Continuous Outcomes and Model-Based Response Estimates

Why Continuous Responses Need an Estimation Step

For a continuous endpoint such as symptom change, blood pressure, or time to recovery, a treatment response is not a success probability. A DBCD still needs current response estimates to form its allocation target, but those estimates may be means on the endpoint's scale or fitted quantities from a model. The important distinction is that the response estimate supplies an input to a prespecified target rule; it is not itself the target allocation.

This tutorial considers two treatment arms. Let $A_j\in\{1,2\}$ denote the treatment assigned to patient $j$, and let $Y_j(k)$ denote that patient's potential continuous outcome under arm $k$. The observed outcome is $Y_j=Y_j(A_j)$, once it is available. As in the earlier material on sequential updates, the allocation decision for a new patient uses information available before that assignment.

Definition: At allocation time $t$, let $O_k(t)$ be the set of patient indices assigned to arm $k$ whose outcomes have been observed by time $t$. The arm-specific observed sample mean is $$ \bar Y_k(t)=\frac{1}{|O_k(t)|}\sum_{j\in O_k(t)}Y_j $$ when $|O_k(t)|>0$. The index set is defined by outcome availability, not by assignment order. A model-based response estimate is a fitted quantity, such as a covariate-standardized arm mean, obtained from the observed outcome records and a stated model.

If all outcomes are available before the next allocation, then $O_k(t)$ consists of all patients assigned to arm $k$ by that time. If follow-up is delayed, it may be only a subset. Writing the set explicitly prevents an unobserved outcome from being included in an estimate just because that patient was assigned earlier.

Assumptions Before Using the Estimates

The response estimates have a meaningful interpretation only in relation to the outcome and the trial's data process. The outcome should be defined consistently for both arms, including its measurement scale, assessment time, and direction of benefit. If a larger value is worse, the target rule must account for that; a mean estimate does not determine the direction of allocation on its own.

For a causal comparison of treatment responses, the usual trial assumptions include randomized assignment with known probabilities conditional on the information used by the design, consistency of the observed outcome with the assigned treatment's potential outcome, and no interference between patients. The response estimates also require outcome data that are available and measured as intended. If outcomes are missing or delayed in a way related to their unobserved values, the observed records may not represent the assigned patients. In that case, the observed-case mean can be misleading unless the data process is addressed with an appropriate model or other prespecified method.

Adaptive randomization does not make every estimate automatically reliable. Assignment probabilities depend on the trial's history, and a sample mean based on a small or selectively observed set can be unstable. For model-based estimates, the assumed conditional mean structure must be adequate for the purpose at hand, and the data must contain enough covariate and treatment overlap to estimate the arm contrast. The allocation rule must also have a defined initialization for times when one arm has no observed outcomes.

Worked Example: A Mean With Delayed Outcomes

Consider a hypothetical trial at time $t$. Four patients have been assigned to arm 1, but only patients with arm-specific indices $1,3,4$ have observed outcomes: 8, 12, and 10 units. Two patients have observed outcomes in arm 2, with indices $1,2$ and values 7 and 9 units. The remaining assigned outcomes are not yet available.

For arm 1, $O_1(t)=\{1,3,4\}$, so $$ \bar Y_1(t)=\frac{8+12+10}{3}=\frac{30}{3}=10\text{ units}. $$ For arm 2, $O_2(t)=\{1,2\}$, so $$ \bar Y_2(t)=\frac{7+9}{2}=8\text{ units}. $$ The observed-case difference is $10-8=2$ units in favor of arm 1 if larger outcomes are favorable. It is an estimate from the outcomes available at time $t$, not proof that the population mean responses differ by exactly 2 units. In particular, it would be incorrect to include an unavailable value or to assume the observed patients are the first patients assigned to their arms.

From Response Estimates to a Two-Arm Allocation

Write $q_1(t)$ and $q_2(t)=1-q_1(t)$ for the current target allocation shares supplied by the trial's prespecified target rule. The rule may use continuous-response estimates, but it must specify how estimates on the outcome scale are translated into target shares. Let $N_k(t)$ be the number assigned to arm $k$ by time $t$, and let $\pi_k(t)=N_k(t)/t$ be the current allocation proportion when $t>0$. For tuning parameter $\gamma\geq 0$, the two-arm DBCD assigns the next patient to arm 1 with probability $$ P(A_{t+1}=1\mid\mathcal F_t)= \frac{q_1(t)\left(q_1(t)/\pi_1(t)\right)^\gamma} {q_1(t)\left(q_1(t)/\pi_1(t)\right)^\gamma+ q_2(t)\left(q_2(t)/\pi_2(t)\right)^\gamma}, $$ where $\mathcal F_t$ denotes the information available for the allocation decision. The arm 2 probability is one minus this probability. The response estimate affects this calculation through the target $q_k(t)$; the current allocation proportion is a separate input.

