main17.pdf
This paper has been submitted for consideration for publication in Biometrics
Introduction
Randomized clinical trials provide a standard basis for evaluating treatment efficacy but can require substantial time and resources. Two complementary strategies can improve their efficiency. Covariate-adaptive randomization balances prognostic covariates across randomized arms and can improve estimation efficiency without additional enrollment [3 ] . Historical-control borrowing increases the amount of control information by reusing data from completed studies (Pocock, 1976; Viele et al., 2014). This strategy is particularly useful when concurrent-control recruitment is ethically or logistically difficult (Ghadessi et al., 2020). Integrating these strategies makes historical borrowing part of trial design, so the borrowed information guides both sequential allocation and final inference according to the degree of compatibility between historical and concurrent controls.
A central consideration in historical borrowing is systematic bias arising from differences in participant selection and trial conduct between historical and concurrent populations. Such bias can compromise Type I error control and the interpretation of the final analysis (Viele et al., 2014; Calderazzo, Wiesenfarth, and Kopp-Schneider, 2024). This concern has motivated a large literature on Bayesian dynamic borrowing where borrowing is regulated through one or more parameters that are prespecified, estimated empirically, or assigned a hyperprior within a fully Bayesian framework (Lesaffre et al., 2024). Prominent examples include the power prior, which scales the historical likelihood (Ibrahim et al., 2015); the commensurate prior, which captures between-trial heterogeneity through a precision hyperparameter (Hobbs et al., 2011); and the unit information prior, which measures borrowed information in single-observation Fisher-information units (Jin and Yin, 2021; Zhang and Yin, 2023; Gu and Yin, 2024). Other adaptive approaches, including self-adapting mixture and elastic priors, use data-driven compatibility functions to regulate cohort-level borrowing (Yang et al., 2023; Jiang, Nie, and Yuan, 2023). A common feature of these approaches is that they assess overall compatibility between historical and concurrent control cohorts and apply the resulting borrowing decision uniformly across participants.
The comparability of historical and concurrent control populations may vary substantially across clinically meaningful patient subgroups, however, even when the studies investigate the same disease and treatment. In osteoporosis trials, for example, vertebral fracture risk depends on prior fracture history, years since menopause, bone mineral density, and other indicators of underlying skeletal vulnerability (Black et al., 2007). Although patients can be matched on these observed prognostic characteristics, residual differences in disease severity, prior treatments, and other unmeasured risk factors may lead to varying degrees of comparability between historical and concurrent controls. Consequently, historical control information may be highly informative for some patient subgroups but less relevant, or potentially misleading, for others. This heterogeneity motivates a flexible borrowing strategy that accounts for subgroup-specific comparability rather than assuming uniform exchangeability across the study population.
When compatibility between historical and concurrent controls differs across participant groups, a global rule may borrow too little for participants whose historical and concurrent data are compatible and thereby lose efficiency. It may also borrow too much for participants whose data differ between the two sources, which can bias treatment-effect estimates. The Covariate-Adaptive Historical Borrowing (CAHB) design of Jin et al. (2023) integrates kernel-based covariate-adaptive design (KBCD) [3 ] with a Bayesian semiparametric model that allows compatibility to vary across the covariate space.
CAHB uses the estimated historical conditional mean as a fixed prior center, so the borrowing model does not propagate its sampling uncertainty. As a result, CAHB may borrow too much information from an imprecisely estimated historical mean, potentially biasing the treatment-effect estimate when historical and concurrent controls differ. To account for historical estimation uncertainty in local borrowing, we propose Robust Adaptive Discrepancy-Informed Shrinkage for Historical Borrowing (RADISH), a covariate-adaptive clinical trial design that integrates historical borrowing into sequential randomization and final inference. RADISH accounts explicitly for the sampling uncertainty of the estimated historical conditional mean and evaluates its discrepancy from the concurrent control mean relative to the standard errors of both estimates. This allows borrowing to depend on both the precision of historical information and its compatibility with concurrent outcomes. Motivated by two-sided conflict tail probabilities (Marshall and Spiegelhalter, 2007; Presanis et al., 2013), RADISH uses a covariate-specific conflict score to discount historical information as local disagreement increases, and the resulting information ratio guides sequential treatment allocation. At trial completion, a marginal conflict score provides an additional safeguard for borrowing in the final analysis to ensure its robustness. For final inference, we use a kernel-standardized estimator for the average treatment effect in the enrolled cohort, with a variance estimator based on kernel-regression residuals and linear-smoother weights. This approach provides analytical standard errors and confidence intervals. Under the stated regularity conditions, the centered and studentized RADISH estimator converges to a standard normal distribution, supporting asymptotically valid Wald inference.
The remainder of the paper is organized as follows. § 1.2 develops the RADISH trial design, covering the three-stage design and its large-sample properties. § 2 compares RADISH with KBCD and CAHB through simulation studies. § 2.2 presents semi-synthetic simulation studies based on data from the HORIZON pivotal fracture trial (Black et al., 2007) and fracture intervention trial (FIT) (Black et al., 1996; Cummings et al., 1998). § 2.3 concludes and outlines extensions to non-Gaussian endpoints, multi-source borrowing, and covariate overlap. Code is publicly available at https://github.com/BobZhangHT/HistInfoBorrow .
Methodology: The RADISH Design
Notation and Overview
RADISH is designed for a two-arm randomized clinical trial that sequentially enrolls participants from a target population. It uses an external historical control dataset to inform both treatment allocation and final inference. Let D H = { ( X j H , Y j H ) } j = 1 N H \mathcal{D}_H = \{(\mathbf{X}_j^H, Y_j^H)\}_{j=1}^{N_H} D H = {( X j H , Y j H ) } j = 1 N H denote the historical control data, where X j H ∈ R d \mathbf{X}_j^H \in\mathbb{R}^d X j H ∈ R d is a baseline covariate vector and Y j H ∈ R Y_j^H \in\mathbb{R} Y j H ∈ R is a continuous outcome. Let D n = { ( X i , Y i , Z i ) } i = 1 n \mathcal{D}_n = \{(\mathbf{X}_i, Y_i, Z_i)\}_{i=1}^n D n = {( X i , Y i , Z i ) } i = 1 n denote the concurrent trial data observed up to interim step n n n . The treatment indicator is Z i = 1 Z_i = 1 Z i = 1 for the treatment arm and Z i = 0 Z_i = 0 Z i = 0 for the concurrent control arm. The superscript H H H distinguishes historical observations from concurrent ones throughout. We write I z , n = { i ≤ n : Z i = z } \mathcal{I}_{z,n} = \{i \leq n : Z_i = z\} I z , n = { i ≤ n : Z i = z } for the index set of arm z ∈ { 0 , 1 } z \in\{0,1\} z ∈ { 0 , 1 } at step n n n , and N z N_z N z for its final size at n = N n = N n = N , so that N 0 + N 1 = N N_0 + N_1 = N N 0 + N 1 = N .
Let Y i ( 1 ) Y_i(1) Y i ( 1 ) and Y i ( 0 ) Y_i(0) Y i ( 0 ) denote the potential outcomes under treatment and control for participant i i i in the current trial. We assume consistency, so that Y i = Z i Y i ( 1 ) + ( 1 − Z i ) Y i ( 0 ) Y_i = Z_iY_i(1) + (1 - Z_i)Y_i(0) Y i = Z i Y i ( 1 ) + ( 1 − Z i ) Y i ( 0 ) , and write the conditional potential-outcome mean as μ z ( x ) = E { Y i ( z ) ∣ X i = x } \mu_z(x) = \mathbb{E}\{Y_i(z) \mid X_i = x\} μ z ( x ) = E { Y i ( z ) ∣ X i = x } for z ∈ { 0 , 1 } z \in\{0,1\} z ∈ { 0 , 1 } .
Our primary causal estimand is the average treatment effect standardized to the enrolled trial cohort,
Δ N = 1 N ∑ i = 1 N { μ 1 ( X i ) − μ 0 ( X i ) } . (1) \Delta_N = \frac{1}{N}\sum_{i=1}^{N}\{\mu_1(X_i) - \mu_0(X_i)\}. \tag*{(1)} Δ N = N 1 i = 1 ∑ N { μ 1 ( X i ) − μ 0 ( X i )} . (1)
This is a sample-specific estimand defined by the covariate distribution of the participants enrolled in the trial. Conditional on X 1 , … , X N \mathbf{X}_1,\ldots,\mathbf{X}_N X 1 , … , X N , Δ N \Delta_N Δ N is fixed, so the variance describes randomization and outcome uncertainty for the enrolled covariates.
The design proceeds in three stages. Stage I constructs a covariate-specific historical reference and quantifies its uncertainty. Stage II compares this reference with the accrued concurrent controls and uses the resulting local borrowing information to guide sequential randomization. After enrollment is complete, Stage III combines concurrent and discounted historical control information to estimate Δ N \Delta_N Δ N and conduct Wald inference. Supplementary Figure S1 and Algorithm 1 summarize this procedure.
For local borrowing, we adopt a Gaussian product kernel K h K_h K h with bandwidth vector h ∈ R + d \boldsymbol{h} \in\mathbb{R}_{+}^{d} h ∈ R + d . For a covariate profile x \boldsymbol{x} x , define w j H ( x ) = K h ( X j H − x ) w_j^H(\boldsymbol{x}) = K_h(\boldsymbol{X}_j^H - \boldsymbol{x}) w j H ( x ) = K h ( X j H − x ) and w i ( x ) = K h ( X i − x ) w_i(\boldsymbol{x}) = K_h(\boldsymbol{X}_i - \boldsymbol{x}) w i ( x ) = K h ( X i − x ) , so that observations with covariates closer to x \boldsymbol{x} x receive greater weight. The same kernel is used to construct the historical reference, summarize concurrent outcomes, determine local arm information in the randomization rule, and form the final kernel-standardized estimator. The bandwidth is computed at the planning stage and then fixed across all three stages. We use Scott’s rule of thumb (Scott, 2015), h k = N H − 1 / ( d + 4 ) σ ^ H , k h_k = N_H^{-1/(d+4)}\widehat{\sigma}_{H,k} h k = N H − 1/ ( d + 4 ) σ H , k , where σ ^ H , k \widehat{\sigma}_{H,k} σ H , k is the historical sample standard deviation of covariate k k k . If σ ^ H , k < 10 − 12 \widehat{\sigma}_{H,k} < 10^{-12} σ H , k < 1 0 − 12 , replace it by one before calculating h k h_k h k . Throughout, the similarity weights are evaluated conditionally on the observed covariates, including the historical covariates { X j H } j = 1 N H \{\boldsymbol{X}_j^H\}_{j=1}^{N_H} { X j H } j = 1 N H and, after enrollment, the concurrent covariates { X i } i ≤ n \{\boldsymbol{X}_i\}_{i\le n} { X i } i ≤ n . Thus, conditional on these covariates, the kernel weights are fixed, whereas treatment assignments and outcomes remain random.
Within a local neighborhood of a covariate profile x \boldsymbol{x} x defined by the kernel weights, we adopt the working Gaussian model
Y ∣ Z = z , X ≈ x ∼ N { μ z ( x ) , σ z 2 ( x ) } , (2) Y \mid Z = z,\ X \approx x \sim\mathcal{N}\left\{\mu_z(x), \sigma_z^2(x)\right\}, \tag*{(2)} Y ∣ Z = z , X ≈ x ∼ N { μ z ( x ) , σ z 2 ( x ) } , (2)
where μ z ( x ) \mu_z(\boldsymbol{x}) μ z ( x ) and σ z 2 ( x ) \sigma_z^2(\boldsymbol{x}) σ z 2 ( x ) denote the local conditional mean and variance in arm z ∈ { 0 , 1 } z \in\{0,1\} z ∈ { 0 , 1 } . For variance estimation and the large-sample theory, we further assume homoskedasticity within each concurrent arm, so that σ z 2 ( x ) ≡ σ z 2 \sigma_z^2(\boldsymbol{x}) \equiv\sigma_z^2 σ z 2 ( x ) ≡ σ z 2 for z ∈ { 0 , 1 } z \in\{0,1\} z ∈ { 0 , 1 } .
Historical controls provide an external reference for the concurrent control mean μ 0 ( x ) \mu_0(\boldsymbol{x}) μ 0 ( x ) . RADISH summarizes the historical data locally by a reference center θ ( x ) \theta(\boldsymbol{x}) θ ( x ) and its estimation variance τ H 2 ( x ) = Var { θ ( x ) } \tau_H^2(\boldsymbol{x}) = \operatorname{Var}\{\theta(\boldsymbol{x})\} τ H 2 ( x ) = Var { θ ( x )} . The quantity τ H 2 ( x ) \tau_H^2(\boldsymbol{x}) τ H 2 ( x ) characterizes the uncertainty in the estimated historical reference, whereas σ 0 2 \sigma_0^2 σ 0 2 characterizes the variability of individual concurrent control outcomes. The construction of θ ( x ) \theta(\boldsymbol{x}) θ ( x ) and τ H 2 ( x ) \tau_H^2(\boldsymbol{x}) τ H 2 ( x ) is described in Stage I.
Stage I: Construction of the Local Historical Reference
Stage I summarizes the historical control data as a covariate-specific reference for the concurrent control mean. We assume that the historical outcomes are conditionally independent given their covariates and adopt the local-constant Gaussian working model Y j H ∣ X j H ≈ x ∼ N { μ H ( x ) , σ H 2 ( x ) } Y_j^H \mid X_j^H \approx x \sim\mathcal{N}\{\mu_H(x),\sigma_H^2(x)\} Y j H ∣ X j H ≈ x ∼ N { μ H ( x ) , σ H 2 ( x )} . Under this model, maximizing the kernel-weighted log-likelihood yields the local historical mean
θ ( x ) = ∑ j = 1 N H S H , j ( x ) Y j H , S H , j ( x ) = w j H ( x ) ∑ ℓ = 1 N H w ℓ H ( x ) \theta(x) = \sum_{j=1}^{N_H} S_{H,j}(x)Y_j^H,\qquad S_{H,j}(x) = \frac{w_j^H(x)}{\sum_{\ell=1}^{N_H} w_\ell^H(x)} θ ( x ) = j = 1 ∑ N H S H , j ( x ) Y j H , S H , j ( x ) = ∑ ℓ = 1 N H w ℓ H ( x ) w j H ( x )
Thus, θ ( x ) \theta(x) θ ( x ) is the Nadaraya–Watson estimator of μ H ( x ) \mu_H(x) μ H ( x ) .
Under the working model, the reference variance is τ H 2 ( x ) = σ H 2 ( x ) ∑ j S H , j ( x ) 2 \tau_H^2(x) = \sigma_H^2(x)\sum_j S_{H,j}(x)^2 τ H 2 ( x ) = σ H 2 ( x ) ∑ j S H , j ( x ) 2 . We use the weighted residual variance σ ^ H 2 ( x ) = ∑ j S H , j ( x ) { Y j H − θ ( x ) } 2 \widehat{\sigma}_H^2(x) = \sum_j S_{H,j}(x)\{Y_j^H-\theta(x)\}^2 σ H 2 ( x ) = ∑ j S H , j ( x ) { Y j H − θ ( x ) } 2 and set τ ^ H 2 ( x ) = σ ^ H 2 ( x ) ∑ j S H , j ( x ) 2 \widehat{\tau}_H^2(x) = \widehat{\sigma}_H^2(x)\sum_j S_{H,j}(x)^2 τ H 2 ( x ) = σ H 2 ( x ) ∑ j S H , j ( x ) 2 . For a fixed local outcome variance, broadly distributed weights give a more precise historical reference than weights concentrated on a few observations. After each calculation, local and residual variance estimates are floored at ϵ v = 10 − 12 \epsilon_v = 10^{-12} ϵ v = 1 0 − 12 for numerical stability. Algorithm S1 (Supplementary § 5.7 ) constructs the historical reference { θ ( x ) , τ ^ H 2 ( x ) } \{\theta(x),\widehat{\tau}_H^2(x)\} { θ ( x ) , τ H 2 ( x )} used in subsequent stages.
Stage II: Randomization with Robust Borrowing
For each incoming covariate profile x x x , Stage II discounts historical precision using a local historical–concurrent conflict score motivated by prior-data conflict diagnostics (Evans and Moshonov, 2006; Nott et al., 2020). Sequential updates use observed outcomes from previously enrolled participants. To assess compatibility, we calculate local concurrent-control summaries using the kernel function, conditional on the observed control covariates { X i } i ∈ I 0 , n \{X_i\}_{i\in I_{0,n}} { X i } i ∈ I 0 , n . Define the local kernel mass in arm z ∈ { 0 , 1 } z \in\{0,1\} z ∈ { 0 , 1 } by M z , n ( x ) = ∑ i ∈ I z , n w i ( x ) M_{z,n}(x) = \sum_{i\in I_{z,n}} w_i(x) M z , n ( x ) = ∑ i ∈ I z , n w i ( x ) . For i ∈ I 0 , n i \in\mathcal{I}_{0,n} i ∈ I 0 , n , write S 0 , n , i ( x ) = w i ( x ) / M 0 , n ( x ) S_{0,n,i}(x) = w_i(x)/M_{0,n}(x) S 0 , n , i ( x ) = w i ( x ) / M 0 , n ( x ) . The local summaries are
Y ˉ 0 , n ( x ) = ∑ i ∈ I 0 , n S 0 , n , i ( x ) Y i , n 0 , n e f f ( x ) = { ∑ i ∈ I 0 , n S 0 , n , i ( x ) 2 } − 1 , \bar{Y}_{0,n}(x) = \sum_{i \in\mathcal{I}_{0,n}} S_{0,n,i}(x)Y_i,\qquad n^{\mathrm{eff}}_{0,n}(x) = \left\{\sum_{i \in\mathcal{I}_{0,n}} S_{0,n,i}(x)^2\right\}^{-1}, Y ˉ 0 , n ( x ) = i ∈ I 0 , n ∑ S 0 , n , i ( x ) Y i , n 0 , n eff ( x ) = ⎩ ⎨ ⎧ i ∈ I 0 , n ∑ S 0 , n , i ( x ) 2 ⎭ ⎬ ⎫ − 1 ,
σ ^ 0 , n 2 ( x ) = ∑ i ∈ I 0 , n S 0 , n , i ( x ) { Y i − Y ˉ 0 , n ( x ) } 2 . \widehat{\sigma}^2_{0,n}(x) = \sum_{i \in\mathcal{I}_{0,n}} S_{0,n,i}(x)\left\{Y_i-\bar{Y}_{0,n}(x)\right\}^2. σ 0 , n 2 ( x ) = i ∈ I 0 , n ∑ S 0 , n , i ( x ) { Y i − Y ˉ 0 , n ( x ) } 2 .
We standardize the difference between θ ( x ) \theta(x) θ ( x ) and Y ˉ 0 , n ( x ) \bar{Y}_{0,n}(x) Y ˉ 0 , n ( x ) using a reference-predictive standard error that combines historical and concurrent uncertainty. Specifically, we define
S E n 2 ( x ) = τ ^ H 2 ( x ) + σ ^ 0 , n 2 ( x ) / n ~ 0 , n e f f ( x ) , n ~ 0 , n e f f ( x ) = max { n 0 , n e f f ( x ) , n min } , \mathrm{SE}_n^2(x) = \widehat{\tau}_H^2(x) + \widehat{\sigma}_{0,n}^2(x)/\widetilde{n}^{\mathrm{eff}}_{0,n}(x),\qquad\widetilde{n}^{\mathrm{eff}}_{0,n}(x) = \max\{n^{\mathrm{eff}}_{0,n}(x), n_{\min}\}, SE n 2 ( x ) = τ H 2 ( x ) + σ 0 , n 2 ( x ) / n 0 , n eff ( x ) , n 0 , n eff ( x ) = max { n 0 , n eff ( x ) , n m i n } ,
where n min = 5 n_{\min}=5 n m i n = 5 stabilizes the concurrent sampling-variance term in sparse neighborhoods. The first term in S E n 2 ( x ) \mathrm{SE}_n^2(x) SE n 2 ( x ) reflects uncertainty in the historical reference, while the second reflects uncertainty in the concurrent local control mean. We then define the local conflict score as
D n ( x ) = ( − log [ max { 2 Φ ( − ∣ Y ˉ 0 , n ( x ) − θ ( x ) ∣ S E n ( x ) ) , ϵ c } ] − 1 ) + (3) D_n(x) = \left(-\log\left[\max\left\{2\Phi\left(-\frac{\left|\bar{Y}_{0,n}(x)-\theta(x)\right|}{\mathrm{SE}_n(x)}\right),\epsilon_c\right\}\right]-1\right)_+ \tag*{(3)} D n ( x ) = ( − log [ max { 2Φ ( − SE n ( x ) Y ˉ 0 , n ( x ) − θ ( x ) ) , ϵ c } ] − 1 ) + (3)
where Φ ( ⋅ ) \Phi(\cdot) Φ ( ⋅ ) is the standard normal cumulative distribution function, ( u ) + = max ( u , 0 ) (u)_+ = \max(u,0) ( u ) + = max ( u , 0 ) , and ϵ c = 10 − 12 \epsilon_c = 10^{-12} ϵ c = 1 0 − 12 is a numerical safeguard. The score increases with standardized disagreement; subtracting one and truncating at zero leaves small discrepancies unpenalized.
