Why Initialization Matters
A doubly adaptive biased coin design (DBCD) uses both an allocation target and the observed allocation counts to determine the probability of assigning the next participant. At the beginning of a trial, however, there may be no trial response estimates, and there may be no participants allocated to one or both arms. Initialization supplies a starting state from which the adaptive rule can be used; it is not a minor bookkeeping detail.
The target and the counts play different roles. A prespecified target, often obtained from current response estimates through a target map, describes the desired allocation shares. The allocation counts describe what has actually happened in the trial. The DBCD ratio correction uses both. Consequently, a starting response estimate cannot by itself replace a positive allocation count, and a positive allocation count does not by itself provide a response estimate.
This rule increases an arm's weight when its observed allocation share is below its target share, and reduces its weight when its share is above target. The initialization problem is to choose a defensible starting target and allocation history so the rule can be applied without confusing estimates, planned assignments, and actual assignments.
What Must Be Initialized
Two pieces of information are needed for the first adaptive assignment. First, both arms must have positive allocation counts, so the allocation shares \(n_i/N\) are defined and their negative powers are finite. Second, the target shares must be available. If the target is response-driven, this usually means that the design must specify how initial response estimates are obtained or what target to use until such estimates are available.
The estimates used to obtain a target can come from trial run-in participants, external information, or a prespecified neutral starting value. For binary outcomes, an estimate might be written \(\hat p_i\), the estimated response probability for arm \(i\). The target map then turns the available estimates into \(\rho_1\) and \(\rho_2\). The map is a design choice: for example, a target proportional to estimated response probabilities is one possible rule, not a universal requirement. The initialization protocol should state the map and the source of its inputs before assignments begin.
Allocation counts are different. In the ordinary DBCD rule above, \(n_i\) counts participants actually assigned to arm \(i\). External pilot participants do not become trial allocation counts merely because their outcomes inform \(\hat p_i\). Likewise, a prior or a pseudo-count may inform a response estimate, but it is not an actual trial assignment and should not be inserted into \(n_i\) unless the design explicitly defines a modified rule.
Valid Starting Strategies
A common approach is a fixed run-in: assign an initial group using a prespecified scheme, often balanced, and begin adaptive assignments only after both arms have participants. The run-in provides actual positive counts and may also provide trial response data. It separates the initial assignments from adaptation, so the protocol must say how many participants are in the run-in and when adaptation starts. If the run-in is randomized, its assignment probabilities should also be specified; it is not itself governed by a DBCD state that has not yet been established.
A second approach is to use external information for the initial response estimates, while establishing positive trial counts through a run-in or other prespecified initial allocation scheme. External data can be useful when relevant and sufficiently comparable, but its population, outcome definition, and uncertainty matter. An apparently precise estimate from a different setting may produce a misleading initial target if the response rates do not transport to the trial.
A third approach is a neutral start. The design can begin with a target that does not favor either arm, or use a prespecified target based on a common initial estimate, while a run-in establishes positive counts. Neutral does not necessarily mean that the final target is equal allocation: it means that the initial target is chosen without allowing unsupported early response estimates to create a strong preference. The appropriate choice depends on the scientific target and the trial protocol.
These approaches can be combined. For example, a balanced run-in may establish the counts while an external estimate supplies the first target. Alternatively, the run-in outcomes may supply initial response estimates, provided the protocol defines how those estimates map to a target. In each case, the first adaptive assignment is calculated from the same DBCD formula; what differs is the source and credibility of the starting state.
Worked Example: Balanced Run-In and a Prespecified Target
Suppose a hypothetical two-arm trial has assigned 4 participants to each arm in a fixed run-in. The protocol specifies the first adaptive target as \(\rho_1=0.60\), \(\rho_2=0.40\), and uses \(\gamma=1\). Find the probability of assigning the next participant to arm 1.
The run-in gives \(n_1=n_2=4\) and \(N=8\), so both allocation shares are positive. The target and counts are therefore available for the DBCD calculation: $$ \pi_1= \frac{0.60(4/8)^{-1}} {0.60(4/8)^{-1}+0.40(4/8)^{-1}} = \frac{0.60(2)} {0.60(2)+0.40(2)} = \frac{1.20}{2.00} =0.600. $$
The next participant has a 0.600 probability of assignment to arm 1 and a 0.400 probability of assignment to arm 2. Because the run-in counts are balanced, the correction does not change the target shares at this state. This calculation does not claim that 60% of all future participants will necessarily be assigned to arm 1; it gives the probability for this next assignment, conditional on the initialized state.
