Why Clipping Estimated Targets Changes the Design

Why a Small Target Adjustment Is a Design Change

A plug-in target is an input to a response-adaptive randomization rule, not merely a summary printed after the assignments. If an estimated target is very small, an implementation may replace it with a positive floor so that every arm retains a specified minimum target share. That can be a sensible safeguard, but it does not preserve the original target: it defines a different target map and therefore a different allocation rule.

This tutorial makes that change explicit. We define a clipping rule that both enforces a lower bound and preserves a total target share of one, derive why its scale factor exists uniquely, and calculate its effect on DBCD assignment probabilities. The distinction between a target share and the probability of the next assignment is essential: clipping changes the former, and the DBCD coin rule then uses that changed input to determine the latter.

A Simplex-Preserving Lower-Floor Map

Suppose there are $K$ treatment arms, and the plug-in target vector is $\mathbf{t}=(t_1,\ldots,t_K)$, where $t_k\geq 0$ and $\sum_{k=1}^K t_k=1$. Choose a floor $\epsilon$ satisfying $0<\epsilon<1/K$. The restriction on $\epsilon$ is necessary: if every arm must have target share at least $\epsilon$, the shares can sum to one only if $K\epsilon\leq 1$. The strict inequality leaves room for at least one share to exceed the floor.

Definition: The simplex-preserving clipped target is $\mathbf{C}_{\epsilon}(\mathbf{t})=(C_{\epsilon,1}(\mathbf{t}),\ldots,C_{\epsilon,K}(\mathbf{t}))$, where $$ C_{\epsilon,k}(\mathbf{t})=\max\{\epsilon,a(\mathbf{t})t_k\}, \qquad \sum_{k=1}^K\max\{\epsilon,a(\mathbf{t})t_k\}=1. $$ The common scale factor $a(\mathbf{t})$ is chosen to make the clipped components sum to one. Thus each component is at least $\epsilon$, and the resulting vector remains a target-share vector.

The scaling is part of the definition. Simply replacing each $t_k$ with $\max(\epsilon,t_k)$ will generally make the components sum to more than one. Normalizing those values afterward is another possible rule, but it can bring a component below $\epsilon$. The map above instead enforces both requirements simultaneously: a lower bound and a unit sum.

Why the Scale Factor Exists and Is Unique

For $a\geq 0$, define $F(a)=\sum_{k=1}^K\max\{\epsilon,at_k\}$. Each term, and hence $F$, is continuous and nondecreasing in $a$. At $a=0$, $F(0)=K\epsilon<1$. At least one $t_k$ is positive because the components of $\mathbf{t}$ sum to one; therefore $F(a)$ eventually exceeds one as $a$ increases. Continuity gives at least one solution of $F(a)=1$.

For uniqueness, let $a_*$ be any solution. Since $F(0)=K\epsilon<1=F(a_*)$, at least one component, say $j$, must be strictly above the floor at $a_*$: $a_*t_j>\epsilon$. If $a>a_*$, that component strictly increases and no other component decreases, so $F(a)>F(a_*)=1$. If $0\leq a<a_*$, component $j$ is strictly smaller than its value at $a_*$, and all other components are no larger, so $F(a)<F(a_*)=1$. Thus no other $a$ solves the equation.

Key result: For any target vector on the simplex and any $0<\epsilon<1/K$, there is exactly one scale factor $a(\mathbf{t})$ that gives the clipped target. At that solution, at least one arm is above the floor; otherwise the clipped shares would sum to $K\epsilon<1$.

The arms strictly above the floor have shares proportional to their original target components: for any two such arms $j$ and $\ell$, $C_{\epsilon,j}(\mathbf{t})/C_{\epsilon,\ell}(\mathbf{t})=t_j/t_\ell$, provided the original components are positive. Arms at the floor no longer retain their original relative proportions. This is why the result is a modified target rather than a harmless numerical adjustment.