RADISH combines the local conflict score D n ( x ) D_n(x) D n ( x ) and the precision of the historical reference into an allocation-time information ratio. At step n n n , using only information accrued up to that point, we define
R n A ( x ) = 1 + σ ^ 0 , n 2 ( x ) n ~ 0 , n e f f ( x ) S E n 2 ( x ) exp { − D n ( x ) } . R_n^A(x) = 1 + \frac{\widehat{\sigma}_{0,n}^2(x)}{\widetilde{n}^{\mathrm{eff}}_{0,n}(x)\,\mathrm{SE}_n^2(x)}\exp\{-D_n(x)\}. R n A ( x ) = 1 + n 0 , n eff ( x ) SE n 2 ( x ) σ 0 , n 2 ( x ) exp { − D n ( x )} .
The information ratio R n A ( x ) R_n^A(x) R n A ( x ) controls borrowing in the randomization rule, with R n A ( x ) = 1 R_n^A(x)=1 R n A ( x ) = 1 corresponding to no borrowing. Its increment above one combines historical precision with the conflict discount exp { − D n ( x ) } \exp\{-D_n(x)\} exp { − D n ( x )} , reducing borrowing as historical uncertainty or local conflict increases; see Supplementary § 5.4 .
The information ratio R n A ( x ) R_n^A(x) R n A ( x ) is then incorporated into the randomization rule by augmenting the local control information. Using the local arm masses M 0 , n ( x ) M_{0,n}(x) M 0 , n ( x ) and M 1 , n ( x ) M_{1,n}(x) M 1 , n ( x ) , we assign the next participant to the treatment arm with probability
ϕ n ( x ) = { R n A ( x ) M 0 , n ( x ) } 2 { R n A ( x ) M 0 , n ( x ) } 2 + { M 1 , n ( x ) } 2 . (4) \phi_n(\mathbf{x}) = \frac{\{R_n^A(\mathbf{x})M_{0,n}(\mathbf{x})\}^2}{\{R_n^A(\mathbf{x})M_{0,n}(\mathbf{x})\}^2+\{M_{1,n}(\mathbf{x})\}^2}. \tag*{(4)} ϕ n ( x ) = { R n A ( x ) M 0 , n ( x ) } 2 + { M 1 , n ( x ) } 2 { R n A ( x ) M 0 , n ( x ) } 2 . (4)
This allocation rule extends the kernel-weighted covariate-balancing scheme of Jiang et al. [3 ] by incorporating the allocation-time borrowing ratio, following the covariate-adaptive borrowing construction of Jin et al. (2023). When R n A ( x ) > 1 R_n^A(\mathbf{x}) > 1 R n A ( x ) > 1 , borrowed information enlarges the effective control arm and the rule assigns fewer participants to the control group. As R n A ( x ) → 1 R_n^A(\mathbf{x}) \to1 R n A ( x ) → 1 , it reduces to the kernel-balanced allocation of Jiang et al. [3 ] . Allocation uses the local discrepancy at the incoming participant’s covariate profile. At trial completion, Stage III reassesses compatibility using all concurrent controls and incorporates a marginal conflict score. Supplementary Algorithm S2 details sequential allocation.
Stage III: Final Estimation and Inference for the Treatment Effect
At n = N n = N n = N , Stage III estimates Δ N \Delta_N Δ N using the completed trial data. Final borrowing weights define a kernel-standardized estimator of Δ N \Delta_N Δ N , whose outcome-weighted representation supports variance estimation and Wald inference. To determine the final amount of historical borrowing, we first reassess historical–concurrent compatibility using the completed trial data. For each covariate profile x \mathbf{x} x , the terminal local conflict score D N ( x ) D_N(\mathbf{x}) D N ( x ) is obtained from (3) by setting n = N n = N n = N , so that the comparison uses all concurrent controls observed during the trial. Although the local conflict score allows borrowing to adapt across the covariate space, relying on local compatibility alone can lead to unstable borrowing in finite samples. In our numerical investigations, such instability can result in excessive historical borrowing and appreciable bias, particularly when the historical and concurrent control populations differ systematically in their overall outcome levels. To provide an additional safeguard in the final analysis, Stage III therefore supplements the local assessment with a marginal measure of historical–concurrent conflict.
Let Y ˉ C \bar{Y}^{C} Y ˉ C and Y ˉ H \bar{Y}^{H} Y ˉ H denote the overall concurrent- and historical-control means, respectively, with sample variances s ^ C 2 \hat{s}_C^2 s ^ C 2 and s ^ H 2 \hat{s}_H^2 s ^ H 2 . We define the marginal conflict score as
D N G = ( − log [ max { 2 Φ ( − ∣ Y ˉ H − Y ˉ C ∣ ( s ^ C 2 / N 0 + s ^ H 2 / N H ) 1 / 2 ) , ϵ c } ] − 1 ) + . D_N^G = \left(-\log\left[\max\left\{2\Phi\left(-\frac{\left|\bar{Y}_H-\bar{Y}_C\right|}{\left(\hat{s}_C^2/N_0+\hat{s}_H^2/N_H\right)^{1/2}}\right),\epsilon_c\right\}\right]-1\right)_+. D N G = ( − log [ max { 2Φ ( − ( s ^ C 2 / N 0 + s ^ H 2 / N H ) 1/2 Y ˉ H − Y ˉ C ) , ϵ c } ] − 1 ) + .
The local and marginal conflict scores are then combined through D N F ( x ) = D N ( x ) + { 1 − exp [ − D N ( x ) ] } D N G D_N^F(\boldsymbol{x}) = D_N(\boldsymbol{x}) + \{1-\exp[-D_N(\boldsymbol{x})]\}D_N^G D N F ( x ) = D N ( x ) + { 1 − exp [ − D N ( x )]} D N G . The multiplier 1 − exp { − D N ( x ) } 1-\exp\{-D_N(\boldsymbol{x})\} 1 − exp { − D N ( x )} controls the influence of the marginal conflict score according to the degree of local incompatibility. When the local score is zero, the marginal score does not affect borrowing at x \boldsymbol{x} x . As the local score increases, the marginal assessment receives greater weight and provides an additional safeguard against borrowing from an overall incompatible historical source.
Using the completed concurrent-control cohort, define n ~ 0 , N e f f ( x ) = max { n 0 , N e f f ( x ) , n min } \tilde{n}_{0,N}^{\mathrm{eff}}(\boldsymbol{x})=\max\{n_{0,N}^{\mathrm{eff}}(\boldsymbol{x}),n_{\min}\} n ~ 0 , N eff ( x ) = max { n 0 , N eff ( x ) , n m i n } and S E N 2 ( x ) = τ ^ H 2 ( x ) + σ ^ 0 , N 2 ( x ) / n ~ 0 , N e f f ( x ) \mathrm{SE}_N^2(\boldsymbol{x})=\hat{\tau}_H^2(\boldsymbol{x})+\hat{\sigma}_{0,N}^2(\boldsymbol{x})/\tilde{n}_{0,N}^{\mathrm{eff}}(\boldsymbol{x}) SE N 2 ( x ) = τ ^ H 2 ( x ) + σ ^ 0 , N 2 ( x ) / n ~ 0 , N eff ( x ) . The final information ratio is
R N F ( x ) = 1 + σ ^ 0 , N 2 ( x ) n ~ 0 , N e f f ( x ) S E N 2 ( x ) exp { − D N F ( x ) } . R_N^F(\boldsymbol{x})=1+\frac{\hat{\sigma}_{0,N}^2(\boldsymbol{x})}{\tilde{n}_{0,N}^{\mathrm{eff}}(\boldsymbol{x})\mathrm{SE}_N^2(\boldsymbol{x})}\exp\{-D_N^F(\boldsymbol{x})\}. R N F ( x ) = 1 + n ~ 0 , N eff ( x ) SE N 2 ( x ) σ ^ 0 , N 2 ( x ) exp { − D N F ( x )} .
As in Stage II, the variance ratio accounts for historical precision. The exponential discount uses the combined local and marginal score D N F ( x ) D_N^F(\boldsymbol{x}) D N F ( x ) . Finally, we convert the information ratio into the borrowing weight W N F ( x ) = ( R N F ( x ) − 1 ) / R N F ( x ) W_N^F(\boldsymbol{x})=(R_N^F(\boldsymbol{x})-1)/R_N^F(\boldsymbol{x}) W N F ( x ) = ( R N F ( x ) − 1 ) / R N F ( x ) , which, by construction, is always between 0 and 1 / 2 1/2 1/2 . A larger W N F ( x ) W_N^F(\boldsymbol{x}) W N F ( x ) corresponds to greater historical borrowing when the historical reference is precise and compatible with the concurrent controls. As historical information becomes less precise or less compatible, W N F ( x ) W_N^F(\boldsymbol{x}) W N F ( x ) decreases toward zero.
The borrowing weight W N F ( x ) W_N^F(\boldsymbol{x}) W N F ( x ) combines the concurrent and historical local control estimates to construct the treatment-effect estimator Δ ^ N \widehat{\Delta}_N Δ N . Let ( Y H , Y 0 , Y 1 ) (Y_H,Y_0,Y_1) ( Y H , Y 0 , Y 1 ) denote the outcome vectors for the historical controls, concurrent controls, and treatment arm, respectively. For each covariate profile x \boldsymbol{x} x , let S 0 ( x ) ∈ R N 0 S_0(\boldsymbol{x})\in\mathbb{R}^{N_0} S 0 ( x ) ∈ R N 0 collect the normalized kernel weights assigned to the concurrent controls, so that S 0 ( x ) ⊤ Y 0 = Y ˉ 0 , N ( x ) S_0(\boldsymbol{x})^\top Y_0=\bar{Y}_{0,N}(\boldsymbol{x}) S 0 ( x ) ⊤ Y 0 = Y ˉ 0 , N ( x ) . Similarly, let S H ( x ) ∈ R N H S_H(\boldsymbol{x})\in\mathbb{R}^{N_H} S H ( x ) ∈ R N H collect the normalized historical kernel weights introduced in Stage I, so that S H ( x ) ⊤ Y H = θ ( x ) S_H(\boldsymbol{x})^\top Y_H=\theta(\boldsymbol{x}) S H ( x ) ⊤ Y H = θ ( x ) . The borrowing-augmented local control mean is
μ ^ 0 a u g ( x ) = { 1 − W N F ( x ) } S 0 ( x ) ⊤ Y 0 + W N F ( x ) S H ( x ) ⊤ Y H . \hat{\mu}_0^{\mathrm{aug}}(\boldsymbol{x})=\{1-W_N^F(\boldsymbol{x})\}S_0(\boldsymbol{x})^\top Y_0+W_N^F(\boldsymbol{x})S_H(\boldsymbol{x})^\top Y_H. μ ^ 0 aug ( x ) = { 1 − W N F ( x )} S 0 ( x ) ⊤ Y 0 + W N F ( x ) S H ( x ) ⊤ Y H .
The fit reduces to concurrent-only estimation when W N F ( x ) = 0 W_N^F(\boldsymbol{x}) = 0 W N F ( x ) = 0 . Since W N F ( x ) ≤ 1 / 2 W_N^F(\boldsymbol{x}) \leq1/2 W N F ( x ) ≤ 1/2 , historical controls receive no more weight than concurrent controls. Similarly, let S 1 ( x ) ∈ R N 1 \boldsymbol{S}_1(\boldsymbol{x}) \in\mathbb{R}^{N_1} S 1 ( x ) ∈ R N 1 denote the row-normalized kernel weight vector over the concurrent treatment arm. The local treatment mean is estimated by μ ^ 1 ( x ) = S 1 ( x ) T Y 1 \widehat{\mu}_1(\boldsymbol{x}) = \boldsymbol{S}_1(\boldsymbol{x})^{\mathsf{T}}\boldsymbol{Y}_1 μ 1 ( x ) = S 1 ( x ) T Y 1 . The treatment-effect estimator averages the fitted treatment contrast over the enrolled trial covariates:
Δ ^ N = 1 N ∑ i = 1 N { μ ^ 1 ( X i ) − μ ^ 0 a u g ( X i ) } . (5) \widehat{\Delta}_N = \frac{1}{N}\sum_{i=1}^{N}\left\{\widehat{\mu}_1(\boldsymbol{X}_i) - \widehat{\mu}_0^{\mathrm{aug}}(\boldsymbol{X}_i)\right\}. \tag*{(5)} Δ N = N 1 i = 1 ∑ N { μ 1 ( X i ) − μ 0 aug ( X i ) } . (5)
This estimator standardizes the fitted treatment effect to the empirical covariate distribution of the enrolled trial cohort and is an empirical-standardization form of the g-computation estimator (Robins, 1986; Snowden, Rose, and Mortimer, 2011).
For variance estimation, we express Δ ^ N \widehat{\Delta}_N Δ N as a weighted sum of the observed outcomes. Define S ~ 0 ( x ) = { 1 − W N F ( x ) } S 0 ( x ) \widetilde{\boldsymbol{S}}_0(\boldsymbol{x}) = \{1-W_N^F(\boldsymbol{x})\}\boldsymbol{S}_0(\boldsymbol{x}) S 0 ( x ) = { 1 − W N F ( x )} S 0 ( x ) , S ~ H ( x ) = W N F ( x ) S H ( x ) \widetilde{\boldsymbol{S}}_H(\boldsymbol{x}) = W_N^F(\boldsymbol{x})\boldsymbol{S}_H(\boldsymbol{x}) S H ( x ) = W N F ( x ) S H ( x ) . Define the cumulative coefficient vectors by a 1 = N − 1 ∑ i = 1 N S 1 ( X i ) \boldsymbol{a}_1 = N^{-1}\sum_{i=1}^{N}\boldsymbol{S}_1(\boldsymbol{X}_i) a 1 = N − 1 ∑ i = 1 N S 1 ( X i ) , a 0 = N − 1 ∑ i = 1 N S ~ 0 ( X i ) \boldsymbol{a}_0 = N^{-1}\sum_{i=1}^{N}\widetilde{\boldsymbol{S}}_0(\boldsymbol{X}_i) a 0 = N − 1 ∑ i = 1 N S 0 ( X i ) , a H = N − 1 ∑ i = 1 N S ~ H ( X i ) \boldsymbol{a}_H = N^{-1}\sum_{i=1}^{N}\widetilde{\boldsymbol{S}}_H(\boldsymbol{X}_i) a H = N − 1 ∑ i = 1 N S H ( X i ) . Then
Δ ^ N = a 1 T Y 1 − a 0 T Y 0 − a H T Y H . (6) \widehat{\Delta}_N = \boldsymbol{a}_1^{\mathsf{T}}\boldsymbol{Y}_1 - \boldsymbol{a}_0^{\mathsf{T}}\boldsymbol{Y}_0 - \boldsymbol{a}_H^{\mathsf{T}}\boldsymbol{Y}_H. \tag*{(6)} Δ N = a 1 T Y 1 − a 0 T Y 0 − a H T Y H . (6)
This is algebraically equivalent to (5) , with a 1 \boldsymbol{a}_1 a 1 , a 0 \boldsymbol{a}_0 a 0 , and a H \boldsymbol{a}_H a H collecting the respective cumulative treatment, concurrent-control, and historical-control contributions. For each source z ∈ { 0 , 1 , H } z \in\{0,1,H\} z ∈ { 0 , 1 , H } , let S z z \boldsymbol{S}_{zz} S z z be its row-normalized kernel matrix evaluated at its own covariates, including self-weights. We estimate the source-specific residual variance by
σ ^ z 2 = ∥ Y z − S z z Y z ∥ 2 2 max { N z − 2 tr ( S z z ) + ∥ S z z ∥ F 2 , 1 } . \widehat{\sigma}_z^2 = \frac{\left\lVert\boldsymbol{Y}_z-\boldsymbol{S}_{zz}\boldsymbol{Y}_z\right\rVert_2^2}{\max\left\{N_z-2\operatorname{tr}(\boldsymbol{S}_{zz})+\left\lVert\boldsymbol{S}_{zz}\right\rVert_F^2,1\right\}}. σ z 2 = max { N z − 2 tr ( S z z ) + ∥ S z z ∥ F 2 , 1 } ∥ Y z − S z z Y z ∥ 2 2 .
The denominator is the smoother’s residual degrees of freedom. Here ∥ ⋅ ∥ F \left\lVert\cdot\right\rVert_F ∥ ⋅ ∥ F denotes the Frobenius norm, and σ ^ H 2 \widehat{\sigma}_H^2 σ H 2 is source-wide, distinct from the local σ ^ H 2 ( x ) \widehat{\sigma}_H^2(\boldsymbol{x}) σ H 2 ( x ) in Stage I. Treating the coefficients as fixed under the homoskedastic working model gives the model-assisted linear-smoother variance
V ^ l i n = σ ^ 1 2 ∥ a 1 ∥ 2 2 + σ ^ 0 2 ∥ a 0 ∥ 2 2 + σ ^ H 2 ∥ a H ∥ 2 2 . (7) \widehat{V}_{\mathrm{lin}} = \widehat{\sigma}_1^2\left\lVert\boldsymbol{a}_1\right\rVert_2^2 + \widehat{\sigma}_0^2\left\lVert\boldsymbol{a}_0\right\rVert_2^2 + \widehat{\sigma}_H^2\left\lVert\boldsymbol{a}_H\right\rVert_2^2. \tag*{(7)} V lin = σ 1 2 ∥ a 1 ∥ 2 2 + σ 0 2 ∥ a 0 ∥ 2 2 + σ H 2 ∥ a H ∥ 2 2 . (7)
The borrowing weights entering a 0 \boldsymbol{a}_0 a 0 and a H \boldsymbol{a}_H a H are estimated from the same outcomes used in Δ ^ N \widehat{\Delta}_N Δ N . Treating the realized weights as fixed may therefore understate finite-sample uncertainty.
To keep the reported variance at least as large as its concurrent-only counterpart, we also calculate the variance corresponding to the concurrent-only control estimator. Let a 0 C = N − 1 ∑ i = 1 N S 0 ( X i ) a_0^C = N^{-1}\sum_{i=1}^{N} S_0(X_i) a 0 C = N − 1 ∑ i = 1 N S 0 ( X i ) , and define V ^ C = σ ^ 1 2 ∥ a 1 ∥ 2 + σ ^ 0 2 ∥ a 0 C ∥ 2 \widehat{V}_C = \widehat{\sigma}_1^2\lVert a_1\rVert^2 + \widehat{\sigma}_0^2\lVert a_0^C\rVert^2 V C = σ 1 2 ∥ a 1 ∥ 2 + σ 0 2 ∥ a 0 C ∥ 2 . For computation, V ^ l i n \widehat{V}_{\mathrm{lin}} V lin and V ^ C \widehat{V}_C V C are also floored at ϵ v \epsilon_v ϵ v . The reported variance is V ^ N = max { V ^ l i n , V ^ C } \widehat{V}_N = \max\{\widehat{V}_{\mathrm{lin}}, \widehat{V}_C\} V N = max { V lin , V C } . Thus, the reported interval is at least as wide as the concurrent-only kernel interval calculated from the same realized trial data.
