Building an Estimator-Driven Allocation Target

From Response Estimates to Desired Shares

An estimator-driven allocation target is a rule for translating current estimates of treatment response into desired proportions of the trial assigned to each arm. This translation is a distinct design choice: estimated responses do not, by themselves, determine allocation shares. A target map must specify how the estimates are to be used, along with any fixed design inputs such as weights or tuning constants.

The map’s output describes a target composition, not necessarily the assignment probability for the next patient. In an adaptive design, a separate allocation rule may use the current arm counts and that target to determine the next assignment probability. Here the focus is the target map: its inputs, output, and interpretation, without prescribing that separate rule or choosing a particular optimality criterion.

The response parameters might be success probabilities, mean outcomes, or parameters from a prespecified response model. The map’s fixed inputs might include exponents, arm-specific constants, or other quantities chosen as part of the design. The map must say exactly which estimates and constants it uses and how they produce shares. Calling a target “response-adaptive” without specifying that translation leaves the allocation objective operationally incomplete.

Inputs, Output, and Assumptions

A clear specification begins with the estimand. State what each component of \(\theta\) means, including the outcome definition, measurement scale, and population or patient group to which it refers. Then state how \(\hat{\theta}_n\) is obtained and which observations it uses. If, for example, a component is a binary-response probability, specify which outcome counts as a success. Otherwise, even a correctly calculated estimate may not correspond to the response quantity intended by the design.

Next specify the map itself, including all fixed constants and its domain. The map must be defined for the estimates it could receive during the design, not merely for a preferred final estimate. Its output must be a valid vector of shares. If the formula uses nonnegative weights, the normalization must have a positive denominator; if it uses a transformation such as a logarithm, its inputs must lie in that transformation’s domain. Any exceptional cases need an explicit handling rule.

The map is a design specification, not a model for estimating responses. Its use therefore relies on the response estimates having a defensible interpretation under the chosen study design and response model. In particular, the outcome must be recorded as specified, the estimator must match the parameter being targeted, and relevant modeling assumptions must be checked for that estimator. A map cannot repair biased, unstable, or poorly defined input estimates.

The particular optimality criterion, if any, is a separate design decision. Earlier material on optimal allocation targets shows that different goals can lead to different targets. Once a target has been chosen, the map expresses how estimates feed into it; the mere use of estimated responses does not establish that the resulting shares are ethically preferable, more precise, or optimal.

Plug-In Targets and a Basic Justification

The plug-in construction evaluates a prespecified target function at the estimated rather than unknown response parameters. If \(T(\theta;d)\) is the target that would be calculated from the true parameter vector, the current estimated target is \(T(\hat{\theta}_n;d)\). This is a definition of the estimator-driven target; it is not a guarantee that the estimates, or the target, are already accurate.

A simple normalization explains how many maps produce valid shares. Suppose the map first assigns each arm a nonnegative weight \(w_i(\hat{\theta}_n;d)\), with at least one weight positive, and then divides each weight by the total. The resulting shares are nonnegative, and their sum is one:

This algebra verifies the share constraint when its stated conditions hold; it does not justify the choice of weights. For example, a designer could choose weights to express a previously selected allocation goal, but the normalization alone supplies no ethical or statistical rationale for those weights.

There is also a useful large-sample justification for plug-in evaluation. Suppose the estimator is consistent, so \(\hat{\theta}_n\) converges in probability to \(\theta\), and suppose \(T(\cdot;d)\) is continuous at \(\theta\) for the fixed design input \(d\). Then the continuous mapping theorem gives:

Thus, under these assumptions, the estimated target approaches the target evaluated at the true response parameters. This conclusion is conditional on consistency and continuity; it does not say that early targets are stable, that the estimator is unbiased at every sample size, or that the chosen map is desirable. Nor does convergence of the target alone establish properties of a separate sequential assignment rule.

Diagnostics, Alternatives, and Limitations

A target map should be checked both as a mathematical function and as part of a statistical design. Useful checks include:

• Input definition: Confirm that each estimated parameter has an unambiguous outcome, scale, and arm interpretation, and that the estimator matches that definition.
• Domain: Verify that every possible input used by the map is allowed by its formula. A map involving a logarithm or square root, for example, may fail at some boundary values.
• Output: Check that shares are nonnegative and sum to one, and that normalization denominators cannot be zero for inputs the design permits.
• Numerics: Evaluate using adequate numerical precision, then round only for reporting. Rounding intermediate weights can change displayed shares, as can happen when arms have similar weights.
• Estimation behavior: Inspect whether response estimates are based on enough information to be meaningful and whether the specified response model is plausible. A formally valid map can still yield unstable targets from unstable estimates.

One practical diagnostic is to evaluate the map at a range of plausible response estimates, not only at a single preferred estimate. This can reveal whether small estimate changes produce target changes that are important for the design. It does not require choosing a different map, but it makes the consequences of the chosen transformation visible. If a map is undefined for plausible inputs, the design needs an explicit remedy rather than an implicit assumption that the problematic inputs will never occur.

The main alternative at the target level is to specify fixed shares that do not depend on estimated responses. A fixed target avoids variation in the target caused by changing estimates, but it does not respond to estimated response differences. An estimator-driven target makes that dependence explicit, while inheriting uncertainty and possible model error from its inputs. A further design choice is whether to use a direct transformation of estimates, such as a power or exponential map, or to encode estimates and constants in weights before normalization. Either approach must be stated fully; neither is automatically superior.

A common mistake is to treat a target map as if it were itself a complete randomization procedure. For example, seeing \(\hat{\rho}_1=0.60\) and concluding that the next patient must have assignment probability 0.60 for arm 1 confuses the target composition with the next-patient allocation probability. The correction is to report what the map actually returns—a desired share—and specify separately how the design uses that share to make assignments. Another mistake is to assume that a map becomes optimal merely because it uses current response estimates. Optimality requires a stated criterion and a justification under that criterion.