Burn-In Randomization Before Adaptation

Why Begin With Fixed Allocation?

Response-adaptive allocation uses accumulating trial outcomes to influence later assignments. Starting that process immediately can leave early allocation decisions dependent on very little information. A burn-in is an initial period in which allocation follows a prespecified rule that does not depend on observed responses. It creates a defined body of randomized data before response-driven allocation begins.

The burn-in is a design choice, not a guarantee that the early data will be precise or that adaptation will be beneficial. Its size determines how many participants are assigned under the fixed rule; the resulting information also depends on the endpoint, response frequency, number of arms, and prior assumptions. The practical task is to specify the period precisely and understand what its data can support.

Definition: A burn-in randomization period is a prespecified initial part of a trial during which participants are randomized using fixed allocation probabilities or a fixed allocation ratio, without using accumulating response outcomes to change those probabilities. Its boundary is defined in the protocol, commonly by a total number $B$ of randomized participants. The allocation rule may change only after that boundary and any stated data-availability conditions are met.

A protocol should state the total burn-in size $B$, the fixed allocation rule, and the intended number $m_a$ assigned to each arm $a$. For example, a 1:1:1 rule with $B=30$ can target 10 participants per arm. If exact counts matter, the randomization implementation must enforce them, such as by using permuted blocks; independent random assignment with equal probabilities targets equal counts but does not guarantee them. The protocol should also state what outcome data are eligible for the first adaptive decision. A count-based boundary alone does not ensure that all outcomes are observed at that point.

What Information Does the Burn-In Supply?

Consider a binary endpoint, such as response by a prespecified assessment time. For arm $a$, let $p_a$ be its response probability, $m_a$ the number of burn-in participants assigned to it, and $S_a$ the number who respond. Assume responses are independent Bernoulli observations within each arm, the response probability is constant for participants in that arm, and the prior for $p_a$ is Beta with parameters $\alpha_a$ and $\beta_a$. For the calculations below, priors across arms are independent.

Beta-Binomial update: Under these assumptions, if $p_a \sim \mathrm{Beta}(\alpha_a,\beta_a)$, then after $S_a=s_a$ responses among $m_a$ participants, $$ p_a \mid \text{burn-in data} \sim \mathrm{Beta}(\alpha_a+s_a,\ \beta_a+m_a-s_a). $$ The posterior mean and variance are $$ E(p_a \mid \text{data})=\frac{\alpha_a+s_a}{\alpha_a+\beta_a+m_a}, \qquad \operatorname{Var}(p_a \mid \text{data})= \frac{(\alpha_a+s_a)(\beta_a+m_a-s_a)} {(\alpha_a+\beta_a+m_a)^2(\alpha_a+\beta_a+m_a+1)}. $$

The update follows directly from multiplying prior and likelihood. The prior density is proportional to $p_a^{\alpha_a-1}(1-p_a)^{\beta_a-1}$, while the likelihood for the observed responses is proportional to $p_a^{s_a}(1-p_a)^{m_a-s_a}$. Their product has powers $p_a^{\alpha_a+s_a-1}$ and $(1-p_a)^{\beta_a+m_a-s_a-1}$, which is the kernel of the stated Beta posterior.

This calculation makes the role of burn-in size explicit: more observations contribute more likelihood information. If the observed response fraction stays near an interior value $q$ as $m_a$ grows, the posterior variance is approximately $q(1-q)/m_a$ for large $m_a$. That is an approximation, not a promise about every realized trial. A particular added observation can increase the realized posterior variance. Under the model, however, expected posterior uncertainty declines as data accrue: for current data $D$ and a subsequent observation $Y$, the law of total variance gives $$ \operatorname{Var}(p_a \mid D) = E\{\operatorname{Var}(p_a \mid D,Y)\mid D\} + \operatorname{Var}\{E(p_a\mid D,Y)\mid D\}. $$ The second term is nonnegative, so the expected posterior variance after observing $Y$ is no greater than the variance before observing it. The realized variance after a particular $Y$ need not be lower.

Worked Example: Defining a Burn-In for Three Arms

Suppose a hypothetical trial has three arms and plans a fixed 1:1:1 burn-in of $B=30$ participants. The binary endpoint is response by the scheduled assessment. The protocol uses permuted blocks to assign exactly 10 participants to each arm and permits response-based allocation only after all 30 are randomized and the prespecified endpoint data for those participants are available.