The large-sample analysis compares RADISH with a concurrent-only oracle. Conditional on covariates and randomization draws, its allocation and smoothing coefficients are outcome-independent. The proof establishes first-order equivalence of the RADISH and oracle estimators and consistency of the reported variance, including the concurrent-only maximum. In Supplementary § 5.5.2 , condition (C4) controls the allocation perturbation and requires terminal borrowing to be first-order negligible, while (C3) and (C6) control smoothing bias and residual-variance estimation. The following Studentized limit supports Wald inference.
Theorem 1 (Asymptotic normality of RADISH estimator): Suppose conditions (C1)–(C6) in Supplementary § 5.5.2 hold. Then, conditional on the historical data D H \mathcal{D}_H D H and the enrolled covariates X 1 : N \mathbf{X}_{1:N} X 1 : N ,
Δ ^ N − Δ N V ^ N → d N ( 0 , 1 ) . \frac{\widehat{\Delta}_N-\Delta_N}{\sqrt{\widehat{V}_N}} \xrightarrow{d} N(0,1). V N Δ N − Δ N d N ( 0 , 1 ) .
Supplementary § 5.5 gives the full conditions and proof. A two-sided 100 ( 1 − α ) % 100(1-\alpha)\% 100 ( 1 − α ) % Wald confidence interval for Δ N \Delta_N Δ N is therefore
Δ ^ N ± z 1 − α / 2 V ^ N , (8) \widehat{\Delta}_N \pm z_{1-\alpha/2}\sqrt{\widehat{V}_N}, \tag*{(8)} Δ N ± z 1 − α /2 V N , (8)
where z 1 − α / 2 z_{1-\alpha/2} z 1 − α /2 is the standard normal ( 1 − α / 2 ) (1-\alpha/2) ( 1 − α /2 ) quantile. The two-sided test rejects H 0 : Δ N = 0 H_0:\Delta_N=0 H 0 : Δ N = 0 at nominal level α \alpha α when this interval excludes zero. § 2 evaluates finite-sample performance. Algorithm 1 summarizes the three stages; Algorithms S1–S3 in Supplementary Section S6 give the detailed computations.
Algorithm 1: The three-stage RADISH design.
Input: Historical controls D H D_H D H , trial size N N N , n i n i t = 20 n_{\mathrm{init}} = 20 n init = 20 , and α = 0.05 \alpha= 0.05 α = 0.05 .
Output: Assignments { Z i } i = 1 N \{Z_i\}_{i=1}^{N} { Z i } i = 1 N , estimate Δ ^ N \widehat{\Delta}_N Δ N , and Wald interval (8) .
Fix h h h and construct { θ ( x ) , τ ^ H 2 ( x ) } \{\theta(x), \widehat{\tau}_H^2(x)\} { θ ( x ) , τ H 2 ( x )} as in Stage I;
Randomize the first n i n i t n_{\mathrm{init}} n init participants with probability 1 / 2 1/2 1/2 and observe their outcomes;
for n = n i n i t , … , N − 1 n = n_{\mathrm{init}}, \ldots, N - 1 n = n init , … , N − 1 do
Observe X n + 1 X_{n+1} X n + 1 ; compute local control summaries, arm masses, D n D_n D n , and R n A R_n^A R n A ;
Compute ϕ n \phi_n ϕ n by (4) ; draw Z n + 1 Z_{n+1} Z n + 1 , observe Y n + 1 Y_{n+1} Y n + 1 , and append the record;
end
Using all concurrent controls, recompute D N ( X i ) D_N(X_i) D N ( X i ) and the marginal score D N G D_N^G D N G ; obtain D N F ( X i ) D_N^F(X_i) D N F ( X i ) , R N F ( X i ) R_N^F(X_i) R N F ( X i ) , and W N F ( X i ) W_N^F(X_i) W N F ( X i ) as in Stage III;
Form the fitted means and estimate Δ ^ N \widehat{\Delta}_N Δ N by (5) ;
Compute the residual variances, V ^ l i n \widehat{V}_{\mathrm{lin}} V lin in (7) , and V ^ C \widehat{V}_C V C ; set V ^ N = max { V ^ l i n , V ^ C } \widehat{V}_N = \max\{\widehat{V}_{\mathrm{lin}}, \widehat{V}_C\} V N = max { V lin , V C } ;
Compute se ^ N = V ^ N \widehat{\operatorname{se}}_N = \sqrt{\widehat{V}_N} se N = V N and return [ Δ ^ N ± z 1 − α / 2 se ^ N ] [\widehat{\Delta}_N \pm z_{1-\alpha/2}\widehat{\operatorname{se}}_N] [ Δ N ± z 1 − α /2 se N ] ;
Simulation Studies
Simulation Settings and Data-Generating Process
We evaluated sequential allocation and final inference using external historical controls and a sequentially enrolled two-arm trial, with fixed sample sizes N H = N = 200 N_H = N = 200 N H = N = 200 and M = 1000 M = 1000 M = 1000 independent replications per scenario. After n i n i t = 20 n_{\mathrm{init}} = 20 n init = 20 equal-probability burn-in assignments, the adaptive allocation rule was activated. We used the Gaussian product kernel and fixed historical-data bandwidth specified in § 1.2.1 , with d = 2 d = 2 d = 2 and Scott’s rule (Scott, 2015).
Each participant was assigned X = ( X 1 , X 2 ) T X = (X_1, X_2)^{\mathsf{T}} X = ( X 1 , X 2 ) T , with X 1 ∼ Bernoulli ( 0.5 ) X_1 \sim\operatorname{Bernoulli}(0.5) X 1 ∼ Bernoulli ( 0.5 ) and X 2 ∼ N ( 0 , 1 ) X_2 \sim\mathcal{N}(0,1) X 2 ∼ N ( 0 , 1 ) . The concurrent outcome was Y = μ 0 ( C ) ( X ) + Δ Z + ε Z Y = \mu_0^{(C)}(X) + \Delta Z + \varepsilon_Z Y = μ 0 ( C ) ( X ) + Δ Z + ε Z , where $\mu_0^{(C)}(X) = \beta_0 + β 1 X 1 + β 2 X 2 + β 3 X 1 X 2 \beta_1 X_1 + \beta_2 X_2 + \beta_3 X_1 X_2 β 1 X 1 + β 2 X 2 + β 3 X 1 X 2 . We set β 0 = 0 \beta_0 = 0 β 0 = 0 , β 1 = β 2 = 0.5 \beta_1 = \beta_2 = 0.5 β 1 = β 2 = 0.5 , and β 3 = 0.3 \beta_3 = 0.3 β 3 = 0.3 , with ε 0 ∼ N ( 0 , 1.0 2 ) \varepsilon_0 \sim\mathcal{N}(0,1.0^2) ε 0 ∼ N ( 0 , 1. 0 2 ) and ε 1 ∼ N ( 0 , 1.2 2 ) \varepsilon_1 \sim\mathcal{N}(0,1.2^2) ε 1 ∼ N ( 0 , 1. 2 2 ) . Historical outcomes followed Y H = μ 0 ( C ) ( X ) + b ( X 1 ) + ε H Y^H = \mu_0^{(C)}(\mathbf{X}) + b(X_1) + \varepsilon_H Y H = μ 0 ( C ) ( X ) + b ( X 1 ) + ε H , where ε H ∼ N ( 0 , σ H 2 ) \varepsilon_H \sim\mathcal{N}(0,\sigma_H^2) ε H ∼ N ( 0 , σ H 2 ) and b ( x 1 ) = b ( 0 ) 1 { x 1 = 0 } + b ( 1 ) 1 { x 1 = 1 } b(x_1) = b(0)\mathbf{1}\{x_1 = 0\} + b(1)\mathbf{1}\{x_1 = 1\} b ( x 1 ) = b ( 0 ) 1 { x 1 = 0 } + b ( 1 ) 1 { x 1 = 1 } . Thus μ H ( x ) = μ 0 ( C ) ( x ) + b ( x 1 ) \mu_H(\mathbf{x}) = \mu_0^{(C)}(\mathbf{x}) + b(x_1) μ H ( x ) = μ 0 ( C ) ( x ) + b ( x 1 ) .
In the simulation, the historical outcome variance is homoskedastic, so that σ H ( x ) ≡ σ H \sigma_H(\mathbf{x}) \equiv\sigma_H σ H ( x ) ≡ σ H for all x \mathbf{x} x . In every nonzero-conflict scenario, the historical surface differs from the concurrent surface in both X 1 X_1 X 1 strata, with a larger shift when X 1 = 1 X_1 = 1 X 1 = 1 . Consequently, the overall mean shift { b ( 0 ) + b ( 1 ) } / 2 \{b(0) + b(1)\}/2 { b ( 0 ) + b ( 1 )} /2 can conceal differences between the two strata.
We considered Δ ∈ { 0 , 0.5 } \Delta\in\{0, 0.5\} Δ ∈ { 0 , 0.5 } . Because the treatment effect is constant, μ 1 ( x ) − μ 0 ( x ) = Δ \mu_1(\mathbf{x}) - \mu_0(\mathbf{x}) = \Delta μ 1 ( x ) − μ 0 ( x ) = Δ for every x \mathbf{x} x , and hence Δ N = Δ \Delta_N = \Delta Δ N = Δ in every trial. The primary study uses a 4 × 2 4 \times2 4 × 2 conflict-by-precision design with ( b ( 0 ) , b ( 1 ) ) ∈ { ( 0 , 0 ) , ( 0.04 , 0.12 ) , ( 0.25 , 0.75 ) , ( 1.5 , 2.5 ) } (b(0), b(1)) \in\{(0, 0), (0.04, 0.12), (0.25, 0.75), (1.5, 2.5)\} ( b ( 0 ) , b ( 1 )) ∈ {( 0 , 0 ) , ( 0.04 , 0.12 ) , ( 0.25 , 0.75 ) , ( 1.5 , 2.5 )} and σ H ∈ { 0.5 , 3.0 } \sigma_H \in\{0.5, 3.0\} σ H ∈ { 0.5 , 3.0 } , defining scenarios S1–S8. S1/S2 are exactly compatible, S3/S4 have small covariate-dependent conflict, S5/S6 have moderate conflict, and S7/S8 have severe conflict. All nonzero-conflict scenarios are covariate dependent, with different historical shifts for X 1 = 0 X_1 = 0 X 1 = 0 and X 1 = 1 X_1 = 1 X 1 = 1 . Within each conflict level, we consider a precise historical source with σ H = 0.5 \sigma_H = 0.5 σ H = 0.5 and a less precise historical source with σ H = 3.0 \sigma_H = 3.0 σ H = 3.0 .
We compared RADISH with kernel-based covariate-adaptive design (KBCD) [3 ] and covariate-adaptive historical borrowing (CAHB) (Jin et al., 2023). KBCD fixes R n K B C D ( x ) = 1 R_n^{\mathrm{KBCD}}(\mathbf{x}) = 1 R n KBCD ( x ) = 1 and uses only concurrent outcomes. Like the other methods, it uses the prespecified planning bandwidth computed from historical covariates. It therefore serves as the concurrent-only inference reference. CAHB uses one fitted local model to calculate posterior variances with its borrowing term disabled and enabled, denoted by Var r e f { μ 0 ( x ) } \operatorname{Var}_{\mathrm{ref}}\{\mu_0(\mathbf{x})\} Var ref { μ 0 ( x )} and Var b o r r o w { μ 0 ( x ) } \operatorname{Var}_{\mathrm{borrow}}\{\mu_0(\mathbf{x})\} Var borrow { μ 0 ( x )} , respectively, and sets R n C A H B ( x ) = Var r e f { μ 0 ( x ) } / Var b o r r o w { μ 0 ( x ) } R_n^{\mathrm{CAHB}}(\mathbf{x}) = \operatorname{Var}_{\mathrm{ref}}\{\mu_0(\mathbf{x})\}/\operatorname{Var}_{\mathrm{borrow}}\{\mu_0(\mathbf{x})\} R n CAHB ( x ) = Var ref { μ 0 ( x )} / Var borrow { μ 0 ( x )} . The implemented comparator uses γ = 0.25 \gamma= 0.25 γ = 0.25 and λ 2 ( n ) = 300 max { log n , 1 } \lambda_2(n) = 300\max\{\log n, 1\} λ 2 ( n ) = 300 max { log n , 1 } . All three methods use the allocation rule in (4) . The CAHB comparator uses the common terminal kernel estimator and residual-based variance calculation, so this comparison evaluates its borrowing and allocation rules within the shared final-analysis framework. RADISH additionally applies the concurrent-only variance floor described in § 1.2.4 .
For each scenario and method, we evaluated the empirical bias, the root mean squared error (RMSE), the Type I error (rejection rate under Δ = 0 \Delta= 0 Δ = 0 ), and calibrated power under Δ = 0.5 \Delta= 0.5 Δ = 0.5 .
Calibrated power was calculated using the empirical 95th percentile of the absolute null Wald statistic as its critical value (Lloyd, 2005). We also evaluated the empirical confidence-interval coverage probability, the average confidence-interval width, and the allocation proportion to the control arm, N 0 / N N_0/N N 0 / N . To describe how borrowing changes along the bias and precision axes, we also recorded three borrowing measures: the mean borrowing weight W ‾ N F = N − 1 ∑ i W N , i F \overline{W}^{F}_{N} = N^{-1}\sum_i W^{F}_{N,i} W N F = N − 1 ∑ i W N , i F , with W N , i F = ( R N , i F − 1 ) / R N , i F W^{F}_{N,i} = (R^{F}_{N,i} - 1)/R^{F}_{N,i} W N , i F = ( R N , i F − 1 ) / R N , i F , the mean local information ratio R ‾ N F = N − 1 ∑ i R N , i F \overline{R}^{F}_{N} = N^{-1}\sum_i R^{F}_{N,i} R N F = N − 1 ∑ i R N , i F , and the mean combined conflict score D ‾ N F = N − 1 ∑ i D N F ( X i ) \overline{D}^{F}_{N} = N^{-1}\sum_i D^{F}_{N}(\mathbf{X}_i) D N F = N − 1 ∑ i D N F ( X i ) . Each measure was averaged across enrolled participants and replications.
Hypothesis Testing and Estimation Performance
Figure 1 shows hypothesis-testing performance and Table 1 reports estimation, interval, allocation, and borrowing results under the alternative. The corresponding null results are given in Supplementary Table S1. With M = 1000 M = 1000 M = 1000 replications, the normal-approximation 95% Monte Carlo reference interval around a nominal Type I error of 0.05 is [0.036, 0.064]. RADISH’s Type I errors range from 0.031 to 0.062 across the eight scenarios. Most of the Type I errors are within the reference interval apart from the conservative S1 result (0.031). CAHB is well calibrated under compatibility but records 0.487/0.183 under moderate conflict with high and low precision historical information, respectively. This inflation may reflect CAHB’s assessment of discrepancy after shrinking the concurrent control mean toward the historical mean, which can make moderate conflict appear small enough to retain substantial borrowing. The concurrent-only KBCD rejection rate is 0.055 throughout since it does not borrow from the historical data.
Table 1. Estimation, interval, allocation, and borrowing performance under Δ = 0.5 \Delta= 0.5 Δ = 0.5 , based on M = 1000 M = 1000 M = 1000 replications with N H = N = 200 N_{H} = N = 200 N H = N = 200 . Scenarios S1–S8 combine standardized historical shifts ( b ( 0 ) , b ( 1 ) ) = ( 0 , 0 ) , ( 0.04 , 0.12 ) , ( 0.25 , 0.75 ) , ( 1.5 , 2.5 ) (b(0), b(1)) = (0, 0), (0.04, 0.12), (0.25, 0.75), (1.5, 2.5) ( b ( 0 ) , b ( 1 )) = ( 0 , 0 ) , ( 0.04 , 0.12 ) , ( 0.25 , 0.75 ) , ( 1.5 , 2.5 ) with σ H = 0.5 , 3.0 \sigma_{H} = 0.5, 3.0 σ H = 0.5 , 3.0 , respectively. RMSE is root mean squared error; Width is mean 95% interval width; Alloc. Proportion is the proportion assigned to the control group; W ˉ N F \bar{W}_{N}^{F} W ˉ N F is the mean Stage III borrowing weight.
Figure 1. Hypothesis-testing performance across the eight simulation scenarios at M = 1000 M = 1000 M = 1000 replications. Left panel: empirical Type I error rate under the null hypothesis (Δ = 0 \Delta= 0 Δ = 0 ); the horizontal dashed line marks the nominal α = 0.05 \alpha= 0.05 α = 0.05 level, and the shaded band is the normal-approximation 95% confidence interval for a correctly calibrated test at this M M M . The right panel shows calibrated power under the alternative hypothesis (Δ = 0.5 \Delta= 0.5 Δ = 0.5 ), using the empirical 0.95 null quantile as the critical value. Error bars are pointwise normal-approximation 95% confidence intervals. Scenarios S1–S8 use ( b ( 0 ) , b ( 1 ) ) = ( 0 , 0 ) , ( 0.04 , 0.12 ) , ( 0.25 , 0.75 ) , ( 1.5 , 2.5 ) (b(0), b(1)) = (0, 0), (0.04, 0.12), (0.25, 0.75), (1.5, 2.5) ( b ( 0 ) , b ( 1 )) = ( 0 , 0 ) , ( 0.04 , 0.12 ) , ( 0.25 , 0.75 ) , ( 1.5 , 2.5 ) , each at σ H = 0.5 \sigma_H = 0.5 σ H = 0.5 and 3.0, with N H = N = 200 N_H = N = 200 N H = N = 200 .
[Figure 1 about here.]
[Table 1 about here.]
RADISH attains calibrated power 0.951/0.907 under compatibility and 0.916/0.903 under small conflict, compared with 0.871 for KBCD. In the moderate-conflict scenarios, RADISH records 0.847/0.865, whereas CAHB falls to 0.070/0.290 as its negative bias offsets the treatment effect. Under severe conflict, RADISH records 0.875/0.879, close to KBCD, while its mean borrowing weights fall to 0.001/0.004.
With precise historical information, RADISH reduces RMSE from 0.1507 for KBCD to 0.1362 under compatibility and 0.1374 under small conflict (Table 1 ). These gains diminish with lower historical precision. Under moderate conflict, RADISH retains small biases and RMSEs near KBCD’s, whereas CAHB exhibits substantially larger errors. Under severe conflict, RADISH approaches KBCD as borrowing vanishes. Coverage under the alternative ranges from 0.956 to 0.974, reflecting the concurrent-only variance safeguard while preserving estimation gains from informative borrowing.
Allocation Behavior and Borrowing Diagnostics
Table 1 and Supplementary Figure S2 summarize allocation, while Figure 2 shows local borrowing. KBCD’s control allocation remains near 0.50. With precise, compatible historical information (S1), CAHB/RADISH allocate 0.407/0.420 to control, with similar proportions under small conflict (S3). RADISH moves toward equal allocation as conflict increases or historical precision decreases. Under moderate conflict (S5), its proportion is 0.456 versus 0.412 for CAHB; under severe conflict (S7/S8), it is nearly 0.50.
Figure 2. Local borrowing performance under σ H = 0.5 \sigma_H = 0.5 σ H = 0.5 , Δ = 0.5 \Delta= 0.5 Δ = 0.5 and N H = N = 200 N_H = N = 200 N H = N = 200 based on M = 1000 M = 1000 M = 1000 replications. The path uses b ( 1 ) ∈ { 0.001 , 0.003 , 0.01 , 0.03 , 0.10 , 0.30 , 1 , 3 } b(1) \in\{0.001, 0.003, 0.01, 0.03, 0.10, 0.30, 1, 3\} b ( 1 ) ∈ { 0.001 , 0.003 , 0.01 , 0.03 , 0.10 , 0.30 , 1 , 3 } and b ( 0 ) = b ( 1 ) / 3 b(0) = b(1)/3 b ( 0 ) = b ( 1 ) /3 , so every nonzero conflict depends on X 1 X_1 X 1 and is stronger in the X 1 = 1 X_1 = 1 X 1 = 1 stratum. Panels show (a) mean absolute error, (b) calibrated power using an empirical Type I error of 0.05 within each scenario and method, (c) the mean terminal borrowing weight by stratum, and (d) the proportion assigned to the control group by stratum. Solid and dashed curves denote the lower- and higher-conflict strata, respectively.