The rule presupposes that the target shares and current proportions are defined. In practice, initialization handles early states with no response estimates or no assignments to an arm. It is also important that the observed response estimates and the target map be available before the next assignment. Replacing an unavailable outcome with a guess, or allowing future information into the decision, changes the design.

Why the Ratio Correction Restores the Target

The ratio form makes the restoring direction visible. Dividing the arm 1 weight by the arm 2 weight gives $$ \frac{P(A_{t+1}=1\mid\mathcal F_t)} {P(A_{t+1}=2\mid\mathcal F_t)} = \left(\frac{q_1(t)}{q_2(t)}\right)^{\gamma+1} \left(\frac{\pi_2(t)}{\pi_1(t)}\right)^\gamma. $$ To obtain this expression, take the ratio of the two unnormalized weights: the target terms contribute $q_1/q_2$, and the correction terms contribute $\{(q_1/\pi_1)/(q_2/\pi_2)\}^{\gamma}$. Multiplying yields the displayed odds. Thus, if arm 1 is below its target share, its relative correction increases its assignment odds; if it is above target, the correction decreases them. When $\gamma=0$, the next assignment probability equals the target share, regardless of current imbalance. For positive $\gamma$, current imbalance affects the probability as well.

This restoring behavior concerns allocation proportions, not the accuracy of the response estimates. A mistaken, noisy, or delayed-data-biased target can still be tracked by the rule. The allocation mechanism cannot repair a poor response estimate merely by correcting the allocation ratio.

Worked Example: Fitting and Standardizing a Continuous-Outcome Model

Suppose a hypothetical trial records a baseline covariate $X$ and an outcome $Y$, both measured in compatible units for this example. The observed records are $(A=1,X,Y)=(1,0,10),(1,1,12),(1,2,14)$ and $(A=2,X,Y)=(2,2,12),(2,3,14),(2,4,16)$. A linear model with a common covariate slope is $$ E(Y\mid A,X)=\beta_0+\beta_1 I(A=1)+\beta_2X. $$ These records lie exactly on the fitted relation $Y=8+2I(A=1)+2X$, so the least-squares estimates are $\hat\beta_0=8$, $\hat\beta_1=2$, and $\hat\beta_2=2$. The raw arm means are 12 units in arm 1 and 14 units in arm 2, but the covariate distributions differ.

Standardize both fitted arm means to the same reference value $X=2$. The estimates are $$ \hat\mu_1(2)=\hat\beta_0+\hat\beta_1+\hat\beta_2(2)=8+2+4=14\text{ units}, $$ $$ \hat\mu_2(2)=\hat\beta_0+\hat\beta_2(2)=8+4=12\text{ units}. $$ On this common covariate reference, the fitted arm 1 mean is 2 units higher. This comparison is conditional on the model and reference value; the crude means answer a different, unadjusted descriptive question. The common-slope model is identifiable in these data, but its exact fit in this small hypothetical example does not demonstrate that a linear mean structure is suitable in a real trial.

Worked Allocation Using a Continuous-Response Target

The model estimates in the preceding example do not by themselves specify an allocation target. Suppose, for illustration, that the trial's prespecified target map takes the current response estimates and yields $q_1=0.60$ and $q_2=0.40$. Assume ten patients have been assigned, five to each arm, so $\pi_1=\pi_2=0.50$, and use $\gamma=1$. Then $$ P(A_{11}=1\mid\mathcal F_{10})= \frac{0.60(0.60/0.50)} {0.60(0.60/0.50)+0.40(0.40/0.50)} = \frac{0.72}{0.72+0.32} = \frac{0.72}{1.04} =0.6923. $$

The next patient therefore has a 0.6923 probability of assignment to arm 1 and a 0.3077 probability of assignment to arm 2. The probability exceeds the target share of 0.60 because arm 1 is currently below its target proportion. The target share is not itself the next-patient probability except in special states or when $\gamma=0$.

