Two Objectives, One Allocation Rule
A response-adaptive trial uses accumulating information to change the probabilities of assigning future participants to treatments. This can improve the expected outcomes of participants enrolled later, especially when one treatment is genuinely better. But the same allocation decisions also determine how much information is collected about each treatment. Assigning fewer participants to one arm can make a treatment-effect estimate less precise, even when the trial’s overall expected number of favorable outcomes increases.
These are different design objectives. Patient benefit concerns the outcomes experienced by participants during the trial. Inferential performance concerns how well the trial estimates a treatment contrast, including its precision and the reliability of uncertainty statements. Neither objective automatically guarantees the other. A design should therefore be assessed on both, rather than judged by the fraction assigned to the apparently better treatment alone.
The treatment contrast is a population quantity, whereas cumulative reward is an outcome summed across the particular trial participants. A policy can improve expected cumulative reward without increasing the precision of the contrast estimate. Conversely, a design that allocates evenly may be valuable for comparison even if it gives fewer participants the treatment with the larger mean reward.
Expected Reward Under Adaptive Assignment
Write \(H_{t-1}\) for the information available just before participant \(t\) is assigned. For two treatments, let \(q_t(B\mid H_{t-1})\) be the probability of assigning treatment \(B\), so the probability of \(A\) is \(1-q_t(B\mid H_{t-1})\). The assignment probability can vary with the history. To express expected reward using fixed treatment means, however, a further condition is needed: history must not change the conditional mean potential outcome for either treatment.
Specifically, suppose that, conditional on \(H_{t-1}\), the mean potential rewards remain \(E\{Y_t(A)\mid H_{t-1}\}=\mu_A\) and \(E\{Y_t(B)\mid H_{t-1}\}=\mu_B\). Also suppose the random assignment is independent of the current participant’s potential outcomes given the history, as in a properly randomized policy. These conditions hold, for example, with independent participant outcomes from stable treatment-specific distributions and assignment probabilities based only on past data. They do not follow merely from stable marginal means and no interference.
Taking expectations over the history and summing over participants gives the expected cumulative reward. This calculation averages over both the random assignment and the trial’s evolving history; the probabilities can themselves be random because they depend on earlier outcomes.
The derivation uses the law of total expectation, participant by participant. Conditional on the history, the assigned treatment is \(A\) with probability \(1-q_t\) and \(B\) with probability \(q_t\). Under the stated conditional-mean and randomization assumptions, the corresponding conditional mean rewards are \(\mu_A\) and \(\mu_B\). Averaging that mixture over histories and adding the expectations yields the displayed expression. If the conditional treatment means vary with history, replace \(\mu_A,\mu_B\) by \(E\{Y_t(A)\mid H_{t-1}\}\) and \(E\{Y_t(B)\mid H_{t-1}\}\); the unconditional means alone are not enough.
Worked Example: Expected Reward With Changing Allocation Probabilities
Suppose a hypothetical 10-participant trial has independent Bernoulli rewards with stable means \(\mu_A=0.50\) and \(\mu_B=0.60\). The first four participants each have assignment probability \(q_t(B\mid H_{t-1})=0.50\). For the last six, the policy assigns \(B\) with probability \(0.80\), regardless of the realized history. These are hypothetical design values, and the assumed independence and stable means make the conditional-mean calculation appropriate.
The expected number assigned to \(B\) is \(4(0.50)+6(0.80)=6.8\); for \(A\), it is \(10-6.8=3.2\). Thus
Under equal allocation probability \(0.50\) for all 10 participants, the expected reward is \(5(0.60)+5(0.50)=5.50\). The adaptive probabilities increase expected reward by \(5.68-5.50=0.18\) successes. This is an expectation, not a promise that the realized adaptive trial will have more successes.
Why More Participant Benefit Need Not Mean More Precision
A common benchmark for precision is the standard error of a difference between two independent sample means. With fixed arm counts \(n_A,n_B\) and independent outcomes with variances \(\sigma_A^2,\sigma_B^2\), the variance of the difference is \(\sigma_A^2/n_A+\sigma_B^2/n_B\). For Bernoulli rewards with success probabilities \(p_A,p_B\), those variances are \(p_A(1-p_A)\) and \(p_B(1-p_B)\). This benchmark makes the allocation tradeoff visible: a small arm count can contribute substantially to the contrast’s uncertainty.
The calculation is exact for the stated fixed-count independent-sampling model. In an adaptive trial, arm counts depend on the evolving history, and simply substituting observed counts into this formula need not give the exact sampling uncertainty of an adaptive estimator. Here it is an allocation-based comparison under a clearly stated benchmark, not a general inferential procedure for adaptive data.
Worked Example: Higher Expected Benefit, Larger Benchmark Standard Error
Consider a hypothetical trial of 100 participants with independent Bernoulli outcomes, \(p_A=0.50\) and \(p_B=0.60\). Compare balanced allocation (50 per arm) with an 80/20 allocation favoring \(B\). For this benchmark, treat the arm counts as fixed. Both designs estimate the same contrast, \(p_B-p_A=0.10\).
With balanced allocation, the variance and standard error of the difference in sample proportions are
With 80 assigned to \(B\) and 20 to \(A\), the corresponding calculation is
Expected successes are \(50(0.60)+50(0.50)=55\) under balanced allocation and \(80(0.60)+20(0.50)=58\) under the 80/20 allocation. In this hypothetical comparison, favoring \(B\) yields three more expected successes, while the fixed-count benchmark standard error rises from 0.0990 to 0.1245. The result illustrates a possible tradeoff; it does not establish the exact uncertainty of an adaptive estimator.