Figure S2. Overall proportion assigned to the control group across M = 1000 M=1000 M = 1000 paired replications under Δ = 0 \Delta=0 Δ = 0 and Δ = 0.5 \Delta=0.5 Δ = 0.5 , with N H = N = 200 N_H=N=200 N H = N = 200 . The eight scenarios cross ( b ( 0 ) , b ( 1 ) ) = ( 0 , 0 ) , ( 0.04 , 0.12 ) , ( 0.25 , 0.75 ) , ( 1.5 , 2.5 ) (b(0),b(1))=(0,0),(0.04,0.12),(0.25,0.75),(1.5,2.5) ( b ( 0 ) , b ( 1 )) = ( 0 , 0 ) , ( 0.04 , 0.12 ) , ( 0.25 , 0.75 ) , ( 1.5 , 2.5 ) with σ H = 0.5 , 3.0 \sigma_H=0.5,3.0 σ H = 0.5 , 3.0 , respectively; colors identify KBCD, CAHB, and RADISH. Boxes show the interquartile range with the median line, and whiskers use the standard 1.5-IQR rule. The horizontal reference at 0.5 denotes equal randomization; values below 0.5 indicate allocating fewer participants to the control group, whose interpretation depends on historical–concurrent compatibility. KBCD is the concurrent-only reference, CAHB is the borrowing comparator, and RADISH is the proposed locally adaptive method.
[Figure 2 about here.]
Figure 2 further illustrates the local borrowing mechanism along an eight-point conflict path with fixed historical precision (σ H = 0.5 \sigma_H = 0.5 σ H = 0.5 ), where b ( 1 ) ∈ { 0.001 , 0.003 , 0.01 , 0.03 , 0.10 , b(1) \in\{0.001, 0.003, 0.01, 0.03, 0.10, b ( 1 ) ∈ { 0.001 , 0.003 , 0.01 , 0.03 , 0.10 , 0.30 , 1 , 3 } 0.30, 1, 3\} 0.30 , 1 , 3 } and b ( 0 ) = b ( 1 ) / 3 b(0) = b(1)/3 b ( 0 ) = b ( 1 ) /3 . Along this path, RADISH retains its estimation and power gains under small conflict, while the borrowing weights for the two X 1 X_1 X 1 strata gradually separate as the covariate-specific discrepancies become more pronounced. As conflict increases further, both borrowing weights approach zero and the performance of RADISH converges toward that of KBCD. CAHB remains efficient near compatibility, but its estimation accuracy and calibrated power deteriorate substantially over the moderate-conflict range. Supplementary Figure S3 further examines how borrowing responds to the mean shift and variation in subgroup-specific shifts. The sensitivity analysis with a larger historical sample size, N H = 400 N_H = 400 N H = 400 , shows a similar qualitative pattern (Supplementary Figure S4), with somewhat greater efficiency gains under compatibility and small conflict while retaining the same conflict-adaptive pattern. The HORIZON–FIT study below evaluates this borrowing behavior using covariate distributions and an outcome model calibrated to clinical trial data.
Semi-Synthetic Simulation Studies Based on HORIZON–FIT Data
HORIZON–FIT Data and Outcome Model
We evaluated borrowing under overall and subgroup-specific mean shifts using semi-synthetic data from HORIZON (Black et al., 2007) and FIT (Black et al., 1996; Cummings et al., 1998). We used HORIZON data to estimate the covariate distribution and fit the outcome regression model for the concurrent trial, and FIT data to estimate the covariate distribution for the historical controls. Both were placebo-controlled randomized trials involving postmenopausal women at elevated fracture risk. We used total hip bone mineral density (BMD) at 24 months as the outcome. This outcome was chosen because previous research supports using treatment-related percentage changes in total hip BMD as a surrogate for reductions in vertebral fracture risk (Black et al., 2020).
We restricted the analysis to White women aged at least 80 years in HORIZON and at least 78 years in FIT. After excluding participants with missing values, the cohorts comprised 620 HORIZON participants and 388 FIT participants. We included four prognostic covariates: fall history (X 1 X_1 X 1 , X 1 = 1 X_1 = 1 X 1 = 1 if there is any fall history, otherwise 0), vertebral fracture history (X 2 X_2 X 2 , X 2 = 1 X_2 = 1 X 2 = 1 if there is any vertebral fracture history, otherwise 0), years since menopause (X 3 X_3 X 3 ) and the baseline hip BMD (X 4 X_4 X 4 ).
We standardized 24-month hip BMD using the mean and standard deviation from the HORIZON trial and modeled the standardized outcome as
Y ~ = β 0 ( X 1 , X 2 ) + β 1 ( X 1 , X 2 ) X ~ 3 + β 2 ( X 1 , X 2 ) X ~ 4 + δ Z + ε C , ε C ∼ N ( 0 , σ C 2 ) . (9) \widetilde{Y} = \beta_0(X_1, X_2) + \beta_1(X_1, X_2)\widetilde{X}_3 + \beta_2(X_1, X_2)\widetilde{X}_4 + \delta Z + \varepsilon_C,\qquad\varepsilon_C \sim\mathcal{N}(0,\sigma_C^2). \tag*{(9)} Y = β 0 ( X 1 , X 2 ) + β 1 ( X 1 , X 2 ) X 3 + β 2 ( X 1 , X 2 ) X 4 + δ Z + ε C , ε C ∼ N ( 0 , σ C 2 ) . (9)
The continuous covariates X ~ 3 \widetilde{X}_3 X 3 and X ~ 4 \widetilde{X}_4 X 4 were also standardized using the corresponding means and standard deviations from the HORIZON trial. The four subgroups were indexed by ( X 1 , X 2 ) ∈ { ( 0 , 0 ) , ( 1 , 0 ) , ( 0 , 1 ) , ( 1 , 1 ) } (X_1, X_2) \in\{(0,0),(1,0),(0,1),(1,1)\} ( X 1 , X 2 ) ∈ {( 0 , 0 ) , ( 1 , 0 ) , ( 0 , 1 ) , ( 1 , 1 )} .
We used ordinary least squares within each HORIZON subgroup to estimate the subgroup-specific intercepts and covariate coefficients in (9) . We also estimated the conditional distributions of ( X 3 , X 4 ) ∣ ( X 1 , X 2 ) (X_3, X_4) \mid(X_1, X_2) ( X 3 , X 4 ) ∣ ( X 1 , X 2 ) separately in HORIZON and FIT using Gaussian kernel density estimation, with bandwidths selected using Scott’s rule of thumb (Scott, 2015). For each concurrent participant, we first sampled ( X 1 , X 2 ) (X_1, X_2) ( X 1 , X 2 ) using the empirical subgroup proportions in HORIZON, approximately ( 0.198 , 0.082 , 0.492 , 0.227 ) (0.198, 0.082, 0.492, 0.227) ( 0.198 , 0.082 , 0.492 , 0.227 ) , and then sampled ( X 3 , X 4 ) (X_3, X_4) ( X 3 , X 4 ) from the corresponding estimated conditional distribution. Given these covariates and treatment assignment, we generated outcomes from (9) , with δ = 0 \delta= 0 δ = 0 under the null hypothesis, and δ = 0.14 \delta= 0.14 δ = 0.14 under the alternative. The residual standard deviation was set to σ C = 0.11 \sigma_C = \sqrt{0.11} σ C = 0.11 , which was estimated from HORIZON data. Historical covariates were sampled analogously from the estimated FIT distribution. We then generated outcomes from (9) with Z = 0 Z = 0 Z = 0 and an additional mean shift b ( X 1 ) b(X_1) b ( X 1 ) . The same residual standard deviation was used for historical outcomes, i.e., σ H = σ C \sigma_H = \sigma_C σ H = σ C . The shift b ( X 1 ) b(X_1) b ( X 1 ) was specified according to the conflict scenarios described in the next subsection. Table 2 summarizes the simulation parameters.
Table 2. Simulation settings and estimated model parameters for the semi-synthetic simulation studies. Subgroup-specific values correspond to ( X 1 , X 2 ) = ( 0 , 0 ) (X_1, X_2)=(0,0) ( X 1 , X 2 ) = ( 0 , 0 ) , ( 1 , 0 ) (1,0) ( 1 , 0 ) , ( 0 , 1 ) (0,1) ( 0 , 1 ) , and ( 1 , 1 ) (1,1) ( 1 , 1 ) , respectively.
[Table 2 about here.]
Simulation Settings and Evaluation
We considered two sets of scenarios. The first varied the overall mean shift and its variation across fall-history subgroups (X 1 = 0 X_1 = 0 X 1 = 0 and X 1 = 1 X_1 = 1 X 1 = 1 ). The second increased subgroup differences while holding the overall mean shift at zero. All overall shifts were averaged over the estimated HORIZON covariate distribution.
For the first set of scenarios, let p 1 = 192 / 620 ≈ 0.310 p_1 = 192/620 \approx0.310 p 1 = 192/620 ≈ 0.310 denote the proportion of HORIZON participants with a history of falls. We specified the historical mean shift as
b ( X 1 ) σ C = g + ℓ X 1 − p 1 p 1 ( 1 − p 1 ) . \frac{b(X_1)}{\sigma_C} = g + \ell\frac{X_1-p_1}{\sqrt{p_1(1-p_1)}}. σ C b ( X 1 ) = g + ℓ p 1 ( 1 − p 1 ) X 1 − p 1 .
Under the HORIZON covariate distribution, the standardized shift b ( X 1 ) / σ C b(X_1)/\sigma_C b ( X 1 ) / σ C has mean g g g and standard deviation ℓ \ell ℓ . Thus, g g g is the HORIZON-averaged standardized conditional mean shift, and ℓ \ell ℓ measures its variation across fall-history subgroups. Because covariate distributions differ, the unadjusted marginal control-mean difference, on the modeled outcome scale, need not equal σ C g \sigma_C g σ C g .
We considered all four combinations of g ∈ { 0 , 2 } g \in\{0, 2\} g ∈ { 0 , 2 } and ℓ ∈ { 0.1 , ( 1 − p 1 ) / p 1 } \ell\in\{0.1, \sqrt{(1-p_1)/p_1}\} ℓ ∈ { 0.1 , ( 1 − p 1 ) / p 1 } , where ( 1 − p 1 ) / p 1 ≈ 1.493 \sqrt{(1-p_1)/p_1} \approx1.493 ( 1 − p 1 ) / p 1 ≈ 1.493 . The two values of g g g correspond to no mean shift and a mean shift of two residual standard deviations. The two values of ℓ \ell ℓ represent small and large variation in subgroup-specific mean shifts. For example, when g = 0 g = 0 g = 0 and ℓ = ( 1 − p 1 ) / p 1 \ell= \sqrt{(1-p_1)/p_1} ℓ = ( 1 − p 1 ) / p 1 , the scenario has substantial subgroup-specific shifts that are cancelled when averaged over the HORIZON population. This scenario examines whether borrowing methods detect subgroup-specific discrepancies when the mean discrepancy is zero.
The second set extended the range of subgroup-specific shifts while preserving a zero overall mean shift. We set
( b ( 0 ) , b ( 1 ) ) σ C = 4 ξ ( − 1 , 1 − p 1 p 1 ) , \frac{(b(0), b(1))}{\sigma_C} = 4\xi\left(-1, \frac{1-p_1}{p_1}\right), σ C ( b ( 0 ) , b ( 1 )) = 4 ξ ( − 1 , p 1 1 − p 1 ) ,
with ξ ∈ { 0 , 0.003162 , 0.01 , 0.031623 , 0.1 , 0.316228 , 1 } \xi\in\{0, 0.003162, 0.01, 0.031623, 0.1, 0.316228, 1\} ξ ∈ { 0 , 0.003162 , 0.01 , 0.031623 , 0.1 , 0.316228 , 1 } . At ξ = 0 \xi= 0 ξ = 0 , the historical and concurrent conditional control means coincide. As ξ \xi ξ increases, the shifts in the two fall-history subgroups increase in magnitude with opposite signs. Their average under the HORIZON covariate distribution remains zero. At ξ = 1 \xi= 1 ξ = 1 , the standardized shifts are approximately ( − 4.00 , 8.91 ) (-4.00, 8.91) ( − 4.00 , 8.91 ) .
For each scenario, we generated N H = 350 N_H = 350 N H = 350 historical controls and N = 200 N = 200 N = 200 concurrent participants. We conducted M = 1000 M = 1000 M = 1000 simulation replications for each method under both the null and alternative hypotheses. Each trial used n i n i t = 20 n_{\mathrm{init}} = 20 n init = 20 equal-probability burn-in assignments before method-specific allocation. Because the treatment effect was constant across covariates, Δ N = δ \Delta_N = \delta Δ N = δ in every replication. Table 2 summarizes the simulation settings.
We compared RADISH, KBCD, and CAHB as in § 2 . RADISH used the same parameter settings as in that section. For CAHB, we set γ = 3 \gamma= \sqrt{3} γ = 3 and λ 2 ( n ) = 300 max { log n , 1 } \lambda_2(n) = 300\max\{\log n, 1\} λ 2 ( n ) = 300 max { log n , 1 } .
We evaluated the proportion assigned to the control group, the mean borrowing weight W ‾ N F \overline{W}_N^F W N F , estimation bias, mean absolute error (MAE), Type I error, and calibrated power. Mean absolute error was calculated as M − 1 ∑ r = 1 M ∣ Δ ^ N , r − δ ∣ M^{-1}\sum_{r=1}^{M}|\widehat{\Delta}_{N,r}-\delta| M − 1 ∑ r = 1 M ∣ Δ N , r − δ ∣ , where Δ ^ N , r \widehat{\Delta}_{N,r} Δ N , r is the treatment-effect estimate in replication r r r . Calibrated power used the empirical null critical value estimated separately for each method and scenario, as described in § 2 . The resulting power comparisons use a common empirical null rejection rate for each scenario.
Results under Semi-Synthetic Simulation Studies
Figure 3 summarizes the results under the first set of scenarios. Compared with concurrent-only KBCD, RADISH improves estimation accuracy and calibrated power when the historical and concurrent controls are similar. With zero mean shift (g = 0 g = 0 g = 0 ) and small variation in subgroup-specific shifts (ℓ = 0.1 \ell= 0.1 ℓ = 0.1 ), RADISH reduces MAE from 0.042 to 0.040 and increases calibrated power from 0.734 to 0.787. The empirical Type I error is 0.029. In the remaining scenarios, RADISH is overall comparable to KBCD, with similar MAE, Type I error, and calibrated power. Nevertheless, when the mean shift is zero and variation in subgroup-specific shifts is large (ℓ ≈ 1.493 \ell\approx1.493 ℓ ≈ 1.493 ), RADISH shows modest Type I error inflation, with a rejection rate of 0.072.
Figure 3. Performance across four HORIZON–FIT simulation scenarios, with N = 200 N = 200 N = 200 , N H = 350 N_H = 350 N H = 350 , and 1000 replications per scenario. The HORIZON-averaged standardized conditional shift is labeled zero (g = 0 g = 0 g = 0 ) or large (g = 2 g = 2 g = 2 ). Variation in subgroup-specific shifts is labeled small (ℓ = 0.1 \ell= 0.1 ℓ = 0.1 ) or large (ℓ = 1.493 \ell= 1.493 ℓ = 1.493 ). Panels show (a) mean absolute error, (b) calibrated power, (c) empirical Type I error, and (d) estimation bias. Panels (a), (b), and (d) use δ = 0.14 \delta= 0.14 δ = 0.14 , and panel (c) uses δ = 0 \delta= 0 δ = 0 . The dotted lines indicate Type I error 0.05 and zero bias. Error bars show pointwise normal-approximation 95% confidence intervals. The gray band in panel (c) is the corresponding Monte Carlo reference at nominal Type I error. Calibrated power uses method- and scenario-specific empirical null critical values.
[Figure 3 about here.]
Compared with CAHB, RADISH borrows more conservatively and is more robust when the shifts vary substantially across subgroups. The two methods perform similarly when variation in subgroup-specific shifts is small. With large variation, however, CAHB shows substantial Type I error inflation: its rejection rates are 0.423 for g = 0 g = 0 g = 0 and 0.436 for g = 2 g = 2 g = 2 . The corresponding RADISH rates are 0.072 and 0.046. RADISH also has lower MAE in both scenarios, with values of 0.047 and 0.044, compared with 0.091 and 0.090 for CAHB.
The borrowing weights support this interpretation. When subgroup-specific shifts vary substantially, RADISH assigns less weight than CAHB to historical information in the X 1 = 0 X_1 = 0 X 1 = 0 subgroup, which represented approximately 69% of the HORIZON population. With zero mean shift, the mean final borrowing weights are 0.173 for RADISH and 0.416 for CAHB. With g = 2 g = 2 g = 2 , the corresponding weights are 0.095 and 0.404, respectively. These smaller weights limit the influence of historical outcomes whose conditional means differed from those of the concurrent controls.
Figure 4 shows how borrowing and performance changed as variation in subgroup-specific shifts increased while the mean shift remained zero. RADISH adapts borrowing as intended. When historical and concurrent controls are compatible (ξ = 0 \xi= 0 ξ = 0 ), RADISH reduces MAE from 0.042 for KBCD to 0.037 and increases calibrated power from 0.743 to 0.827. As compatibility decreases, RADISH reduces borrowing, and its gains over KBCD diminish. Although RADISH’s MAE is slightly higher than KBCD’s in some scenarios, RADISH retains gains in calibrated power over the range of ξ \xi ξ . At ξ = 1 \xi= 1 ξ = 1 , its mean final borrowing weight is only 0.0016, and its estimation accuracy and calibrated power approach those of KBCD. The control allocation proportion is 0.499, indicating nearly equal allocation.
Figure 4. Performance for HORIZON–FIT-based semi-synthetic simulation studies as subgroup-specific shifts increase while the mean shift remains zero, with N = 200 N = 200 N = 200 , N H = 350 N_H = 350 N H = 350 , and 1000 replications per scenario. The standardized shifts are ( b ( 0 ) , b ( 1 ) ) / σ C = 4 ξ − 1 , ( 1 − p 1 ) / p 1 (b(0), b(1))/\sigma_C = 4\xi^{-1}, (1-p_1)/p_1 ( b ( 0 ) , b ( 1 )) / σ C = 4 ξ − 1 , ( 1 − p 1 ) / p 1 . For ξ > 0 \xi> 0 ξ > 0 , the two shifts have opposite signs. Positive values of ξ \xi ξ are displayed on a logarithmic scale, with ξ = 0 \xi= 0 ξ = 0 shown at the left. Panels show (a) mean absolute error under δ = 0.14 \delta= 0.14 δ = 0.14 , (b) calibrated power, and (c) the proportion assigned to the control group. Curves show point estimates, and shaded regions show pointwise 95% confidence intervals.
Supplementary Figure S5 shows how borrowing and control allocation adapt to subgroup-specific compatibility. Repeating the second set of scenarios with N H = 30 N_H = 30 N H = 30 yielded smaller efficiency gains, while Type I error remained below 0.05 across scenarios (Supplementary Figure S6).
Figure S5. Subgroup-specific borrowing diagnostics for the HORIZON–FIT-based semi-synthetic simulation, with N = 200 N = 200 N = 200 , N H = 350 N_H = 350 N H = 350 , δ = 0.14 \delta= 0.14 δ = 0.14 , and 1000 replications per scenario. The standardized shifts are ( b ( 0 ) , b ( 1 ) ) / σ C = 4 ξ { − 1 , ( 1 − p 1 ) / p 1 } (b(0), b(1))/\sigma_C = 4\xi\{ -1, (1-p_1)/p_1 \} ( b ( 0 ) , b ( 1 )) / σ C = 4 ξ { − 1 , ( 1 − p 1 ) / p 1 } , where p 1 ≈ 0.310 p_1 \approx0.310 p 1 ≈ 0.310 . Their magnitudes increase with ξ \xi ξ while the mean shift remains zero. Panels (a) and (b) show mean final borrowing weights for the X 1 = 0 X_1 = 0 X 1 = 0 and X 1 = 1 X_1 = 1 X 1 = 1 subgroups, respectively. KBCD provides the zero-borrowing reference. Panels (c) and (d) show the corresponding proportions assigned to the control group. The dotted line marks equal allocation at 0.5. Error bars show pointwise normal-approximation 95% confidence intervals. Positive values of ξ \xi ξ are displayed on a logarithmic scale, with ξ = 0 \xi= 0 ξ = 0 shown at the left.