Choosing Between Means and Models

An observed arm-specific mean is transparent and requires relatively few modeling choices. It is often a reasonable response estimate when the outcome is measured consistently, observed cases are representative for the relevant estimand, and covariate adjustment is not needed for the target calculation. Its weakness is that it can be noisy with few observed outcomes and can reflect differing covariate compositions across arms.

A fitted model can use baseline covariates to improve precision or place arm estimates on a common covariate distribution. Its advantage depends on the model being appropriate: an incorrect functional form, omitted treatment-by-covariate interaction when effects vary, or extrapolation beyond covariate overlap can distort the fitted response. A common-slope model, for example, assumes the covariate-outcome relationship is the same in both arms. When that assumption is implausible, a model allowing arm-specific slopes or interactions may be more suitable, provided the available data support fitting it.

Neither approach is universally preferable. The analysis should identify the estimand (for example, an observed arm mean or a standardized mean), the outcome set used at each update, the model if one is fitted, and how the estimates enter the target rule. Allocation estimates are operational inputs, not automatically the final treatment-effect analysis. The latter may require methods that account for the sequential design and missing outcomes.

Diagnostics, Failure Modes, and Limitations

Check which outcomes are included. At each update, verify that the estimate uses exactly the observed outcomes eligible under the stated rule. With delayed follow-up, assignment order and observation order need not agree.
Monitor availability by arm. A large difference in pending outcomes, especially if delay is associated with outcome or treatment, can make observed-case means unrepresentative. A rapidly changing estimate as delayed outcomes arrive is a warning to review the data process.
Inspect the model, not only its coefficients. Check residual patterns, influential records, covariate overlap, and whether the fitted mean structure is plausible. Estimates based on extrapolating one arm into covariate regions seen only in the other are particularly fragile.
Separate instability from a genuine response change. Early estimates based on few observations can move sharply. A target map that reacts strongly to small estimate changes may amplify random fluctuation into allocation changes.
Keep the target and the next-patient probability distinct. The target $q_1(t)$ describes a desired allocation share at the current estimated response state. The DBCD probability also depends on the current allocation proportion and $\gamma$.

A further limitation is that treatment assignment and response estimation form a feedback system: early outcomes affect later allocation, and later allocation determines which additional responses are observed. The ratio correction regulates allocation imbalance relative to the current target; it does not ensure unbiased response estimation, correct a misspecified model, or guarantee balanced covariates in a small trial. Continuous outcomes also have scale and distributional features—such as outliers or heavy tails—that can make means unstable. Those features should be considered when selecting the response summary and the prespecified target map.

Common Mistake

Common pitfall: “Four patients have been assigned to arm 1, and three outcomes are observed, so the observed mean is the mean of the first three patients assigned.” This is not justified when outcome times differ. Define $O_1(t)$ as the set of indices with outcomes actually observed by time $t$, and average $Y_j$ over that set. Separately assess whether the observed cases are representative; correct indexing prevents a data error but does not eliminate bias from informative delay or missingness.

Check Your Understanding

Answer each question, then compare with the short solutions below.

At time $t$, arm 2 has observed outcomes 5, 8, and 11 for indices in $O_2(t)=\{1,3,4\}$. What is the observed arm 2 mean, and why should the index set be stated?
A common-slope model is fitted as $E(Y\mid A,X)=\beta_0+\beta_1I(A=1)+\beta_2X$. Write the fitted arm means standardized to $X=x^\star$.
If $q_1=0.70$, $q_2=0.30$, current proportions are $\pi_1=0.50$, $\pi_2=0.50$, and $\gamma=0$, what is the next-patient probability for arm 1?
Name one reason an observed-case mean may be misleading when outcomes are delayed, and one diagnostic for a fitted model.

Solutions. (1) $\bar Y_2(t)=(5+8+11)/3=8$ units. The set identifies the observed records and avoids assuming observed patients are the first assigned. (2) The fitted means are $\hat\mu_1(x^\star)=\hat\beta_0+\hat\beta_1+\hat\beta_2x^\star$ and $\hat\mu_2(x^\star)=\hat\beta_0+\hat\beta_2x^\star$. (3) With $\gamma=0$, the correction factors equal 1, so the probability is $q_1=0.70$. (4) If delay depends on outcome, observed cases may not represent assigned patients; a model diagnostic could examine residual patterns or covariate overlap.