Why Positive Counts Are Required
The need for positive counts follows directly from the ratio form of the DBCD rule. Dividing the arm 1 weight by the arm 2 weight gives $$ \frac{\pi_1}{\pi_2} = \frac{\rho_1}{\rho_2} \left(\frac{n_1}{n_2}\right)^{-\gamma} = \frac{\rho_1}{\rho_2} \left(\frac{n_2}{n_1}\right)^\gamma. $$
This expression is finite and positive when both counts and both target shares are positive. If \(n_1=0\) and \(\gamma>0\), the factor \((n_1/N)^{-\gamma}\) is undefined as a finite weight; informally treating it as infinite would force a degenerate assignment rather than implement the intended finite-probability rule. If a target share is zero, the target ratio also requires special handling and is outside the positive-target formula stated here. A positive run-in count in each arm and a target strictly between zero and one avoid these problems.
When \(\gamma=0\), the count factors equal one for positive shares, so the rule assigns with probabilities equal to the target shares. That algebraic fact does not remove the practical need for a well-defined initialization protocol: response estimates may still be unavailable, and the design still needs to state how its target is selected. For \(\gamma>0\), positive counts are essential to compute the ratio correction directly.
Worked Example: Unequal Starting Counts Correct Toward the Target
Suppose a hypothetical run-in leaves \(n_1=6\) and \(n_2=2\), so \(N=8\). The initial target is \(\rho_1=0.60\), \(\rho_2=0.40\), and \(\gamma=1\). Calculate the next assignment probability.
Arm 1 has an observed share of \(6/8=0.75\), above its target; arm 2 has a share of \(2/8=0.25\), below its target. The DBCD weights and probability are $$ w_1=0.60(6/8)^{-1}=0.60(8/6)=0.800, \qquad w_2=0.40(2/8)^{-1}=0.40(4)=1.600, $$ $$ \pi_1=\frac{w_1}{w_1+w_2} =\frac{0.800}{0.800+1.600} =\frac{1}{3} \approx 0.333. $$
The next participant is assigned to arm 1 with probability about 0.333 and arm 2 with probability about 0.667. The probability for arm 1 is below its target share of 0.60 because arm 1 is already overrepresented in the run-in. This is the restoring direction of the biased coin, not a change in the target itself.
Starting Estimates and Starting Counts Are Not Interchangeable
A response estimate can be used to determine a target, but it does not establish the allocation history. Conversely, the fact that several participants were assigned to an arm does not guarantee that its response estimate is informative: outcomes may be missing, delayed, or highly variable. The initialization plan should therefore record these items separately: actual counts by arm; the source and value of each initial response estimate; the target map; the resulting target shares; and the value of \(\gamma\).
Early estimates can be unstable because they are based on few observations. One response can produce a large change in a raw proportion, especially when an arm has only one or two observed outcomes. A practical design can reduce the influence of this instability by specifying a neutral or externally informed starting target, by requiring a run-in before adaptation, or by prespecifying a rule for when response information is considered available. These are design choices, not corrections that can be improvised after seeing early results.
Worked Example: Trial Estimates Used With a Specified Target Map
In a hypothetical binary-response trial, a run-in has assigned 4 participants to arm 1 and 2 to arm 2. There are 2 responses among the 4 arm 1 participants and 1 response among the 2 arm 2 participants, giving \(\hat p_1=2/4=0.50\) and \(\hat p_2=1/2=0.50\). For this example only, suppose the protocol defines the target as proportional to these estimates, so \(\rho_i=\hat p_i/(\hat p_1+\hat p_2)\), and sets \(\gamma=1\). Find the next assignment probability.
The stated target map gives \(\rho_1=0.50/(0.50+0.50)=0.50\) and \(\rho_2=0.50\). The allocation counts are \(n_1=4\), \(n_2=2\), and \(N=6\), so $$ \pi_1= \frac{0.50(4/6)^{-1}} {0.50(4/6)^{-1}+0.50(2/6)^{-1}} = \frac{0.50(1.5)} {0.50(1.5)+0.50(3)} = \frac{0.75}{2.25} =0.333. $$
Although the response-based target is equal allocation, arm 1 already has twice as many trial assignments as arm 2. The ratio correction therefore gives the next participant a one-third probability of assignment to arm 1. This example illustrates the distinction between the target (derived here from response estimates) and the probability for the next participant. It also relies on the example's explicitly stated target map; another scientific objective could prescribe a different map.