[Figure 4 about here.]
CAHB borrows more aggressively than RADISH. Under compatibility, CAHB produces larger gains, with MAE 0.035 and calibrated power 0.883. However, CAHB continues to borrow in some scenarios where the historical and concurrent conditional control means differed substantially, resulting in large bias and low calibrated power. At ξ = 0.316228 \xi= 0.316228 ξ = 0.316228 , its MAE is 0.077 and Type I error is 0.317, compared with 0.045 and 0.048 for RADISH. At ξ = 1 \xi= 1 ξ = 1 , CAHB also reduces borrowing and approaches KBCD.
Discussion
RADISH incorporates historical controls into covariate-adaptive randomization and final inference, adjusting borrowing to local compatibility and historical precision. Precise, compatible information can improve efficiency and reduce control enrollment, while increasing uncertainty or conflict reduces borrowing. Under the stated regularity conditions, the Studentized RADISH estimator is asymptotically normal, supporting Wald inference. Simulations show estimation and power gains with precise, compatible historical information and a return toward concurrent-only performance as incompatibility increases. Semi-synthetic studies based on HORIZON ((Black et al., 2007)) and FIT ((Black et al., 1996)) yield qualitatively similar results. The 24-month BMD endpoint illustrates a practical constraint: outcome-dependent allocation requires sufficiently slow recruitment or delayed updates; the simulations assume outcomes are available before the next assignment.
The local Gaussian working model currently accommodates continuous outcomes. Extending its likelihood and conflict measure to binary, count, and time-to-event endpoints remains to be studied. RADISH also uses a single historical source, although multiple studies with differing compatibility may be available (Jin and Yin, 2021; Lesaffre et al., 2024; Zhang and Yin, 2023). Source-specific compatibility assessments and weights could accommodate such studies. The method assumes consistently defined and measured covariates across sources and adequate historical support over the concurrent covariate space. Extensions addressing limited overlap, inconsistent definitions, and sparse historical information would broaden its applicability.
Conflict of Interest
The authors declare no conflicts of interest.
Data Availability
The HORIZON and FIT participant-level data are subject to the original data-use agreements and are not distributed with this article. Access requires authorization under the applicable data-use agreements. Simulation and analysis code is available at https://github.com/BobZhangHT/HistInfoBorrow .
Supplementary Material
Supplementary Sections S1–§ 5.7 , Table S1 , and Figure S1 –Figure S6 referenced in § 1.2 –§ 2.2 are available with this paper at the Biometrics website on Oxford Academic.
Figure S1. The RADISH trial-design procedure. Stage I constructs the local historical reference before enrollment. Stage II begins with equal-probability burn-in randomization for the first n i n i t n_{\mathrm{init}} n init participants and then updates the allocation sequentially using accrued outcomes. Stage III uses the completed trial data for final estimation and Wald inference.
Table S1. Estimation, interval, allocation, and borrowing performance under Δ = 0 \Delta= 0 Δ = 0 , based on M = 1000 M = 1000 M = 1000 paired replications with N H = N = 200 N_{H}=N=200 N H = N = 200 . Scenarios S1–S8 combine standardized historical shifts ( b ( 0 ) , b ( 1 ) ) = ( 0 , 0 ) , ( 0.04 , 0.12 ) , ( 0.25 , 0.75 ) , ( 1.5 , 2.5 ) (b(0), b(1)) = (0, 0), (0.04, 0.12), (0.25, 0.75), (1.5, 2.5) ( b ( 0 ) , b ( 1 )) = ( 0 , 0 ) , ( 0.04 , 0.12 ) , ( 0.25 , 0.75 ) , ( 1.5 , 2.5 ) with σ H = 0.5 , 3.0 \sigma_{H}=0.5, 3.0 σ H = 0.5 , 3.0 , respectively. RMSE is root mean squared error; Width is mean 95% interval width; Alloc. Proportion is the proportion assigned to the control group; W ‾ N F \overline{W}^{F}_{N} W N F is the mean Stage III borrowing weight.
Figure S6. Sensitivity analysis of the historical sample size N H N_H N H for the HORIZON–FIT-based semi-synthetic simulation as subgroup-specific shifts increase while the mean shift remains zero, with N = 200 N=200 N = 200 and 1000 replications per scenario. The left and right columns use N H = 30 N_H=30 N H = 30 and N H = 350 N_H=350 N H = 350 , respectively. The standardized shifts are ( b ( 0 ) , b ( 1 ) ) / σ C = 4 ξ − 1 { ( 1 − p 1 ) / p 1 } (b(0),b(1))/\sigma_C=4\xi^{-1}\{(1-p_1)/p_1\} ( b ( 0 ) , b ( 1 )) / σ C = 4 ξ − 1 {( 1 − p 1 ) / p 1 } , where p 1 ≈ 0.310 p_1\approx0.310 p 1 ≈ 0.310 . For ξ > 0 \xi>0 ξ > 0 , the two shifts have opposite signs. Positive values of ξ \xi ξ are displayed on a logarithmic scale, with ξ = 0 \xi=0 ξ = 0 shown at the left. Panels (a) and (b) show relative efficiency under δ = 0.14 \delta=0.14 δ = 0.14 , defined as ( R M S E K B C D / R M S E ) 2 (\mathrm{RMSE}_{\mathrm{KBCD}}/\mathrm{RMSE})^2 ( RMSE KBCD / RMSE ) 2 . Panels (c) and (d) show empirical Type I error. Dotted lines mark relative efficiency 1 and Type I error 0.05. The gray band shows the normal-approximation 95% Monte Carlo confidence interval under Type I error 0.05.
Supplementary Materials for RADISH: A Robust Covariate-Adaptive Trial Design with Historical Control Borrowing
Hengtao Zhang; Yuanke Qu; Mi-Ok Kim; Huaqing Jin
supplement17.pdf
Overview
Figure S1 summarizes the RADISH trial-design procedure. These supplementary materials provide additional numerical results, followed by technical details and proofs.
In § 5.2 , we report the results of the simulation study. Table S1 reports performance under the null hypothesis Δ = 0 \Delta= 0 Δ = 0 , and Figure S2 shows the allocation proportion to the control group. Figure S3 examines borrowing under different overall mean shifts and variation in subgroup-specific shifts. Figure S4 assesses sensitivity to historical sample size.
Figure S3. Marginal-shift-by-local-heterogeneity diagnostic based on 1000 paired replications with N H = N = 200 N_H = N = 200 N H = N = 200 and σ H = 0.5 \sigma_H = 0.5 σ H = 0.5 . The four horizontal-axis categories cross weak/strong marginal shift with weak/strong local heterogeneity. The horizontal-axis labels use g g g for the marginal shift g = E { b ( X 1 ) } g = \mathrm{E}\{b(X_1)\} g = E { b ( X 1 )} and ℓ \ell ℓ for the heterogeneity parameter. In plot order, the four pairs ( b ( 0 ) , b ( 1 ) ) (b(0), b(1)) ( b ( 0 ) , b ( 1 )) are ( − 0.05 , 0.10 ) (-0.05, 0.10) ( − 0.05 , 0.10 ) , ( 0.55 , 0.70 ) (0.55, 0.70) ( 0.55 , 0.70 ) , ( − 0.60 , 1.20 ) (-0.60, 1.20) ( − 0.60 , 1.20 ) , and ( 0 , 1.80 ) (0, 1.80) ( 0 , 1.80 ) . Here Δ = 0.5 \Delta= 0.5 Δ = 0.5 . In the g + ℓ + g^{+}\ell^{+} g + ℓ + category, ( b ( 0 ) , b ( 1 ) ) = ( 0 , 1.80 ) (b(0), b(1)) = (0, 1.80) ( b ( 0 ) , b ( 1 )) = ( 0 , 1.80 ) , so X 1 = 0 X_1 = 0 X 1 = 0 remains compatible even though the marginal shift and between-stratum heterogeneity are strong. Panels show (a) terminal borrowing weights by X 1 X_1 X 1 stratum, (b) the allocation proportion to the control group by stratum, (c) RMSE under Δ = 0.5 \Delta= 0.5 Δ = 0.5 , and (d) calibrated power using the cell-specific empirical null critical value. Solid and dashed curves denote X 1 = 0 X_1 = 0 X 1 = 0 and X 1 = 1 X_1 = 1 X 1 = 1 .
Figure S4. Sensitivity of the simulation results to doubling the historical sample size from N H = 200 N_H = 200 N H = 200 to 400, based on 1000 paired replications per cell with N = 200 N = 200 N = 200 . Columns correspond to the two historical sample sizes. Panels (a) and (b) show relative efficiency, ( R M S E K B C D / R M S E ) 2 (\mathrm{RMSE}_{\mathrm{KBCD}}/\mathrm{RMSE})^2 ( RMSE KBCD / RMSE ) 2 , so values above one favor the borrowing method. Panels (c) and (d) show calibrated-power differences relative to KBCD, so values above zero favor the borrowing method. The dashed lines are the KBCD references. S1–S2 are compatible, S3–S4 have small local conflict, S5–S6 have moderate local conflict, and S7–S8 have severe local conflict; within each pair, the first and second scenarios use σ H = 0.5 \sigma_H = 0.5 σ H = 0.5 and 3.0, respectively. All nonzero conflicts depend on X 1 X_1 X 1 . KBCD is concurrent-only, CAHB is the borrowing comparator, and RADISH is the proposed method.
In § 5.3 , we report the results of the HORIZON–FIT-based semi-synthetic study. Figure S5 shows subgroup-specific borrowing weights and the allocation proportion to the control group, and Figure S6 examines sensitivity to historical sample size.
In § 5.4 , we provide additional details for Stage III. In § 5.5 , we provide the asymptotic theory and proofs. § 5.7 gives detailed Algorithms S1–S3 for the three-stage overview in Algorithm 1 of the main text.
Additional Results for the Simulation Study
Null-Hypothesis Results and Control Allocation
In the main text, we reported the results of the simulation study under the alternative hypothesis Δ = 0.5 \Delta= 0.5 Δ = 0.5 . Table S1 summarizes estimation accuracy, confidence-interval coverage and width, and borrowing under the null hypothesis Δ = 0 \Delta= 0 Δ = 0 . Figure S2 compares the proportions assigned to the control group across the eight simulation scenarios.
Mean Shift and Variation in Subgroup-Specific Shifts
To distinguish the marginal shift from local heterogeneity, we write b ( X 1 ) = g + ℓ s ( X 1 ) b(X_{1}) = g + \ell s(X_{1}) b ( X 1 ) = g + ℓ s ( X 1 ) with P ( X 1 = 1 ) = 1 / 3 P(X_{1}=1)=1/3 P ( X 1 = 1 ) = 1/3 , s ( 0 ) = − 1 / 2 s(0)=-1/2 s ( 0 ) = − 1/2 , and s ( 1 ) = 1 s(1)=1 s ( 1 ) = 1 , so E { s ( X 1 ) } = 0 E\{s(X_{1})\}=0 E { s ( X 1 )} = 0 and E { b ( X 1 ) } = g E\{b(X_{1})\}=g E { b ( X 1 )} = g . Thus g g g is the marginal historical shift and ℓ \ell ℓ controls how strongly that shift varies between the two covariate strata. Each ordered pair below is ( b ( 0 ) , b ( 1 ) ) (b(0), b(1)) ( b ( 0 ) , b ( 1 )) : its first entry is the historical mean shift for X 1 = 0 X_{1}=0 X 1 = 0 , and its second entry is the shift for X 1 = 1 X_{1}=1 X 1 = 1 , both relative to the concurrent control mean. For example, ( − 0.05 , 0.10 ) (-0.05, 0.10) ( − 0.05 , 0.10 ) means a shift of − 0.05 -0.05 − 0.05 in the X 1 = 0 X_{1}=0 X 1 = 0 stratum and 0.10 0.10 0.10 in the X 1 = 1 X_{1}=1 X 1 = 1 stratum; these shifts average to zero under the target distribution.
Crossing g ∈ { 0 , 0.6 } g \in\{0,0.6\} g ∈ { 0 , 0.6 } with ℓ ∈ { 0.1 , 1.2 } \ell\in\{0.1,1.2\} ℓ ∈ { 0.1 , 1.2 } gives four regimes: weak marginal shift and weak local heterogeneity ( − 0.05 , 0.10 ) (-0.05,0.10) ( − 0.05 , 0.10 ) , strong marginal shift and weak local heterogeneity ( 0.55 , 0.70 ) (0.55,0.70) ( 0.55 , 0.70 ) , weak marginal shift and strong local heterogeneity ( − 0.60 , 1.20 ) (-0.60,1.20) ( − 0.60 , 1.20 ) , and strong marginal shift and strong local heterogeneity ( 0 , 1.80 ) (0,1.80) ( 0 , 1.80 ) . In the last regime, the overall shift and the between-stratum contrast are both strong: X 1 = 0 X_1=0 X 1 = 0 remains exactly compatible, whereas X 1 = 1 X_1=1 X 1 = 1 has a shift of 1.80. Every regime remains covariate dependent, while g g g and ℓ \ell ℓ vary independently. Here N H = N = 200 N_H=N=200 N H = N = 200 and σ H = 0.5 \sigma_H=0.5 σ H = 0.5 ; the remaining settings match the primary simulation, and each cell uses 1000 paired replications.
Figure S3 illustrates the behavior of the combined score: broadly distributed conflict activates the marginal safeguard, while the local score distinguishes covariate strata when marginal averaging masks heterogeneous conflict. In the two regimes where only one conflict axis is strong, RADISH remains close to the concurrent-only RMSE and calibrated-power references, whereas CAHB is sensitive to the unresolved conflict. In the strong-marginal-shift, strong-local-heterogeneity regime, RADISH retains mean weights 0.328/0.001 for X 1 = 0 / 1 X_1=0/1 X 1 = 0/1 . The nonzero weight in the first stratum is expected because b ( 0 ) = 0 b(0)=0 b ( 0 ) = 0 , whereas the nearly zero second weight reflects the severe local conflict b ( 1 ) = 1.80 b(1)=1.80 b ( 1 ) = 1.80 . In the weak-marginal-shift, strong-local-heterogeneity regime, the corresponding weights are 0.080/0.021. These results support the rationale for combining local and marginal conflict through a local gate.
Sensitivity Analysis of the Historical Sample Size
The primary simulation fixes N H = 200 N_H = 200 N H = 200 . We repeated the eight scenarios with N H = 400 N_H = 400 N H = 400 using nested historical cohorts and paired concurrent-trial random-number streams, as shown in Figure S4 . Doubling the historical cohort produces a modest additional efficiency gain under compatible and small-conflict settings, whereas RADISH remains close to KBCD under moderate and severe conflict because its borrowing weights decrease as incompatibility becomes detectable. Thus the larger historical cohort increases the gains from compatible borrowing while preserving the return toward concurrent-only analysis as conflict increases.
Additional Results for the Semi-Synthetic Simulation Study
Figure S5 complements the aggregate operating characteristics in the main text by separating the two X 1 X_1 X 1 regions. It shows how subgroup-specific borrowing and randomization produce the overall response along the localized-conflict path.
The main HORIZON–FIT conflict path uses N H = 350 N_H = 350 N H = 350 . This sensitivity analysis repeats the same seven levels ξ ∈ { 0 , 0.003162 , 0.01 , 0.031623 , 0.1 , 0.316228 , 1 } \xi\in\{0, 0.003162, 0.01, 0.031623, 0.1, 0.316228, 1\} ξ ∈ { 0 , 0.003162 , 0.01 , 0.031623 , 0.1 , 0.316228 , 1 } at N H = 30 N_H = 30 N H = 30 , as shown in Figure S6 . Along the path, ( b ( 0 ) , b ( 1 ) ) / σ C = 4 ξ { − 1 , ( 1 − p 1 ) / p 1 } (b(0), b(1))/\sigma_C = 4\xi\{-1, (1-p_1)/p_1\} ( b ( 0 ) , b ( 1 )) / σ C = 4 ξ { − 1 , ( 1 − p 1 ) / p 1 } , so every positive ξ \xi ξ remains covariate dependent while the marginal shift is zero. The comparison assesses whether the local-conflict response persists at lower historical precision.
Additional Computational Details
Historical Reference Variance
Because θ ( x ) \theta(x) θ ( x ) is a linear combination of conditionally independent observations, its variance has a closed form under the working model after conditioning on the historical covariates { X j H } j = 1 N H \{X_j^H\}_{j=1}^{N_H} { X j H } j = 1 N H :
τ H 2 ( x ) = σ H 2 ( x ) ∑ j = 1 N H { S H , j ( x ) } 2 = σ H 2 ( x ) n H e f f ( x ) , n H e f f ( x ) = { ∑ j w j H ( x ) } 2 ∑ j { w j H ( x ) } 2 , \tau_H^2(x)=\sigma_H^2(x)\sum_{j=1}^{N_H}\{S_{H,j}(x)\}^2=\frac{\sigma_H^2(x)}{n_H^{\mathrm{eff}}(x)},\qquad n_H^{\mathrm{eff}}(x)=\frac{\left\{\sum_j w_j^H(x)\right\}^2}{\sum_j\{w_j^H(x)\}^2}, τ H 2 ( x ) = σ H 2 ( x ) j = 1 ∑ N H { S H , j ( x ) } 2 = n H eff ( x ) σ H 2 ( x ) , n H eff ( x ) = ∑ j { w j H ( x ) } 2 { ∑ j w j H ( x ) } 2 ,
where n H e f f ( x ) n_H^{\mathrm{eff}}(x) n H eff ( x ) is the local effective historical sample size for constructing θ ( x ) \theta(x) θ ( x ) . The unknown σ H 2 ( x ) \sigma_H^2(x) σ H 2 ( x ) is replaced by the weighted residual estimator
σ ^ H 2 ( x ) = [ ∑ j w j H ( x ) { Y j H − θ ( x ) } 2 ] / { ∑ j w j H ( x ) } . \widehat{\sigma}_H^2(x)=\left[\sum_j w_j^H(x)\{Y_j^H-\theta(x)\}^2\right]\Big/\left\{\sum_j w_j^H(x)\right\}. σ H 2 ( x ) = [ j ∑ w j H ( x ) { Y j H − θ ( x ) } 2 ] / { j ∑ w j H ( x ) } .
Substituting this estimator gives τ ^ H 2 ( x ) = σ ^ H 2 ( x ) / n H e f f ( x ) \widehat{\tau}_H^2(x)=\widehat{\sigma}_H^2(x)/n_H^{\mathrm{eff}}(x) τ H 2 ( x ) = σ H 2 ( x ) / n H eff ( x ) .
Information-Ratio Interpretation
The increment above one reflects both the precision and compatibility of the historical reference. The variance ratio { σ ^ 0 , n 2 ( x ) / n ^ 0 , n e f f ( x ) } / S E n 2 ( x ) \{\widehat{\sigma}_{0,n}^2(x)/\widehat{n}_{0,n}^{\mathrm{eff}}(x)\}/\mathrm{SE}_n^2(x) { σ 0 , n 2 ( x ) / n 0 , n eff ( x )} / SE n 2 ( x ) compares the uncertainty in the concurrent local control mean with the total uncertainty in the historical–concurrent comparison. For a fixed concurrent sampling variance, this ratio is larger when the historical reference is estimated more precisely and decreases as its estimation uncertainty increases. The factor exp { − D n ( x ) } \exp\{-D_n(x)\} exp { − D n ( x )} further discounts historical information as local incompatibility increases. In particular, it equals one within the tolerance region where D n ( x ) = 0 D_n(x)=0 D n ( x ) = 0 and decreases as the historical and concurrent local means become less compatible.
