives each design an expected total of \(6(0.50)=3\) successes. The observed comparison is therefore compatible with chance variation around the same expectation.

For context, under either design the probability of exactly five successes is \(\binom{6}{5}(0.50)^6=6/64=0.09375\), while the probability of exactly two is \(\binom{6}{2}(0.50)^6=15/64=0.2344\). These calculations describe possible results under the model; they do not show that the adaptive policy caused the higher total in the first trial.

Allocation Counts and Successes by Treatment

Although expected total successes agree, expected successes on each treatment need not agree across allocation rules. Let \(N_A\) be the number assigned to treatment \(A\), and let \(S_A\) be the number of successes among them. Since each assignment to \(A\) has success probability \(p\), \(E(S_A)=pE(N_A)\); similarly, \(E(S_B)=pE(N_B)\). If an adaptive rule assigns more patients to \(A\) on average, it can produce more successes on \(A\) simply because more patients received it, not because \(A\) is more effective.

Worked Example: More Patients Assigned Does Not Mean Greater Effectiveness

In a hypothetical ten-patient trial, suppose both treatments have success probability \(p=0.60\). Under an adaptive rule, assume the expected allocation counts are \(E(N_A)=7\) and \(E(N_B)=3\). Under a fixed rule, five patients are assigned to each treatment.

For the adaptive rule, the expected successes by treatment are \(0.60(7)=4.20\) on \(A\) and \(0.60(3)=1.80\) on \(B\), for \(4.20+1.80=6.00\) total. For fixed allocation, the corresponding expectations are \(0.60(5)=3.00\) on each treatment, again totaling \(6.00\).

The larger expected success count on \(A\) under the adaptive rule reflects its larger expected patient count on \(A\). It is not a higher per-patient success probability: both remain \(0.60\).

When the Comparison Can Mislead

The equal expected-success conclusion depends on how success probabilities behave. If outcomes depend on calendar time, patient characteristics, or previous patients' outcomes, the homogeneous binomial result may not apply. Equal overall treatment rates alone do not ensure that an adaptive rule and a fixed rule have equal expected totals if the rule is responding to information related to future patients' prognosis. The relevant question is whether the assigned treatment leaves each patient's success probability unchanged under the comparison, given the information available at assignment.

Other departures also matter. Missing outcomes can distort the observed success total if their occurrence differs by assignment or outcome. Carryover, interference between patients, or an adaptation rule that uses information unavailable at the time of assignment can invalidate the stated model. In practice, assess whether outcome definitions and follow-up are consistent, whether response probabilities plausibly remain stable, and whether the allocation rule uses only information already observed. These checks do not prove the assumptions, but they help identify where the simple result may fail.

A useful comparison is therefore:

Question	Response-adaptive allocation	Fixed randomization
Expected total successes under the stated model	\(np\)	\(np\)
Assignments across treatments	May change with observed history	Set by the fixed allocation scheme
Realized total in one trial	Can vary by chance	Can vary by chance

The comparison is deliberately narrow: it concerns expected successful outcomes under equal treatment success probabilities and the stated assumptions. It does not imply that the allocation rules have identical treatment counts, identical observed data, or identical performance for every other trial objective.

Common Mistake

Wrong reasoning: “The adaptive trial observed more successes, so adaptive allocation benefited patients.” A single realized difference does not establish an expected benefit. Under the equal-probability model, both designs have expected total \(np\), and their realized totals fluctuate around that value.

Correction: Separate the design-level expectation from the outcome of one trial. State the assumed success probabilities and allocation conditions, calculate the expected total under each rule, and interpret observed differences as realizations unless there is evidence that the assumptions or treatment effects differ.

Key takeaway: When every patient has the same success probability \(p\) on either treatment, expected total successes equal \(np\) under response-adaptive and fixed randomization. Adaptive assignment can change treatment counts, but chance still determines the realized outcomes in any one trial.

Check Your Understanding

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

A hypothetical trial enrolls eight patients; both treatments have success probability \(0.25\). What is the expected total number of successes under either allocation rule?
Under the homogeneous model, does response-adaptive allocation change the distribution of the total number of successes? State the distribution for \(n=4\) and \(p=0.70\).
An adaptive design expects to assign six patients to \(A\) and four to \(B\), with \(p=0.50\) on both. What are the expected successes by treatment and in total?
Why does observing more successes in one adaptive trial than in a separate fixed-randomization trial not establish an expected benefit?
Name one assumption whose failure could make the binomial comparison inappropriate.

Solutions. (1) \(E(S)=8(0.25)=2.00\) successes. (2) No. \(S\sim\operatorname{Binomial}(4,0.70)\) under either rule, with expectation \(4(0.70)=2.80\). (3) The expected counts are \(0.50(6)=3.00\) on \(A\) and \(0.50(4)=2.00\) on \(B\), totaling \(5.00\). (4) Each observed total is random; under the stated assumptions both designs have the same expected total, so a difference can arise by chance. (5) Examples include changing success probabilities over time, carryover, interference between patients, or incomplete outcome observation.