Assume independent $\mathrm{Beta}(1,1)$ priors. In this hypothetical burn-in, the observed responses are 6 of 10 in arm A, 4 of 10 in arm B, and 3 of 10 in arm C. The posteriors are $$ p_A\mid\text{data}\sim\mathrm{Beta}(7,5),\quad p_B\mid\text{data}\sim\mathrm{Beta}(5,7),\quad p_C\mid\text{data}\sim\mathrm{Beta}(4,8). $$ For example, arm A has $1+6=7$ posterior successes and $1+10-6=5$ posterior failures. Its posterior mean is $7/12\approx0.5833$, and its posterior variance is $$ \frac{7(5)}{12^2(13)}=\frac{35}{1872}\approx0.01870. $$ Similarly, arm B has posterior mean $5/12\approx0.4167$ and variance $35/1872\approx0.01870$; arm C has posterior mean $4/12=0.3333$ and variance $32/1872\approx0.01709$. These summaries describe the information available from this specified burn-in. They do not, by themselves, prescribe an allocation for the next participants.

The boundary is operationally explicit: assignment does not adapt during the first 30 participants, and the stated endpoint data must be available before the first response-driven decision. If the protocol instead meant “begin adapting as soon as 30 participants are randomized,” outcomes still pending at that point would not be observed responses and could not simply be counted as failures.

How Burn-In Size Changes the Evidence

A larger burn-in usually gives more observations per arm before an allocation update, but it also keeps the trial on its fixed rule for longer. In a balanced design with $K$ arms, an exact equal allocation requires $B$ to be divisible by $K$; otherwise, the protocol must specify how the remainder is assigned. For unequal allocation ratios, the same planning issue applies to the target counts. These details matter because a total $B$ does not alone reveal how much information is available for each arm.

Worked Example: Comparing Two Burn-In Sizes

Consider a hypothetical two-arm trial with independent $\mathrm{Beta}(1,1)$ priors and a fixed 1:1 burn-in. Compare data from 4 participants per arm with data from 12 per arm. Suppose the observed response fractions are 1/4 in control and 3/4 in treatment at the smaller size, and 3/12 in control and 9/12 in treatment at the larger size.

With 4 participants per arm, the posteriors are $\mathrm{Beta}(2,4)$ for control and $\mathrm{Beta}(4,2)$ for treatment. Their posterior means are $2/6=0.3333$ and $4/6=0.6667$; the difference in means is exactly $4/6-2/6=2/6=1/3\approx0.3333$, not the difference obtained by subtracting separately rounded values. Each posterior variance is $$ \frac{2(4)}{6^2(7)}=\frac{8}{252}\approx0.03175. $$

With 12 participants per arm, the posteriors are $\mathrm{Beta}(4,10)$ for control and $\mathrm{Beta}(10,4)$ for treatment. Their posterior means are $4/14\approx0.2857$ and $10/14\approx0.7143$, a difference of $6/14=3/7\approx0.4286$. Each posterior variance is $$ \frac{4(10)}{14^2(15)}=\frac{40}{2940}\approx0.01361. $$ In this hypothetical comparison, the larger burn-in provides more precise arm-specific estimates. The posterior mean difference also changes because the observed fractions differ from the smaller-sample posterior means. A larger burn-in does not guarantee a larger estimated treatment difference; its main design implication is more outcome data before the first adaptive decision.

The example illustrates why size should be judged against the information needed for the planned allocation rule, rather than selected only as a round number. If early responses are rare, a modest burn-in may yield very few successes and limited ability to distinguish arms. If there are many arms, a fixed total can leave few observations per arm. Conversely, a very large burn-in postpones response-driven allocation and may assign many participants without using the emerging response data.

Worked Example: Choosing a Fixed Ratio and Boundary

A hypothetical four-arm platform plans an initial total of 40 participants. Its protocol specifies a fixed 2:1:1:1 allocation ratio during burn-in, using blocks that produce exactly 16 participants in the designated control arm and 8 in each experimental arm. The binary primary response is assessed at a fixed time. The protocol allows the first response-adaptive allocation only after all 40 have been assigned and the primary-response records for the burn-in cohort have been finalized.

This design fixes both the total and the arm-specific exposure: $16+8+8+8=40$. It supplies twice as many control observations as observations in any one experimental arm, which can be appropriate if precise control estimation is an explicit design priority. It also means each experimental arm has less burn-in information than control. The protocol should therefore not describe this as “40 participants per arm” or imply equal precision across arms. The fixed ratio determines assignment, while the posterior information still depends on the realized outcomes and the model.