Thus, a larger R n A ( x ) R_n^A(x) R n A ( x ) corresponds to greater historical borrowing when the historical reference is both precise and locally compatible with the concurrent controls. As the historical reference becomes less precise or less compatible, R n A ( x ) R_n^A(x) R n A ( x ) approaches one, corresponding to progressively weaker historical borrowing.
Residual Variance Calculations
We estimate these source-specific outcome variances using residuals from the same kernel smoothing procedure used to construct the point estimator. The ordinary within-source sample variance is not appropriate here because it contains both variation in the conditional mean across covariates and residual variation around that mean.
For z ∈ { 0 , 1 } z\in\{0,1\} z ∈ { 0 , 1 } , let X z X_z X z and Y z Y_z Y z denote the covariate matrix and outcome vector for concurrent participants in arm z z z , and define the within-arm smoothing matrix
S z z = normRows { K h ( X z , X z ) } , S_{zz}=\operatorname{normRows}\left\{K_h(X_z,X_z)\right\}, S z z = normRows { K h ( X z , X z ) } ,
where normRows ( S ) \operatorname{normRows}(S) normRows ( S ) divides each row of S S S by its row sum. Here, K h \mathbf{K}_h K h is the pairwise kernel matrix whose ( i , j ) (i,j) ( i , j ) th entry is { K h ( X z , X z ) } i j = K h ( x i − x j ) \{\mathbf{K}_h(\mathbf{X}_z,\mathbf{X}_z)\}_{ij}=K_h(\mathbf{x}_i-\mathbf{x}_j) { K h ( X z , X z ) } ij = K h ( x i − x j ) , where x i \mathbf{x}_i x i and x j \mathbf{x}_j x j denote the corresponding row vectors.
The fitted outcomes are S z z Y z S_{zz}Y_z S z z Y z , and we estimate the residual variance by
σ ^ z 2 = ∥ Y z − S z z Y z ∥ 2 2 max { N z − 2 tr ( S z z ) + tr ( S z z T S z z ) , 1 } , z ∈ { 0 , 1 } . \widehat{\sigma}_z^2=\frac{\lVert Y_z-S_{zz}Y_z\rVert_2^2}{\max\{N_z-2\operatorname{tr}(S_{zz})+\operatorname{tr}(S_{zz}^{\mathsf{T}}S_{zz}),1\}},\qquad z\in\{0,1\}. σ z 2 = max { N z − 2 tr ( S z z ) + tr ( S z z T S z z ) , 1 } ∥ Y z − S z z Y z ∥ 2 2 , z ∈ { 0 , 1 } .
The denominator is the residual degrees of freedom associated with the linear smoother, since
tr [ ( I N z − S z z ) T ( I N z − S z z ) ] = N z − 2 tr ( S z z ) + tr ( S z z T S z z ) . \operatorname{tr}\left[(I_{N_z}-S_{zz})^{\mathsf{T}}(I_{N_z}-S_{zz})\right]=N_z-2\operatorname{tr}(S_{zz})+\operatorname{tr}(S_{zz}^{\mathsf{T}}S_{zz}). tr [ ( I N z − S z z ) T ( I N z − S z z ) ] = N z − 2 tr ( S z z ) + tr ( S z z T S z z ) .
The lower bound of one is a finite-sample safeguard that ensures a positive residual-degrees-of-freedom denominator.
The historical residual variance is estimated analogously. Let X H \mathbf{X}_H X H and Y H Y_H Y H denote the historical covariate matrix and outcome vector, and define S H H = normRows { K h ( X H , X H ) } S_{HH}=\operatorname{normRows}\{K_h(\mathbf{X}_H,\mathbf{X}_H)\} S H H = normRows { K h ( X H , X H )} . Then
σ ^ H 2 = ∥ Y H − S H H Y H ∥ 2 2 max { N H − 2 tr ( S H H ) + tr ( S H H T S H H ) , 1 } . \widehat{\sigma}_H^2=\frac{\lVert Y_H-S_{HH}Y_H\rVert_2^2}{\max\{N_H-2\operatorname{tr}(S_{HH})+\operatorname{tr}(S_{HH}^{\mathsf{T}}S_{HH}),1\}}. σ H 2 = max { N H − 2 tr ( S H H ) + tr ( S H H T S H H ) , 1 } ∥ Y H − S H H Y H ∥ 2 2 .
Here σ ^ H 2 \widehat{\sigma}_H^2 σ H 2 is the source-wide residual variance used for final inference and is distinct from the local Stage I variance σ ^ H 2 ( x ) \widehat{\sigma}_H^2(x) σ H 2 ( x ) . For numerical evaluation, residual variances and the computed values of V ^ l i n \widehat{V}_{\mathrm{lin}} V lin and V ^ C \widehat{V}_C V C are floored at ϵ v = 10 − 12 \epsilon_v=10^{-12} ϵ v = 1 0 − 12 before forming their final maximum.
Implementation Properties
For Scott’s rule, the historical covariate standard deviation σ ^ H , k \widehat{\sigma}_{H,k} σ H , k is replaced by one if it is less than 10 − 12 10^{-12} 1 0 − 12 , keeping the bandwidth positive. Several properties of the kernel calculations are also useful for implementation and interpretation. Multiplying the kernel by any positive constant leaves all row-normalized smoothing weights and local effective sample sizes unchanged. The same multiplier applies to both arm-specific kernel masses in the randomization probability and therefore cancels. Thus a common rescaling of the kernel does not change the local means, information ratios, borrowing weights, randomization probabilities, or final estimator.
Within each data source, the outcome vector, covariate matrix, and smoothing coefficients use the same observation ordering. This convention ensures that the entries of the cumulative coefficient vectors multiply the corresponding outcomes in the weighted-sum representation of the estimator.
The trace in the residual degrees of freedom has a direct working-model interpretation. For a fixed smoother S S S and Y = μ + ϵ Y=\mu+\epsilon Y = μ + ϵ , with independent mean-zero homoskedastic errors of variance σ 2 \sigma^2 σ 2 ,
E { ∥ ( I − S ) Y ∥ 2 2 } = ∥ ( I − S ) μ ∥ 2 2 + σ 2 tr { ( I − S ) T ( I − S ) } . \mathbb{E}\{\lVert(I-S)Y\rVert_2^2\}=\lVert(I-S)\mu\rVert_2^2+\sigma^2\operatorname{tr}\{(I-S)^{\mathsf{T}}(I-S)\}. E {∥( I − S ) Y ∥ 2 2 } = ∥( I − S ) μ ∥ 2 2 + σ 2 tr {( I − S ) T ( I − S )} .
The trace therefore normalizes the noise contribution to the residual sum of squares; consistency also requires the squared mean-approximation error, divided by the residual degrees of freedom, to vanish. For RADISH, conditions (C4) and (C6) in § 5.5.2 transfer this argument from the outcome-independent oracle smoother to the realized adaptive smoother.
The probability floor bounds the conflict scores whenever their standardized discrepancies are defined, and the local variance floors keep the local standard errors positive. Well-definedness of the marginal discrepancy is imposed separately in condition (C4). The resulting exponential discount is strictly positive, although it can be numerically negligible under severe conflict.
Asymptotic Theory and Proofs
Large-Sample Properties
The asymptotic analysis conditions on both the historical dataset and the enrolled covariates. Under this conditional formulation, Δ N \Delta_N Δ N in Equation (1) of the main text is fixed, and repeated-trial uncertainty arises from sequential randomization and the outcome errors. The working variance therefore quantifies uncertainty for the treatment effect in the enrolled cohort conditional on the observed covariate distribution.
The central limit theorem accounts for the dependence of RADISH allocation probabilities and terminal borrowing weights on concurrent control outcomes. The smoothing coefficients in Δ ^ N \widehat{\Delta}_N Δ N are therefore dependent on the outcomes they multiply. We separate this dependence from the outcome variation used for inference through a hypothetical concurrent-only version of RADISH. This oracle retains the covariate-adaptive sequential structure and sets historical borrowing to zero in allocation and final estimation. Comparing the actual and oracle paths under common randomization draws isolates the effect of historical borrowing on allocation and estimation. Under the coupling rates stated below, the oracle large-sample distribution also describes the RADISH estimator to first order.
Let U 1 , … , U N U_1,\ldots,U_N U 1 , … , U N be independent standard uniform variables, independent of the historical data, enrolled covariates, and potential-outcome errors. These variables represent the random draws used to implement the sequential allocation probabilities, so the actual assignment has the exact representation Z i = 1 { U i ≤ ϕ i − 1 ( X i ) } Z_i = 1\{U_i \le\phi_{i-1}(X_i)\} Z i = 1 { U i ≤ ϕ i − 1 ( X i )} . Using the same U i U_i U i for the actual and oracle paths couples their assignments participant by participant. Conditional on their allocation probabilities, the two assignments differ only when U i U_i U i falls between those probabilities.
On this common probability space, the oracle uses the same burn-in and the same U i U_i U i , but sets the historical information ratio to one. After burn-in, its treatment probability is
ϕ i − 1 ∘ ( X i ) = { ∑ k < i K h ( X k − X i ) 1 ( Z k ∘ = 0 ) } 2 { ∑ k < i K h ( X k − X i ) 1 ( Z k ∘ = 0 ) } 2 + { ∑ k < i K h ( X k − X i ) 1 ( Z k ∘ = 1 ) } 2 \phi^\circ_{i-1}(X_i)=\frac{\left\{\sum_{k<i}K_h(X_k-X_i)1(Z_k^\circ=0)\right\}^{2}}{\left\{\sum_{k<i}K_h(X_k-X_i)1(Z_k^\circ=0)\right\}^{2}+\left\{\sum_{k<i}K_h(X_k-X_i)1(Z_k^\circ=1)\right\}^{2}} ϕ i − 1 ∘ ( X i ) = { ∑ k < i K h ( X k − X i ) 1 ( Z k ∘ = 0 ) } 2 + { ∑ k < i K h ( X k − X i ) 1 ( Z k ∘ = 1 ) } 2 { ∑ k < i K h ( X k − X i ) 1 ( Z k ∘ = 0 ) } 2
and Z i ∘ = 1 { U i ≤ ϕ i − 1 ∘ ( X i ) } Z_i^\circ= 1\{U_i \le\phi^\circ_{i-1}(X_i)\} Z i ∘ = 1 { U i ≤ ϕ i − 1 ∘ ( X i )} . The oracle probability depends only on the covariates and preceding oracle assignments.
Applying the concurrent part of Stage III to the oracle path gives α 1 N ∘ , α 0 N ∘ ∈ R N \alpha^\circ_{1N},\alpha^\circ_{0N}\in\mathbb{R}^N α 1 N ∘ , α 0 N ∘ ∈ R N . Each vector has one position for every enrolled participant. The i i i th entry of α z N ∘ \alpha^\circ_{zN} α z N ∘ is zero if oracle participant i i i is not assigned to arm z z z and otherwise equals that participant’s averaged kernel coefficient. Using the common participant index 1 , … , N 1,\ldots,N 1 , … , N , the oracle estimator is Δ ^ N ∘ = ∑ i = 1 N α 1 N , i ∘ Y i ( 1 ) − ∑ i = 1 N α 0 N , i ∘ Y i ( 0 ) \widehat{\Delta}^\circ_N=\sum_{i=1}^{N}\alpha^\circ_{1N,i}Y_i(1)-\sum_{i=1}^{N}\alpha^\circ_{0N,i}Y_i(0) Δ N ∘ = ∑ i = 1 N α 1 N , i ∘ Y i ( 1 ) − ∑ i = 1 N α 0 N , i ∘ Y i ( 0 ) . Conditional on the enrolled covariates and the variables U 1 , … , U N U_1,\ldots,U_N U 1 , … , U N , the oracle coefficients are fixed with respect to the outcome errors. The stochastic part of Δ ^ N ∘ \widehat{\Delta}^\circ_N Δ N ∘ is therefore a weighted sum of independent, mean-zero errors, which permits a conditional weighted central limit theorem. The proof then establishes first-order equivalence between the RADISH estimator and this oracle estimator.
Proposition S1. (Large-sample reversion to concurrent-only design and analysis). Suppose conditions ( C 1 ) (C1) ( C 1 ) , ( C 2 ) (C2) ( C 2 ) , and ( C 4 ) (C4) ( C 4 ) in § 5.5.2 hold. Then
1 N ∑ i = 1 N 1 ( Z i ≠ Z i ∘ ) ⟶ p 0 , 1 N ∑ i = 1 N W N F ( X i ) ⟶ p 0 , \frac{1}{N}\sum_{i=1}^{N}\mathbf{1}(Z_i \ne Z_i^\circ) \overset{p}{\longrightarrow} 0,\qquad\frac{1}{N}\sum_{i=1}^{N}W_N^{F}(X_i) \overset{p}{\longrightarrow} 0, N 1 i = 1 ∑ N 1 ( Z i = Z i ∘ ) ⟶ p 0 , N 1 i = 1 ∑ N W N F ( X i ) ⟶ p 0 , and
Δ ^ N − Δ ^ N ∘ = o p ( N − 1 / 2 ) . \widehat{\Delta}_N-\widehat{\Delta}_N^\circ=o_p(N^{-1/2}). Δ N − Δ N ∘ = o p ( N − 1/2 ) .
Proposition S1 has implications for both trial design and inference. Because the two paths use the same randomization variables, the first limit measures the effect of historical borrowing on treatment allocation. This limit shows that borrowing changes the assignments of a vanishing proportion of participants. The second limit shows that the average contribution of the fixed historical dataset to the final control analysis vanishes as concurrent information accumulates. The third limit gives the rate needed for inference: the difference between the RADISH and oracle estimators is smaller than the root-N N N sampling scale, allowing them to share the same first-order limit under the variance conditions stated below. Together, these limits describe how borrowing changes with concurrent enrollment. Historical information can influence allocation and contribute additional precision at trial sizes used in practice. With a fixed historical dataset, its contribution vanishes as concurrent information accumulates. The proposition links this gradual return to concurrent-only analysis with the proof of asymptotic normality. The empirical results quantify the finite-sample contribution of borrowing.
Theorem 1 of the main text combines the conditional central limit theorem for the oracle with first-order equivalence of the RADISH and oracle estimators, stability of their coefficient vectors, and consistency of the working variance. Thus the proof explicitly tracks the outcome dependence of the RADISH allocation, borrowing weights, and realized coefficients. The regularity conditions and proofs are given in § 5.5.2 –§ 5.6.4 .
Corollary S1. (Large-sample validity of Wald inference). Under the conditions of Theorem 1, the interval in Equation (8) of the main text has conditional limiting coverage 1 − α 1-\alpha 1 − α . The two-sided Wald test has conditional limiting size α \alpha α under H 0 : Δ N = 0 H_0:\Delta_N=0 H 0 : Δ N = 0 .
Asymptotic Framework and Regularity Conditions
The theory considers a triangular sequence of concurrent trials indexed by N N N , conditional on the historical dataset D H \mathcal{D}_H D H observed before enrollment and on the enrolled covariates X 1 : N \mathbf{X}_{1:N} X 1 : N . Thus N H N_H N H , the historical outcomes, and Δ N \Delta_N Δ N are fixed in the conditional analysis, and the variance targets randomization and outcome uncertainty for the enrolled cohort. The argument also conditions on the auxiliary randomization variables when fixing the oracle coefficients. Unless stated otherwise, all probabilities, expectations, stochastic orders, and limits in this asymptotic analysis are understood under the law conditional on ( D H , X 1 : N ) (\mathcal{D}_H,\mathbf{X}_{1:N}) ( D H , X 1 : N ) . Let F i − 1 X = σ { D H , ( X k , Z k , Y k ) k < i , X i } \mathcal{F}^{X}_{i-1}=\sigma\{\mathcal{D}_H,(\mathbf{X}_k,Z_k,Y_k)_{k<i},\mathbf{X}_i\} F i − 1 X = σ { D H , ( X k , Z k , Y k ) k < i , X i } denote the information available immediately before participant i i i is randomized, augmented by the incoming covariate X i \mathbf{X}_i X i . The actual RADISH allocation is nonanticipating because Z i Z_i Z i is generated before Y i Y_i Y i is observed.
For the proofs, embed the two concurrent coefficient vectors in R N \mathbb{R}^{N} R N . Define
α ^ z N , i = { ( a z ) ∑ k = 1 i − 1 1 ( Z k = z ) , Z i = z , 0 , Z i ≠ z , z ∈ { 0 , 1 } , \widehat{\alpha}_{zN,i} =
\begin{cases}
(a_z) \sum_{k=1}^{i-1} 1(Z_k=z), & Z_i=z,\\
0, & Z_i\ne z,
\end{cases}
\qquad z\in\{0,1\}, α z N , i = { ( a z ) ∑ k = 1 i − 1 1 ( Z k = z ) , 0 , Z i = z , Z i = z , z ∈ { 0 , 1 } ,
and set α ^ H N = a H ∈ R N H \widehat{\alpha}_{HN}=\mathbf{a}_H\in\mathbb{R}^{N_H} α H N = a H ∈ R N H . The zero entries place both concurrent coefficient vectors on the common participant index, while the estimator continues to use the observed outcome from each assigned arm. With this notation,
Δ ^ N = ∑ i = 1 N α ^ 1 N , i Y i ( 1 ) − ∑ i = 1 N α ^ 0 N , i Y i ( 0 ) − ∑ j = 1 N H α ^ H N , j Y j H . \widehat{\Delta}_N=\sum_{i=1}^{N}\widehat{\alpha}_{1N,i}Y_i(1)-\sum_{i=1}^{N}\widehat{\alpha}_{0N,i}Y_i(0)-\sum_{j=1}^{N_H}\widehat{\alpha}_{HN,j}Y_j^H. Δ N = i = 1 ∑ N α 1 N , i Y i ( 1 ) − i = 1 ∑ N α 0 N , i Y i ( 0 ) − j = 1 ∑ N H α H N , j Y j H .
Exactly one of the first two terms is observed for each concurrent participant, because the unused coefficient is zero.
Conditions (C1)–(C6) are organized in the order in which they enter the proof. Condition (C4) is stated in terms of allocation-probability perturbations and normalized kernel rows. From these conditions, Lemma S4 establishes the first-order oracle equivalence used by the main theorem.
(C1) Concurrent outcomes, sequential randomization, and moments Concurrent potential outcomes satisfy Y i ( z ) = μ z ( X i ) + ε i ( z ) Y_i(z)=\mu_z(X_i)+\varepsilon_i(z) Y i ( z ) = μ z ( X i ) + ε i ( z ) for z ∈ { 0 , 1 } z\in\{0,1\} z ∈ { 0 , 1 } , and the observed concurrent outcome obeys consistency, Y i = Z i Y i ( 1 ) + ( 1 − Z i ) Y i ( 0 ) Y_i=Z_iY_i(1)+(1-Z_i)Y_i(0) Y i = Z i Y i ( 1 ) + ( 1 − Z i ) Y i ( 0 ) . Conditional on ( D H , X 1 : N ) (D_H,\mathbf{X}_{1:N}) ( D H , X 1 : N ) , the concurrent error pairs are independent across participants and satisfy E { ε i ( z ) ∣ D H , X 1 : N } = 0 \mathbb{E}\{\varepsilon_i(z)\mid D_H,\mathbf{X}_{1:N}\}=0 E { ε i ( z ) ∣ D H , X 1 : N } = 0 . Dependence between ε i ( 0 ) \varepsilon_i(0) ε i ( 0 ) and ε i ( 1 ) \varepsilon_i(1) ε i ( 1 ) for the same concurrent participant is unrestricted. For some q > 2 q>2 q > 2 and a fixed finite constant C C C , the potential outcomes satisfy, almost surely, the uniform conditional moment bound sup i , z E { ∣ Y i ( z ) ∣ q ∣ D H , X 1 : N } ≤ C \sup_{i,z}\mathbb{E}\{|Y_i(z)|^q\mid D_H,\mathbf{X}_{1:N}\}\le C sup i , z E { ∣ Y i ( z ) ∣ q ∣ D H , X 1 : N } ≤ C . The inequality ∣ a − b ∣ q ≤ 2 q − 1 ( ∣ a ∣ q + ∣ b ∣ q ) |a-b|^q\le2^{q-1}(|a|^q+|b|^q) ∣ a − b ∣ q ≤ 2 q − 1 ( ∣ a ∣ q + ∣ b ∣ q ) , together with Jensen’s inequality, gives the corresponding bound for ε i ( z ) \varepsilon_i(z) ε i ( z ) . To match the working variance in Equation (7) of the main text, assume sourcewise homoskedasticity: Var { ε i ( z ) ∣ D H , X 1 : N } = σ z 2 \operatorname{Var}\{\varepsilon_i(z)\mid D_H,\mathbf{X}_{1:N}\}=\sigma_z^2 Var { ε i ( z ) ∣ D H , X 1 : N } = σ z 2 for z ∈ { 0 , 1 } z\in\{0,1\} z ∈ { 0 , 1 } , where both variances are bounded away from zero and infinity. Treatment is assigned nonanticipatively: Pr ( Z i = 1 ∣ F i − 1 X ) = ϕ i − 1 ( X i ) \Pr(Z_i=1\mid\mathcal{F}_{i-1}^{X})=\phi_{i-1}(X_i) Pr ( Z i = 1 ∣ F i − 1 X ) = ϕ i − 1 ( X i ) and Z i ⊥ ⊥ { Y i ( 0 ) , Y i ( 1 ) } ∣ F i − 1 X Z_i\perp\!\!\!\perp\{Y_i(0),Y_i(1)\}\mid\mathcal{F}_{i-1}^{X} Z i ⊥ ⊥ { Y i ( 0 ) , Y i ( 1 )} ∣ F i − 1 X . The common randomized burn-in supplies the initial concurrent observations; condition (C1) requires positive support for both arms before the adaptive rule is used.