Conditional Means Matter
The expected-reward mixture using \(\mu_A\) and \(\mu_B\) requires conditional-mean stability. Correlated participant outcomes can violate it even when treatment-specific marginal means are stable and one participant’s treatment does not affect another’s outcome. For example, a shared unobserved factor may raise or lower rewards for everyone. Earlier outcomes then reveal information about that factor, changing the expected reward for the next participant conditional on the history.
Worked Example: When Marginal Means Give the Wrong Conditional Calculation
Suppose a hypothetical shared factor \(U\) is low or high with probability \(0.50\) each. Conditional success probabilities are \(0.40\) on \(A\) and \(0.50\) on \(B\) when \(U\) is low, and \(0.80\) on \(A\) and \(0.90\) on \(B\) when \(U\) is high. The marginal means are \(\mu_A=(0.40+0.80)/2=0.60\) and \(\mu_B=(0.50+0.90)/2=0.70\). Assume the assignment decision uses a history that reveals \(U\), and assigns \(B\) with probability \(0.80\).
Given a high-factor history, the correct conditional expected reward is \(0.20(0.80)+0.80(0.90)=0.88\). Plugging in the marginal means instead gives \(0.20(0.60)+0.80(0.70)=0.68\), which is not the conditional expectation. The discrepancy occurs because the history changes the conditional means: they are \(0.80\) and \(0.90\) in this history. To calculate expected reward correctly, use the history-conditional means first, then average across histories.
This distinction is important for simulations and design calculations. A simulation with shared latent factors should reproduce that dependence; otherwise it may overstate or understate expected reward and mischaracterize how much the trial’s history predicts future outcomes. Independence is not the only possible model, but the relevant conditional outcome process must be specified.
Assessing the Tradeoff in Practice
A useful design comparison reports patient-oriented and inferential summaries side by side. Expected reward, expected number of successes, or expected utility describes participant outcomes under the chosen reward definition. Precision can be summarized by the sampling variability of a treatment-effect estimator or the typical width of its interval. Reliability requires more: uncertainty procedures should be checked for their behavior under the assignment rule and outcome model, rather than assumed valid from a familiar fixed-allocation formula.
| Question | Useful design summary | What it does not establish |
|---|---|---|
| How well do participants fare on average? | Expected cumulative reward or expected successes | That the treatment contrast is precisely estimated |
| How much information is available for comparison? | Arm counts and an appropriate estimate of contrast variability | That an ordinary fixed-allocation standard error is valid under adaptation |
| Are uncertainty statements dependable? | Repeated-design assessment of interval coverage and error rates | That one realized interval contains the true contrast with a known probability |
A fixed balanced design is a natural comparison when precise head-to-head information is important. A response-adaptive design may shift assignment toward treatments with promising observed outcomes, potentially improving cumulative reward, but its benefit depends on how accurately those outcomes identify better treatments and how much allocation imbalance results. A randomized adaptive policy retains chance variation in assignments; that randomization can protect against deterministic starvation of an arm, but does not by itself guarantee adequate precision or valid inference.
Useful diagnostics include tracking the number and proportion assigned to each arm over time, checking that each treatment retains enough assignments for the intended comparison, and examining whether the policy reacts strongly to small early samples. In design evaluation, simulate repeated trials under plausible treatment effects, outcome variability, dependence structures, and chance histories. Report both expected participant outcomes and inferential summaries. If conclusions change sharply across plausible assumptions, the design’s apparent advantage is fragile.
The main limitation is that patient reward and inferential performance cannot be reduced to a universal single score without making value judgments. The reward definition determines which outcomes count as benefit, while precision requirements depend on the scientific decision the trial is meant to support. A policy optimized for cumulative reward may be unsuitable when reliable comparative evidence is the primary objective. Conversely, an allocation chosen solely to balance information may not maximize outcomes for participants enrolled during the trial.
Common Mistake
Incorrect reasoning: “The treatments have stable marginal means and do not interfere with one another, so the expected reward conditional on the trial history is always the assignment-probability-weighted average of those marginal means.” This is false when earlier outcomes are informative about a shared factor that changes the next participant’s conditional treatment means.
Correction: Either assume explicitly that each treatment’s conditional mean given the history equals its marginal mean, with assignment independent of current potential outcomes given that history, or use the history-conditional treatment means in the reward calculation. Also keep the reward calculation separate from the precision calculation: better expected participant outcomes do not, on their own, demonstrate a more precise or reliable treatment-effect estimate.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- In a 20-person hypothetical trial, \(\mu_A=0.40\) and \(\mu_B=0.50\). If 12 participants are expected to receive \(B\) and 8 to receive \(A\), what is expected cumulative reward under stable conditional means?
- For independent Bernoulli outcomes with \(p_A=0.40\), \(p_B=0.50\), and fixed counts \(n_A=n_B=25\), calculate the benchmark standard error of the difference in sample proportions.
- Why can a shared latent factor invalidate a calculation that weights marginal treatment means by the assignment probabilities conditional on history?
- Name one patient-benefit measure and one inferential-performance measure that should be reported in a design comparison.
Solutions. (1) \(12(0.50)+8(0.40)=6.00+3.20=9.20\) expected rewards. (2) The variance is \(0.50(0.50)/25+0.40(0.60)/25=0.0100+0.0096=0.0196\), so the standard error is \(\sqrt{0.0196}=0.1400\). This is a fixed-count independent-outcome benchmark. (3) History can reveal the shared factor, changing the conditional treatment means; the correct mixture then uses those conditional means, not the marginal means. (4) Examples are expected cumulative successes for patient benefit and the sampling standard deviation or interval width for inferential precision. Reliability additionally requires checking whether uncertainty statements behave as intended under the design.