Worked Example: Finding the Scale Factor

Consider the hypothetical three-arm target $\mathbf{t}=(0.70,0.20,0.10)$ and floor $\epsilon=0.15$. The floor is feasible because $3(0.15)=0.45<1$. The scale factor must satisfy $$ \max(0.15,0.70a)+\max(0.15,0.20a)+\max(0.15,0.10a)=1. $$

Suppose the third arm is at the floor while the first two are above it. Under that active-set assumption, the equation becomes $0.70a+0.20a+0.15=1$, so $a=0.85/0.90=0.9444$. Substitution checks the assumed cases: $0.70a=0.6611>0.15$, $0.20a=0.1889>0.15$, and $0.10a=0.09444<0.15$. Therefore the clipped target is $(0.6611,0.1889,0.1500)$, whose components sum to $1.000$. The original target assigned 10% to arm 3; the modified target assigns it 15%, with the other target shares reduced to keep the total at one.

From the Modified Target to DBCD Probabilities

To make the design change concrete, write $x_k=N_k(n)/n$ for the current allocation share of arm $k$ after $n$ assignments, where $N_k(n)$ is its assignment count. For positive current shares, a standard DBCD rule with tuning parameter $\gamma>0$ uses target $\mathbf{t}$ to assign the next participant to arm $k$ with probability. At $\gamma=0$, define $\pi_k(\mathbf{t},\mathbf{x})=t_k$.

\pi_k(\mathbf{t},\mathbf{x}) = \frac{t_k(t_k/x_k)^\gamma} {\sum_{\ell=1}^K t_\ell(t_\ell/x_\ell)^\gamma}.

When the plug-in target is clipped, the rule uses $\mathbf{C}_{\epsilon}(\mathbf{t})$ in place of $\mathbf{t}$:

\pi_k\bigl(\mathbf{C}_{\epsilon}(\mathbf{t}),\mathbf{x}\bigr) = \frac{C_{\epsilon,k}(\mathbf{t}) \left(C_{\epsilon,k}(\mathbf{t})/x_k\right)^\gamma} {\sum_{\ell=1}^K C_{\epsilon,\ell}(\mathbf{t}) \left(C_{\epsilon,\ell}(\mathbf{t})/x_\ell\right)^\gamma}.

This expression assumes the current shares are positive; a design must specify how its allocation rule is initialized before every arm has a positive count. The clipping map itself changes the target, not the coin rule. But because the coin rule uses target shares in its numerator and in its allocation-ratio correction, changing those shares can change the next-assignment probabilities.

Worked Example: The Same Current Shares, a Different Coin

Suppose a hypothetical two-arm trial has estimated target $\mathbf{t}=(0.90,0.10)$, current shares $\mathbf{x}=(0.50,0.50)$, and $\gamma=1$. Without clipping, the unnormalized DBCD weights are $0.90(0.90/0.50)=1.62$ and $0.10(0.10/0.50)=0.0200$. The probability of assigning the next participant to arm 1 is $1.62/(1.62+0.0200)=0.9878$.

Now use $\epsilon=0.20$. The clipped-target equation is $\max(0.20,0.90a)+\max(0.20,0.10a)=1$. With arm 2 at the floor, $0.90a+0.20=1$, giving $a=0.80/0.90=0.8889$. The resulting target is $(0.8000,0.2000)$, since $0.10a=0.08889<0.20$. Its unnormalized weights are $0.80(0.80/0.50)=1.28$ and $0.20(0.20/0.50)=0.0800$. The next-assignment probability for arm 1 is $1.28/(1.28+0.0800)=0.9412$. Thus the floor changes the target share for arm 2 from 0.10 to 0.20 and, in this state, changes the probability of assigning to arm 1 from 0.9878 to 0.9412. The target share and the next-assignment probability are different quantities.

Interpreting the Resulting Target

If the plug-in target is itself an estimate of an ideal allocation, clipping replaces that estimated ideal with a constrained target. The floor can protect against extremely small target shares, but the procedure does not merely guarantee that every arm has a positive chance of assignment. It changes the allocation objective supplied to the DBCD rule. In particular, when a component is below the floor after scaling, that arm receives target share $\epsilon$, regardless of how much smaller its original estimate was.

In a response-adaptive design, let $\hat{\boldsymbol{\rho}}_n$ denote the estimated target based on available response data. Applying the map means the rule at that stage uses $\mathbf{C}_{\epsilon}(\hat{\boldsymbol{\rho}}_n)$, not $\hat{\boldsymbol{\rho}}_n$. If the estimates settle near some vector $\boldsymbol{\rho}$, the clipping map describes the target to which the plug-in inputs are being transformed. It is important not to report $\boldsymbol{\rho}$ as the design's target without also stating whether the floor is applied.