A valid specification would also define the endpoint window, the handling of missing or unevaluable responses, and the exact data cutoff for the first adaptive decision. These rules prevent the meaning of “burn-in complete” from changing after outcomes are seen.

Design Checks, Alternatives, and Limitations

Before implementation, check that the burn-in rule is reproducible and that the planned information is relevant to the later allocation decision. In particular, verify that the arm-specific counts match the stated ratio, that the randomization system does not use responses during the burn-in, and that outcome eligibility at the boundary is clear. Examine plausible response rates and prior settings to see whether the resulting posterior uncertainty is likely to be useful. For a complex adaptive rule, simulation under plausible scenarios can reveal whether the proposed burn-in leaves the rule unstable or delays adaptation excessively.

A burn-in can fail to provide the intended information if endpoint definitions vary across sites, assessments are missing, follow-up is incomplete, or response probabilities differ materially across patient groups. A simple independent Bernoulli model can also be inadequate if outcomes are clustered or the endpoint is not binary. In those cases, the posterior precision calculated under the simple model may overstate what the data genuinely establish. The model and data-quality assumptions should be reviewed rather than treating a prespecified participant count as evidence of adequacy.

The main alternatives have different trade-offs:

Approach	Potential advantage	Main trade-off
Immediate response adaptation	Uses accumulating outcomes to alter assignments from the outset.	Early updates may rely on little information and can produce imbalanced exposure.
Fixed-allocation burn-in, then adaptation	Provides a prespecified initial sample under a stable assignment rule before responses influence allocation.	Delays response-driven allocation, and its information may still be insufficient or unrepresentative.
Fixed allocation throughout	Keeps assignment probabilities independent of accumulating responses.	Does not use response data to change later allocation probabilities.

These are design alternatives, not universally ranked choices. The appropriate burn-in depends on the trial’s scientific and operational aims, the number of arms, the endpoint and its observation schedule, and the allocation rule planned after the boundary. It cannot eliminate uncertainty, protect against model misspecification, or ensure adequate evidence for every arm.

Common Mistakes

Defining only a total, not the rule. “Burn in 30 participants” is incomplete if the allocation ratio, endpoint-data requirement, and transition condition are unclear.
Assuming a larger burn-in must reduce the realized posterior variance. More data reduce expected posterior variance under the model, but a particular new response can increase the realized variance.
Equating a fixed allocation ratio with equal arm information. Unequal counts, different response patterns, and different prior specifications can yield different posterior precision.
Treating pending outcomes as observed nonresponses. A response not yet recorded is not a failure. The protocol must define the data eligible for the first update.
Choosing a burn-in solely to make an early response difference look decisive. The burn-in size and transition rule should be prespecified, not selected after inspecting outcomes.

Key takeaway: A burn-in is a prespecified fixed-allocation period that creates an initial set of randomized observations before responses can influence allocation. Specify its total size, arm-specific allocation, and data boundary; then evaluate the information it is expected to provide without assuming any particular realized posterior result.

Check Your Understanding

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

In a three-arm trial, a protocol specifies $B=24$ and exact equal allocation. How many participants are assigned to each arm during burn-in?
With a $\mathrm{Beta}(1,1)$ prior, an arm has 5 responses among 10 burn-in participants. Give its posterior distribution and posterior mean.
Why is a burn-in of 30 randomized participants not necessarily equivalent to 30 participants with observed endpoint outcomes at the first adaptive decision?
Under the Bernoulli-Beta model, does every additional observed response necessarily reduce the realized posterior variance? Explain.
Name two design details, beyond total burn-in size, that should be prespecified.

Solutions. (1) $24/3=8$ participants per arm. (2) The posterior is $\mathrm{Beta}(1+5,1+10-5)=\mathrm{Beta}(6,6)$, with mean $6/12=0.5$. (3) Some endpoint assessments may still be pending or records may not be finalized when the 30th participant is randomized; the protocol must specify what data are eligible for the update. (4) No. The expected posterior variance is no greater than the variance before the observation, by the law of total variance, but a particular response can increase the realized variance. (5) Examples include the fixed allocation ratio, exact arm counts or randomization blocks, endpoint assessment window, missing-outcome handling, and the data cutoff required before adaptation starts.