(C2) Outcome-independent concurrent-only KBCD oracle Let H N \mathcal{H}_N H N denote the sigma-field generated by D H D_H D H and X 1 , … , X N X_1,\ldots,X_N X 1 , … , X N . The common bandwidth and every allocation input fixed before the first randomization are H N \mathcal{H}_N H N -measurable. Let U 1 , … , U N U_1,\ldots,U_N U 1 , … , U N be independent standard uniform variables that are jointly independent of D H D_H D H , X 1 : N \mathbf{X}_{1:N} X 1 : N , and all potential-outcome errors. The actual assignment can be represented exactly as Z i = 1 { U i ≤ ϕ i − 1 ( X i ) } Z_i=1\{U_i\le\phi_{i-1}(X_i)\} Z i = 1 { U i ≤ ϕ i − 1 ( X i )} . Starting from the same burn-in, define the oracle design by setting the historical information ratio to one in the Stage II allocation rule and using the same U i U_i U i . Its assignment is Z i ∘ = 1 { U i ≤ ϕ i − 1 ∘ ( X i ) } Z_i^\circ= 1\{U_i \le\phi_{i-1}^\circ(X_i)\} Z i ∘ = 1 { U i ≤ ϕ i − 1 ∘ ( X i )} .
After burn-in, the oracle probability is
ϕ i − 1 ∘ ( X i ) = { ∑ k < i K h ( X k − X i ) 1 ( Z k ∘ = 0 ) } 2 { ∑ k < i K h ( X k − X i ) 1 ( Z k ∘ = 0 ) } 2 + { ∑ k < i K h ( X k − X i ) 1 ( Z k ∘ = 1 ) } 2 . \phi_{i-1}^\circ(X_i) =
\frac{\left\{\sum_{k<i} K_h(X_k-X_i)1(Z_k^\circ=0)\right\}^2}
{\left\{\sum_{k<i} K_h(X_k-X_i)1(Z_k^\circ=0)\right\}^2+\left\{\sum_{k<i} K_h(X_k-X_i)1(Z_k^\circ=1)\right\}^2}. ϕ i − 1 ∘ ( X i ) = { ∑ k < i K h ( X k − X i ) 1 ( Z k ∘ = 0 ) } 2 + { ∑ k < i K h ( X k − X i ) 1 ( Z k ∘ = 1 ) } 2 { ∑ k < i K h ( X k − X i ) 1 ( Z k ∘ = 0 ) } 2 .
The oracle probability is computed from the covariates and preceding oracle assignments and is therefore outcome-independent. Let G N = σ ( D H , X 1 , … , X N , U 1 , … , U N ) \mathcal{G}_N=\sigma(D_H,X_1,\ldots,X_N,U_1,\ldots,U_N) G N = σ ( D H , X 1 , … , X N , U 1 , … , U N ) . For z ∈ { 0 , 1 } z\in\{0,1\} z ∈ { 0 , 1 } , let S z N ∘ ( x ) ∈ R N S_{zN}^\circ(x)\in\mathbb{R}^N S z N ∘ ( x ) ∈ R N be the normalized kernel row based on the oracle participants assigned to arm z z z , with zeros outside that arm. Define α z N ∘ = N − 1 ∑ i = 1 N S z N ∘ ( X i ) \alpha_{zN}^\circ=N^{-1}\sum_{i=1}^N S_{zN}^\circ(X_i) α z N ∘ = N − 1 ∑ i = 1 N S z N ∘ ( X i ) for z ∈ { 0 , 1 } z\in\{0,1\} z ∈ { 0 , 1 } , and let Δ ^ N ∘ = ∑ i α 1 N , i ∘ Y i ( 1 ) − ∑ i α 0 N , i ∘ Y i ( 0 ) \widehat{\Delta}_N^\circ=\sum_i\alpha_{1N,i}^\circ Y_i(1)-\sum_i\alpha_{0N,i}^\circ Y_i(0) Δ N ∘ = ∑ i α 1 N , i ∘ Y i ( 1 ) − ∑ i α 0 N , i ∘ Y i ( 0 ) . Both oracle coefficient vectors are G N \mathcal{G}_N G N -measurable. Because their supports are disjoint, exactly one potential outcome from each participant enters the oracle estimator. Define the oracle smoothing bias as B N ∘ = ∑ i = 1 N α 1 N , i ∘ μ 1 ( X i ) − ∑ i = 1 N α 0 N , i ∘ μ 0 ( X i ) − Δ N B_N^\circ=\sum_{i=1}^N\alpha_{1N,i}^\circ\mu_1(X_i)-\sum_{i=1}^N\alpha_{0N,i}^\circ\mu_0(X_i)-\Delta_N B N ∘ = ∑ i = 1 N α 1 N , i ∘ μ 1 ( X i ) − ∑ i = 1 N α 0 N , i ∘ μ 0 ( X i ) − Δ N .
(C3) Negligible integrated smoothing bias The oracle smoothing bias satisfies
B N ∘ = o p ( N − 1 / 2 ) . B_N^\circ=o_p(N^{-1/2}). B N ∘ = o p ( N − 1/2 ) .
This condition concerns the signed average prediction error over the enrolled trial covariates, which is distinct from the pointwise error of either kernel regression. Root-N N N theory for averages of kernel estimators commonly controls this type of integrated bias by undersmoothing Newey, [5 ] . A closely related precedent within kernel-based adaptive randomization is the second-order-kernel condition of Jiang et al. [3 ] , which combines a diverging local-information rate with a vanishing root-N N N bias rate. Lemma S1 below records exact-reproduction and undersmoothing regimes that imply (C3). These regimes provide sufficient conditions for the required integrated-bias control. In the implemented design, Scott’s rule is a prespecified finite-sample bandwidth choice rather than an asymptotic undersmoothing prescription. With fixed N H N_H N H , the implemented Scott bandwidth satisfies (C3) under the exact-reproduction regime described below. More generally, (C3) specifies the integrated bias along the trial sequence, whose verification depends on the response surface and bandwidth regime.
(C4) Allocation perturbation, terminal borrowing, and kernel regularity Write Λ N = ∑ i = 1 N ∣ ϕ i − 1 ( X i ) − ϕ i − 1 ∘ ( X i ) ∣ \Lambda_N=\sum_{i=1}^N\lvert\phi_{i-1}(X_i)-\phi_{i-1}^\circ(X_i)\rvert Λ N = ∑ i = 1 N ∣ ϕ i − 1 ( X i ) − ϕ i − 1 ∘ ( X i )∣ . For the final analysis, let W N , i F = W N F ( X i ) W_{N,i}^F=W_N^F(X_i) W N , i F = W N F ( X i ) and n ~ 0 , N , i e f f = max { n 0 , N e f f ( X i ) , n min } \widetilde n_{0,N,i}^{\mathrm{eff}}=\max\{n_{0,N}^{\mathrm{eff}}(X_i),n_{\min}\} n 0 , N , i eff = max { n 0 , N eff ( X i ) , n m i n } , where σ ^ 0 , N 2 ( X i ) \widehat{\sigma}_{0,N}^2(X_i) σ 0 , N 2 ( X i ) is the completed-cohort variance defined by setting n = N n=N n = N in the Stage II local summary. Then
S E N , i 2 = τ ^ H 2 ( X i ) + σ ^ 0 , N 2 ( X i ) n ~ 0 , N , i e f f , G N , i F = σ ^ 0 , N 2 ( X i ) n ~ 0 , N , i e f f S E N , i 2 . \mathrm{SE}_{N,i}^2=\widehat{\tau}_H^2(X_i)+\frac{\widehat{\sigma}_{0,N}^2(X_i)}{\widetilde n_{0,N,i}^{\mathrm{eff}}},
\qquad
G_{N,i}^F=\frac{\widehat{\sigma}_{0,N}^2(X_i)}{\widetilde n_{0,N,i}^{\mathrm{eff}}\mathrm{SE}_{N,i}^2}. SE N , i 2 = τ H 2 ( X i ) + n 0 , N , i eff σ 0 , N 2 ( X i ) , G N , i F = n 0 , N , i eff SE N , i 2 σ 0 , N 2 ( X i ) .
Assume S E N , i 2 > 0 \mathrm{SE}_{N,i}^2>0 SE N , i 2 > 0 for every evaluation profile and that the completed-cohort marginal discrepancy is well defined with probability tending to one. The local variance quantities entering denominators or square roots are floored at ϵ ν = 10 − 12 \epsilon_\nu=10^{-12} ϵ ν = 1 0 − 12 in Algorithms S1–S3 in § 5.7 ; well-definedness of the marginal discrepancy is imposed separately. Quantities may be assigned arbitrary finite values on the exceptional events where the marginal discrepancy is undefined. There are deterministic sequences d N ≥ 1 d_N \ge1 d N ≥ 1 and r N ↓ 0 r_N \downarrow0 r N ↓ 0 such that
Λ N = O p ( d N ) , 1 N ∑ i = 1 N G N , i F = O p ( r N ) , N − 1 / 2 + 1 / q d N → 0 , N 1 / 2 + 1 / q r N → 0. (10) \Lambda_N = O_p(d_N), \qquad\frac{1}{N}\sum_{i=1}^{N} G^F_{N,i} = O_p(r_N), \qquad N^{-1/2+1/q}d_N \to0, \qquad N^{1/2+1/q}r_N \to0. \tag*{(10)} Λ N = O p ( d N ) , N 1 i = 1 ∑ N G N , i F = O p ( r N ) , N − 1/2 + 1/ q d N → 0 , N 1/2 + 1/ q r N → 0. (10)
Let S z N C ( X i ) S^C_{zN}(X_i) S z N C ( X i ) denote the normalized concurrent-only kernel row formed from the actual assignment path, and retain S z N ∘ ( X i ) S^\circ_{zN}(X_i) S z N ∘ ( X i ) for the oracle path. The subscript N N N and superscripts C C C or ∘ \circ ∘ denote full-length rows on the common participant index; the unindexed S z S_z S z and S H S_H S H in Stage III are arm-compressed rows used for computation. Both paths use the same bandwidth specified for the N N N th trial; this bandwidth is H N \mathcal{H}_N H N -measurable and hence fixed before the auxiliary randomization is realized. All row denominators are positive, and the normalized kernel matrices have bounded column load:
max z ∈ { 0 , 1 } , 1 ≤ j ≤ N ∑ i = 1 N { S z N , j C ( X i ) + S z N , j ∘ ( X i ) } = O p ( 1 ) . (11) \max_{z\in\{0,1\},\,1\le j\le N}\sum_{i=1}^{N}\left\{S^C_{zN,j}(X_i)+S^\circ_{zN,j}(X_i)\right\}=O_p(1). \tag*{(11)} z ∈ { 0 , 1 } , 1 ≤ j ≤ N max i = 1 ∑ N { S z N , j C ( X i ) + S z N , j ∘ ( X i ) } = O p ( 1 ) . (11)
Entries outside the relevant arm are zero. Condition (S2) prevents one observed outcome from receiving unbounded aggregate weight across all evaluation profiles. It is an average leverage condition. Its relation to a uniform density lower bound over the entire covariate support depends on additional kernel and design assumptions. Together, these requirements state that the normalized kernel regressions are stable and that the cumulative difference between the actual and oracle assignment probabilities grows more slowly than the scale specified in (S1).
The existing KBCD theory establishes balance properties under smooth, stabilizing allocation maps for outcome-independent kernel biased-coin designs [3 ] , and classical biased-coin theory develops the same feedback principle [6 ] . For the outcome-adaptive RADISH rule, the first relation in (S1) summarizes the required design stability. Lemma S2 verifies the remaining rate requirements when this one-over-information coupling rate and the stated borrowing-rate bound hold.
Variance scale and no dominant oracle coefficient
Let V N ∘ = σ 1 2 ∥ α 1 N ∘ ∥ 2 2 + σ 0 2 ∥ α 0 N ∘ ∥ 2 2 V_N^\circ=\sigma_1^2\lVert\alpha^\circ_{1N}\rVert_2^2+\sigma_0^2\lVert\alpha^\circ_{0N}\rVert_2^2 V N ∘ = σ 1 2 ∥ α 1 N ∘ ∥ 2 2 + σ 0 2 ∥ α 0 N ∘ ∥ 2 2 . For some constant v ∈ ( 0 , ∞ ) v\in(0,\infty) v ∈ ( 0 , ∞ ) ,
N V N ∘ ⟶ p v , max z ∈ { 0 , 1 } , 1 ≤ i ≤ N ∣ α z N , i ∘ ∣ V N ∘ ⟶ p 0. NV_N^\circ\overset{p}{\longrightarrow} v, \qquad\max_{z\in\{0,1\},\,1\le i\le N}\frac{|\alpha^\circ_{zN,i}|}{\sqrt{V_N^\circ}}\overset{p}{\longrightarrow}0. N V N ∘ ⟶ p v , z ∈ { 0 , 1 } , 1 ≤ i ≤ N max V N ∘ ∣ α z N , i ∘ ∣ ⟶ p 0.
The second relation is a Lindeberg-type condition requiring that no single coefficient dominates the oracle linear smoother. It is motivated by classical kernel-regression regimes with diverging local effective sample sizes and covariate density bounded away from zero on the analysis region [2 , 8 ] .
(C6) Oracle residual-smoother regularity For z ∈ { 0 , 1 } z\in\{0,1\} z ∈ { 0 , 1 } , let A z N ∘ A^\circ_{zN} A z N ∘ be the N × N N\times N N × N diagonal matrix with i i i th diagonal entry 1 ( Z i = z ) \mathbb{1}(Z_i=z) 1 ( Z i = z ) , and let S z N s e l f , ∘ S^{\mathrm{self},\circ}_{zN} S z N self , ∘ be the full-length oracle self-smoothing matrix whose i i i th row equals S z N ∘ ( X i ) S^{\circ}_{zN}(X_i) S z N ∘ ( X i ) when Z i ∘ = z Z_i^{\circ}=z Z i ∘ = z and is zero otherwise. Let μ z N = { μ z ( X 1 ) , … , μ z ( X N ) } T \boldsymbol{\mu}_{zN}=\{\mu_z(X_1),\ldots,\mu_z(X_N)\}^{\mathsf T} μ z N = { μ z ( X 1 ) , … , μ z ( X N ) } T and N z ∘ = tr ( A z N ∘ ) N_z^{\circ}=\operatorname{tr}(A_{zN}^{\circ}) N z ∘ = tr ( A z N ∘ ) . For constants π z ∈ ( 0 , 1 ) \pi_z\in(0,1) π z ∈ ( 0 , 1 ) ,
N z ∘ N → p π z , 1 N ∥ S z N s e l f , ∘ ∥ F 2 → p 0 , 1 N ∥ ( A z N ∘ − S z N s e l f , ∘ ) μ z N ∥ 2 2 → p 0. (12) \frac{N_z^{\circ}}{N}\xrightarrow{p}\pi_z,\qquad\frac{1}{N}\lVert S_{zN}^{\mathrm{self},\circ}\rVert_F^2\xrightarrow{p}0,\qquad\frac{1}{N}\lVert(A_{zN}^{\circ}-S_{zN}^{\mathrm{self},\circ})\boldsymbol{\mu}_{zN}\rVert_2^2\xrightarrow{p}0. \tag*{(12)} N N z ∘ p π z , N 1 ∥ S z N self , ∘ ∥ F 2 p 0 , N 1 ∥( A z N ∘ − S z N self , ∘ ) μ z N ∥ 2 2 p 0. (12)
The historical residual degrees of freedom in Section S4.3 are positive. The second requirement in (S3) is a vanishing-average-leverage condition, and the third is mean-square consistency of the oracle mean fit on its assigned arm. Both follow under familiar kernel regression conditions such as diverging local effective sample sizes, vanishing maximum normalized weight, and sufficient smoothness with a shrinking bandwidth [1 , 4 ] . With a fixed nonzero bandwidth, the third relation holds, for example, when the smoother exactly reproduces each arm-specific conditional mean at the enrolled covariates. More generally, it requires the mean-square approximation error to vanish. The positive limiting arm fractions are the usual positivity requirement and arise when the oracle biased-coin design converges to an interior allocation target, as in the balance theory of Jiang et al. [3 ] . Condition (C6) is stated in terms of the smoothing operators. Lemma S6 uses this condition together with the RADISH-to-oracle stability in (C4) to prove consistency of the actual residual variance estimators. Under heteroskedastic outcomes, the scalar source-specific variance estimators in Equation (7) of the main text would instead need an observation-level robust replacement.
Lemma S1 (Sufficient regimes for (C3)). For z ∈ { 0 , 1 } z\in\{0,1\} z ∈ { 0 , 1 } , define the oracle fitted conditional mean as m z N ∘ ( X i ) = S z N ∘ ( X i ) T μ z N m_{zN}^{\circ}(X_i)=S_{zN}^{\circ}(X_i)^{\mathsf T}\boldsymbol{\mu}_{zN} m z N ∘ ( X i ) = S z N ∘ ( X i ) T μ z N . Condition (C3) holds in either of the following regimes:
the oracle smoother reproduces both conditional means at every enrolled covariate, so that m z N ∘ ( X i ) = μ z ( X i ) m_{zN}^{\circ}(X_i)=\mu_z(X_i) m z N ∘ ( X i ) = μ z ( X i ) for all i i i and z z z ;
for some smoothness order s > 0 s>0 s > 0 and bandwidth vector h N \mathbf{h}_N h N ,
1 N ∑ i = 1 N { m z N ∘ ( X i ) − μ z ( X i ) } = O p ( ∥ h N ∥ s ) , z ∈ { 0 , 1 } , \frac{1}{N}\sum_{i=1}^{N}\{m_{zN}^{\circ}(X_i)-\mu_z(X_i)\}=O_p(\lVert\mathbf{h}_N\rVert^s),\qquad z\in\{0,1\}, N 1 i = 1 ∑ N { m z N ∘ ( X i ) − μ z ( X i )} = O p (∥ h N ∥ s ) , z ∈ { 0 , 1 } , and N ∥ h N ∥ s → 0 \sqrt{N}\lVert\mathbf{h}_N\rVert^s\to0 N ∥ h N ∥ s → 0 .
The first regime includes arm-specific constant conditional means because every normalized kernel row reproduces constants. In an isotropic d d d -dimensional setting with h N = N − a \mathbf{h}_N=N^{-a} h N = N − a , the usual requirements N h N d → ∞ Nh_N^d\to\infty N h N d → ∞ and N h N s → 0 \sqrt{N}h_N^s\to0 N h N s → 0 can hold simultaneously whenever s > d / 2 s>d/2 s > d /2 , by choosing 1 / ( 2 s ) < a < 1 / d 1/(2s)<a<1/d 1/ ( 2 s ) < a < 1/ d . Taking s = 2 s=2 s = 2 , an integrated-bias expansion therefore permits such a choice for d < 4 d<4 d < 4 .