Worked Example: Several Arms at the Floor

Consider the hypothetical target $\mathbf{t}=(0.80,0.15,0.05)$ with $\epsilon=0.20$. Since $3(0.20)=0.60<1$, the floor is feasible. If arms 2 and 3 are at the floor, the equation for the scale factor is $0.80a+0.20+0.20=1$, so $a=0.60/0.80=0.7500$.

The scaled values are $0.80a=0.6000$, $0.15a=0.1125$, and $0.05a=0.03750$. The last two are both below 0.20, confirming the assumed active set. The modified target is therefore $(0.6000,0.2000,0.2000)$, which sums to one. The two low-target arms are raised to equal shares even though their original target shares, 0.15 and 0.05, were unequal. The first arm is reduced from 0.80 to 0.60 to make room for both floors.

Diagnostics, Alternatives, and Limitations

An implementation should make the clipping operation auditable. For each update, record the raw plug-in target, the chosen floor, the scale factor, the clipped target, and the arms at the floor. Check that the raw target components are nonnegative and sum to one, that $0<\epsilon<1/K$, and that the resulting components sum to one up to numerical precision. After computing a candidate scale factor, verify each component against the assumed floor status, as in the worked examples. A solution based on an incorrect active set can produce a plausible-looking vector that is not the defined map.

A second diagnostic is to compare the raw and modified targets over time. Repeatedly floored arms or large differences between the two vectors indicate that clipping is an active part of the design, not an occasional numerical safeguard. The floor should be specified before the trial and included in design documentation. Choosing it after inspecting outcomes changes the allocation rule in a data-dependent way.

This lower-floor map is not the only stabilization strategy. A uniform-mixture rule, for example, replaces $\mathbf{t}$ by $(1-K\epsilon)\mathbf{t}+\epsilon\mathbf{1}$, where $\mathbf{1}$ is the vector of ones. It also sums to one and gives every arm at least $\epsilon$, but it moves every component toward the uniform allocation, not only components below a threshold. Another approach is to truncate components and renormalize; unless specifically designed to preserve the floor after normalization, that procedure may fail to leave every final component at least $\epsilon$. These are distinct target maps and should not be treated as interchangeable implementations of the same rule.

Clipping has limitations. It can substantially alter the target when several estimates are small, and it may hide instability in the underlying response estimates if only the clipped values are monitored. The floor also does not make the estimate more accurate: it changes how the design reacts to it. Finally, a target floor is not a promise that an arm will receive exactly that fraction of participants. DBCD assignment probabilities depend on current allocation shares and the coin parameter as well as on the target.

Common Mistake

Common mistake: “I replaced every estimated target below $\epsilon$ by $\epsilon$, so the result is the same target with a safety floor.” The replacement may no longer sum to one; normalizing it afterward can also push a component back below $\epsilon$. The correction is to define the full map, including the common scale factor, and then use that modified vector as the DBCD target. Report both the raw and clipped target when explaining the design.

Check Your Understanding

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

For $K=4$, is a floor $\epsilon=0.26$ feasible? Explain.
For $\mathbf{t}=(0.60,0.30,0.10)$ and $\epsilon=0.20$, arm 3 is assumed to be at the floor and the other two above it. Find the scale factor and clipped target, and check the assumption.
Why does the uniqueness argument need more than the statement that $F(a)$ is nondecreasing?
If a clipped target share is 0.20, must the next-assignment probability for that arm be 0.20? Explain.

Solutions. (1) No. Feasibility requires $K\epsilon<1$ for this rule, but $4(0.26)=1.04>1$, so four shares each at least 0.26 cannot sum to one. (2) The equation is $0.60a+0.30a+0.20=1$, giving $a=0.80/0.90=0.8889$. The scaled values are 0.5333, 0.2667, and 0.08889, so the assumed active set is valid and the target is $(0.5333,0.2667,0.2000)$, summing to one. (3) A nondecreasing function can be flat over an interval, so monotonicity alone does not rule out multiple solutions. At a solution, at least one component is above the floor; that component makes the sum strictly larger for any larger scale and strictly smaller for any smaller scale. (4) No. The target is an input to the DBCD coin rule; the next-assignment probability also depends on the current allocation shares and $\gamma$, followed by normalization across arms.