Comparing Initialization Choices and Checking the Starting State
| Approach | How it supplies a starting state | Main consideration |
|---|---|---|
| Fixed run-in | Creates positive counts from actual trial assignments; may also yield trial estimates | Requires a prespecified size and assignment scheme before adaptation begins |
| External information | Supplies response estimates or a starting target; counts still need to come from trial assignments | Relevance to the trial population and uncertainty of the external information must be assessed |
| Neutral start | Uses a prespecified non-favoring estimate or target while counts are established | May avoid strong early adaptation, but the neutral choice must be defined in advance |
Before using the first adaptive probability, check that the specified counts are actual assignments and that \(n_1>0\) and \(n_2>0\); verify that the target shares sum to one and are both positive; confirm that any response estimates came from the stated source; and recompute the weights and their normalization. Sensitivity calculations are useful when external estimates or a target map are uncertain: compare plausible starting targets at the same counts and \(\gamma\), and report whether the first probability changes materially.
A run-in has a different purpose from adding fictitious counts to the formula. A run-in creates an observed allocation history. A pseudo-count may be a device in an estimate, but treating it as if a participant had been randomized changes the allocation correction and therefore changes the design. Such a modification is not automatically invalid, but it must be defined and evaluated as part of the design rather than silently substituted for the ordinary DBCD counts.
Common Mistakes and Limitations
- Starting with a zero count. The negative-power correction is not finite at a zero share when \(\gamma>0\). Use a prespecified initialization that assigns at least one participant to each arm before applying this formula.
- Using an estimate as if it were a count. An external response estimate can inform the target, but it does not represent a trial assignment. Keep target inputs and actual allocation counts distinct.
- Assuming a balanced run-in makes the target equal. Balanced counts make the correction leave the specified target unchanged for that state; they do not determine what the target should be.
- Letting a noisy early estimate dictate the target without a stated rule. A small amount of response data can produce an extreme estimate. Prespecify the estimate source and target map, and examine sensitivity to plausible starting values.
- Interpreting a next-assignment probability as a final allocation proportion. The calculated \(\pi_1\) is conditional on the current counts, target, and \(\gamma\). Later probabilities can differ as the allocation history and target change.
Initialization cannot guarantee that early response estimates are stable, that external evidence is transportable, or that realized allocation proportions will equal the target at every point in the trial. It can make the starting state transparent, mathematically defined, and reproducible. That is especially important because early assignments affect the counts used by subsequent ratio corrections.
Check Your Understanding
Answer each question, then compare your work with the solutions below.
- A trial has \(n_1=3\), \(n_2=3\), target \((\rho_1,\rho_2)=(0.70,0.30)\), and \(\gamma=1\). What is \(\pi_1\)?
- With \(n_1=8\), \(n_2=2\), target \((0.50,0.50)\), and \(\gamma=1\), calculate \(\pi_1\). Explain the direction of the correction.
- Why can an external response estimate help initialize a DBCD while failing to supply a valid \(n_i\)?
- A design has \(n_1=0\), \(n_2=5\), and \(\gamma=1\). Can the stated DBCD formula be used directly? Give the reason and one design-level remedy.
Solutions. (1) The counts are balanced, so the common count factor cancels: \(\pi_1=0.70/(0.70+0.30)=0.700\). (2) Here \(N=10\). The weights are \(w_1=0.50(8/10)^{-1}=0.625\) and \(w_2=0.50(2/10)^{-1}=2.50\), so \(\pi_1=0.625/(0.625+2.50)=0.200\). Arm 1 is overrepresented relative to its 0.50 target, so its next-assignment probability is reduced. (3) External information may supply an estimated response rate used to construct the target, but \(n_i\) counts actual trial assignments. The external observations were not assigned in this trial. (4) No. The factor \((n_1/N)^{-1}\) is undefined at zero. A prespecified run-in that assigns at least one participant to each arm before adaptive assignment is one remedy.