Proof. First express the smoothing bias in terms of fitted conditional means. For each z ∈ { 0 , 1 } z\in\{0,1\} z ∈ { 0 , 1 } , ( α z N ∘ ) T μ z N = N − 1 ∑ i S z N ∘ ( X i ) T μ z N = N − 1 ∑ i m z N ∘ ( X i ) (\boldsymbol{\alpha}_{zN}^{\circ})^{\mathsf T}\boldsymbol{\mu}_{zN}=N^{-1}\sum_i S_{zN}^{\circ}(X_i)^{\mathsf T}\boldsymbol{\mu}_{zN}=N^{-1}\sum_i m_{zN}^{\circ}(X_i) ( α z N ∘ ) T μ z N = N − 1 ∑ i S z N ∘ ( X i ) T μ z N = N − 1 ∑ i m z N ∘ ( X i ) . Substituting this equality and the definition of Δ N \Delta_N Δ N into B N ∘ B_N^{\circ} B N ∘ gives B N ∘ = N − 1 ∑ i [ { m 1 N ∘ ( X i ) − μ 1 ( X i ) } − { m 0 N ∘ ( X i ) − μ 0 ( X i ) } ] B_N^{\circ}=N^{-1}\sum_i[\{m_{1N}^{\circ}(X_i)-\mu_1(X_i)\}-\{m_{0N}^{\circ}(X_i)-\mu_0(X_i)\}] B N ∘ = N − 1 ∑ i [{ m 1 N ∘ ( X i ) − μ 1 ( X i )} − { m 0 N ∘ ( X i ) − μ 0 ( X i )}] . In the exact-reproduction regime, both bracketed errors vanish for every i i i , so B N ∘ = 0 B_N^{\circ}=0 B N ∘ = 0 . In the second regime, each signed average is O p ( ∥ h N ∥ s ) O_p(\lVert\mathbf{h}_N\rVert^s) O p (∥ h N ∥ s ) . Their difference has the same order by the triangle inequality, and N B N ∘ = O p ( N ∥ h N ∥ s ) = o p ( 1 ) \sqrt{N}B_N^{\circ}=O_p(\sqrt{N}\lVert\mathbf{h}_N\rVert^s)=o_p(1) N B N ∘ = O p ( N ∥ h N ∥ s ) = o p ( 1 ) . This is equivalent to B N ∘ = o p ( N − 1 / 2 ) B_N^{\circ}=o_p(N^{-1/2}) B N ∘ = o p ( N − 1/2 ) . The two bandwidth requirements are compatible. If h N = N − a h_N = N^{-a} h N = N − a , then N h N d = N 1 − a d → ∞ Nh_N^d = N^{1-ad} \to\infty N h N d = N 1 − a d → ∞ exactly when a < 1 / d a < 1/d a < 1/ d , whereas N h N s = N 1 / 2 − a s → 0 \sqrt{N}h_N^s = N^{1/2-as} \to0 N h N s = N 1/2 − a s → 0 exactly when a > 1 / ( 2 s ) a > 1/(2s) a > 1/ ( 2 s ) . A value of a a a in this interval exists if and only if 1 / ( 2 s ) < 1 / d 1/(2s) < 1/d 1/ ( 2 s ) < 1/ d , or equivalently s > d / 2 s > d/2 s > d /2 .
Lemma S2 (A sufficient one-over-information regime for (C4)).
Suppose, in addition to the kernel-row requirement (S2), that
sup i 0 ≤ i ≤ N i ∣ ϕ i − 1 ( X i ) − ϕ i − 1 ∞ ( X i ) ∣ = O p ( 1 ) , max 1 ≤ i ≤ N N G N , i F = O p ( 1 ) , \sup_{i_0\leq i\leq N} i\left|\phi_{i-1}(X_i)-\phi^\infty_{i-1}(X_i)\right|=O_p(1),\qquad\max_{1\leq i\leq N}NG^F_{N,i}=O_p(1), i 0 ≤ i ≤ N sup i ϕ i − 1 ( X i ) − ϕ i − 1 ∞ ( X i ) = O p ( 1 ) , 1 ≤ i ≤ N max N G N , i F = O p ( 1 ) , where i 0 i_0 i 0 is the fixed end of burn-in. Then the rate requirements in (S1) hold with d N = log ( e N ) d_N=\log(eN) d N = log ( e N ) and r N = N − 1 r_N=N^{-1} r N = N − 1 for every q > 2 q>2 q > 2 .
Proof. Define A N = sup i 0 ≤ i ≤ N i ∣ ϕ i − 1 ( X i ) − ϕ i − 1 ∞ ( X i ) ∣ A_N=\sup_{i_0\leq i\leq N}i\left|\phi_{i-1}(X_i)-\phi^\infty_{i-1}(X_i)\right| A N = sup i 0 ≤ i ≤ N i ϕ i − 1 ( X i ) − ϕ i − 1 ∞ ( X i ) . The two designs share the burn-in, so their allocation probabilities agree before i 0 i_0 i 0 . For i ≥ i 0 i\geq i_0 i ≥ i 0 , the definition of A N A_N A N gives a pointwise bound by A N / i A_N/i A N / i . Hence Λ N ≤ A N ∑ i = i 0 N i − 1 ≤ A N { 1 + log N } = O p { log ( e N ) } \Lambda_N\leq A_N\sum_{i=i_0}^N i^{-1}\leq A_N\{1+\log N\}=O_p\{\log(eN)\} Λ N ≤ A N ∑ i = i 0 N i − 1 ≤ A N { 1 + log N } = O p { log ( e N )} . For the borrowing term, the second premise of the lemma gives max i N G N , i F = O p ( 1 ) \max_i NG^F_{N,i}=O_p(1) max i N G N , i F = O p ( 1 ) . The mean borrowing weight is bounded by its maximum, and hence N − 1 ∑ i G N , i F = O p ( N − 1 ) N^{-1}\sum_iG^F_{N,i}=O_p(N^{-1}) N − 1 ∑ i G N , i F = O p ( N − 1 ) . Thus d N = log ( e N ) d_N=\log(eN) d N = log ( e N ) and r N = N − 1 r_N=N^{-1} r N = N − 1 provide the two stochastic bounds in (S1). Since q > 2 q>2 q > 2 , the exponent − 1 / 2 + 1 / q -1/2+1/q − 1/2 + 1/ q is negative. Therefore N − 1 / 2 + 1 / q log ( e N ) → 0 N^{-1/2+1/q}\log(eN)\to0 N − 1/2 + 1/ q log ( e N ) → 0 and N 1 / 2 + 1 / q N − 1 = N − 1 / 2 + 1 / q → 0 N^{1/2+1/q}N^{-1}=N^{-1/2+1/q}\to0 N 1/2 + 1/ q N − 1 = N − 1/2 + 1/ q → 0 . This verifies the remaining deterministic rate restrictions.
The preceding regime has a direct design interpretation. If local concurrent information in both arms grows proportionally to enrollment, the fixed historical reference changes the effective local control information by order one, whereas the allocation totals are of order i i i . At fixed local totals, a smooth allocation rule then changes the allocation probability by order i − 1 i^{-1} i − 1 . The preceding lemma assumes this uniform rate for the full coupled outcome-adaptive recursion, while condition (C4) allows the more general cumulative perturbation bound. Similarly, local control information of order N N N , a bounded concurrent variance estimate, and a historical reference variance bounded away from zero give G N , i F = O p ( N − 1 ) G^F_{N,i}=O_p(N^{-1}) G N , i F = O p ( N − 1 ) .
Oracle Equivalence
For z ∈ { 0 , 1 } z\in\{0,1\} z ∈ { 0 , 1 } , let α z N C \alpha^C_{zN} α z N C be the full-length concurrent-only coefficient vector obtained by averaging S z N C ( X i ) S^C_{zN}(X_i) S z N C ( X i ) over i = 1 , … , N i=1,\ldots,N i = 1 , … , N , and let Δ ^ N C = ∑ i α ^ 1 N , i C Y i ( 1 ) − ∑ i α ^ 0 N , i C Y i ( 0 ) \widehat{\Delta}^C_N=\sum_i\widehat{\alpha}^C_{1N,i}Y_i(1)-\sum_i\widehat{\alpha}^C_{0N,i}Y_i(0) Δ N C = ∑ i α 1 N , i C Y i ( 1 ) − ∑ i α 0 N , i C Y i ( 0 ) . This intermediate estimator uses the actual RADISH assignment path but sets the terminal borrowing weights to zero. It separates the effect of terminal borrowing from the effect of the outcome-dependent allocation path.
Lemma S3 (Normalization of two kernel rows). Let k 1 , … , k N k_1,\ldots,k_N k 1 , … , k N be nonnegative, let A A A and B B B be two nonempty index sets, and define p A , j = k j 1 ( j ∈ A ) / s A p_{A,j}=k_j1(j\in A)/s_A p A , j = k j 1 ( j ∈ A ) / s A and p B , j = k j 1 ( j ∈ B ) / s B p_{B,j}=k_j1(j\in B)/s_B p B , j = k j 1 ( j ∈ B ) / s B , where s A = ∑ j ∈ A k j > 0 s_A=\sum_{j\in A}k_j>0 s A = ∑ j ∈ A k j > 0 and s B = ∑ j ∈ B k j > 0 s_B=\sum_{j\in B}k_j>0 s B = ∑ j ∈ B k j > 0 . Then
∥ p A − p B ∥ 1 ≤ 2 ( ∑ j ∈ A ∖ B p A , j + ∑ j ∈ B ∖ A p B , j ) . \lVert p_A-p_B\rVert_1\leq2\left(\sum_{j\in A\setminus B}p_{A,j}+\sum_{j\in B\setminus A}p_{B,j}\right). ∥ p A − p B ∥ 1 ≤ 2 j ∈ A ∖ B ∑ p A , j + j ∈ B ∖ A ∑ p B , j .
Proof. Let u A = ∑ j ∈ A ∖ B p A , j u_A=\sum_{j\in A\setminus B}p_{A,j} u A = ∑ j ∈ A ∖ B p A , j and u B = ∑ j ∈ B ∖ A p B , j u_B=\sum_{j\in B\setminus A}p_{B,j} u B = ∑ j ∈ B ∖ A p B , j . The indices outside A ∩ B A\cap B A ∩ B contribute exactly u A + u B u_A+u_B u A + u B to the ℓ 1 \ell_1 ℓ 1 distance. On A ∩ B A\cap B A ∩ B , p A , j − p B , j = k j ( s A − 1 − s B − 1 ) p_{A,j}-p_{B,j}=k_j(s_A^{-1}-s_B^{-1}) p A , j − p B , j = k j ( s A − 1 − s B − 1 ) , so all nonzero differences have the same sign. Their absolute values can therefore be summed before taking the absolute value, yielding ∑ j ∈ A ∩ B ∣ p A , j − p B , j ∣ = ∣ ∑ j ∈ A ∩ B p A , j − ∑ j ∈ A ∩ B p B , j ∣ = ∣ u A − u B ∣ \sum_{j\in A\cap B}|p_{A,j}-p_{B,j}|=\left|\sum_{j\in A\cap B}p_{A,j}-\sum_{j\in A\cap B}p_{B,j}\right|=|u_A-u_B| ∑ j ∈ A ∩ B ∣ p A , j − p B , j ∣ = ∑ j ∈ A ∩ B p A , j − ∑ j ∈ A ∩ B p B , j = ∣ u A − u B ∣ . Consequently, ∥ p A − p B ∥ 1 = u A + u B + ∣ u A − u B ∣ ≤ 2 ( u A + u B ) \lVert p_A-p_B\rVert_1=u_A+u_B+|u_A-u_B|\leq2(u_A+u_B) ∥ p A − p B ∥ 1 = u A + u B + ∣ u A − u B ∣ ≤ 2 ( u A + u B ) , as claimed.
Lemma S4 (First-order equivalence of RADISH and its oracle). Under (C1), (C2), and (C4),
Δ ^ N − Δ ^ N ∘ = o p ( N − 1 / 2 ) , (13) \widehat{\Delta}_N-\widehat{\Delta}^{\circ}_N=o_p(N^{-1/2}), \tag*{(13)} Δ N − Δ N ∘ = o p ( N − 1/2 ) , (13) and, on defining α H N ∘ \alpha^{\circ}_{HN} α H N ∘ as the zero vector in R N H \mathbb{R}^{N_H} R N H ,
∑ z ∈ { 1 , 0 , H } ∥ α ^ z N − α z N ∘ ∥ 2 = o p ( N − 1 / 2 ) . (14) \sum_{z\in\{1,0,H\}}\lVert\widehat{\alpha}_{zN}-\alpha^{\circ}_{zN}\rVert_2=o_p(N^{-1/2}). \tag*{(14)} z ∈ { 1 , 0 , H } ∑ ∥ α z N − α z N ∘ ∥ 2 = o p ( N − 1/2 ) . (14)
Proof. To compare the allocation paths, let I i = 1 ( Z i ≠ Z i ∘ ) I_i=\mathbf{1}(Z_i\ne Z_i^\circ) I i = 1 ( Z i = Z i ∘ ) and J N = ∑ i = 1 N I i J_N=\sum_{i=1}^N I_i J N = ∑ i = 1 N I i . Let H i − 1 X \mathcal{H}^{X}_{i-1} H i − 1 X be the joint information generated by the historical data, the covariates observed through participant i i i , the preceding actual outcomes and assignments, and the preceding oracle assignments. Because the two paths use the same U i U_i U i , E ( I i ∣ H i − 1 X ) = ∣ ϕ i − 1 ( X i ) − ϕ i − 1 ∘ ( X i ) ∣ \mathbb{E}(I_i\mid\mathcal{H}^{X}_{i-1})=|\phi_{i-1}(X_i)-\phi^\circ_{i-1}(X_i)| E ( I i ∣ H i − 1 X ) = ∣ ϕ i − 1 ( X i ) − ϕ i − 1 ∘ ( X i ) ∣ . Let p N , i = ∣ ϕ i − 1 ( X i ) − ϕ i − 1 ∘ ( X i ) ∣ p_{N,i}=|\phi_{i-1}(X_i)-\phi^\circ_{i-1}(X_i)| p N , i = ∣ ϕ i − 1 ( X i ) − ϕ i − 1 ∘ ( X i ) ∣ and M N , k = ∑ i = 1 k ( I i − p N , i ) M_{N,k}=\sum_{i=1}^k(I_i-p_{N,i}) M N , k = ∑ i = 1 k ( I i − p N , i ) . Its increment is Δ M N , i = I i − p N , i \Delta M_{N,i}=I_i-p_{N,i} Δ M N , i = I i − p N , i . Because I i I_i I i is conditionally Bernoulli with success probability p N , i p_{N,i} p N , i , E ( Δ M N , i ∣ H i − 1 X ) = 0 \mathbb{E}(\Delta M_{N,i}\mid\mathcal{H}^{X}_{i-1})=0 E ( Δ M N , i ∣ H i − 1 X ) = 0 and E { ( Δ M N , i ) 2 ∣ H i − 1 X } = p N , i ( 1 − p N , i ) \mathbb{E}\{(\Delta M_{N,i})^2\mid\mathcal{H}^{X}_{i-1}\}=p_{N,i}(1-p_{N,i}) E {( Δ M N , i ) 2 ∣ H i − 1 X } = p N , i ( 1 − p N , i ) . Thus { M N , k : 0 ≤ k ≤ N } \{M_{N,k}:0\leq k\leq N\} { M N , k : 0 ≤ k ≤ N } is a square-integrable martingale. Its predictable quadratic variation is, by definition,
⟨ M N ⟩ k : = ∑ i = 1 k E { ( Δ M N , i ) 2 ∣ H i − 1 X } = ∑ i = 1 k p N , i ( 1 − p N , i ) ≤ ∑ i = 1 k p N , i . \langle M_N\rangle_k:=\sum_{i=1}^k\mathbb{E}\{(\Delta M_{N,i})^2\mid\mathcal{H}^{X}_{i-1}\}=\sum_{i=1}^k p_{N,i}(1-p_{N,i})\leq\sum_{i=1}^k p_{N,i}. ⟨ M N ⟩ k := i = 1 ∑ k E {( Δ M N , i ) 2 ∣ H i − 1 X } = i = 1 ∑ k p N , i ( 1 − p N , i ) ≤ i = 1 ∑ k p N , i .
Fix c s t o p > 0 c_{\mathrm{stop}}>0 c stop > 0 , and let τ N \tau_N τ N be the first k ≤ N k\leq N k ≤ N for which ∑ i = 1 k p N , i > c s t o p d N \sum_{i=1}^k p_{N,i}>c_{\mathrm{stop}}d_N ∑ i = 1 k p N , i > c stop d N , with τ N = N \tau_N=N τ N = N when no crossing occurs. Immediately before a crossing, the cumulative sum is at most c s t o p d N c_{\mathrm{stop}}d_N c stop d N , and the crossing increment is at most one. Consequently, ⟨ M N ⟩ N ∧ τ N ≤ c s t o p d N + 1 \langle M_N\rangle_{N\wedge\tau_N}\leq c_{\mathrm{stop}}d_N+1 ⟨ M N ⟩ N ∧ τ N ≤ c stop d N + 1 . Because p N , k p_{N,k} p N , k is measurable before the k k k th randomization, τ N \tau_N τ N is a bounded stopping time for the associated martingale filtration. Let E N ( c s t o p ) = { Λ N ≤ c s t o p d N } \mathcal{E}_N(c_{\mathrm{stop}})=\{\Lambda_N\leq c_{\mathrm{stop}}d_N\} E N ( c stop ) = { Λ N ≤ c stop d N } . No crossing occurs on E N ( c s t o p ) \mathcal{E}_N(c_{\mathrm{stop}}) E N ( c stop ) , so M N , N = M N , N ∧ τ N M_{N,N}=M_{N,N\wedge\tau_N} M N , N = M N , N ∧ τ N there. For every L > 0 L>0 L > 0 , split the probability according to E N ( c s t o p ) \mathcal{E}_N(c_{\mathrm{stop}}) E N ( c stop ) :
Pr { ∣ M N , N ∣ > L d N } ≤ Pr { E N ( c s t o p ) c } + Pr { ∣ M N , N ∣ > L d N , E N ( c s t o p ) } ≤ Pr ( Λ N > c s t o p d N ) + Pr { ∣ M N , N ∧ τ N ∣ > L d N } ≤ Pr ( Λ N > c s t o p d N ) + E ( M N , N ∧ τ N 2 ) L 2 d N 2 = Pr ( Λ N > c s t o p d N ) + E { ⟨ M N ⟩ N ∧ τ N } L 2 d N 2 ≤ Pr ( Λ N > c s t o p d N ) + c s t o p d N + 1 L 2 d N 2 ≤ Pr ( Λ N > c s t o p d N ) + c s t o p + 1 L 2 d N . \begin{aligned}
\Pr\{|M_{N,N}|>Ld_N\}
&\leq\Pr\{\mathcal{E}_N(c_{\mathrm{stop}})^c\}+\Pr\{|M_{N,N}|>Ld_N,\mathcal{E}_N(c_{\mathrm{stop}})\}\\
&\leq\Pr(\Lambda_N>c_{\mathrm{stop}}d_N)+\Pr\{|M_{N,N\wedge\tau_N}|>Ld_N\}\\
&\leq\Pr(\Lambda_N>c_{\mathrm{stop}}d_N)+\frac{\mathbb{E}(M_{N,N\wedge\tau_N}^2)}{L^2d_N^2}\\
&=\Pr(\Lambda_N>c_{\mathrm{stop}}d_N)+\frac{\mathbb{E}\{\langle M_N\rangle_{N\wedge\tau_N}\}}{L^2d_N^2}\\
&\leq\Pr(\Lambda_N>c_{\mathrm{stop}}d_N)+\frac{c_{\mathrm{stop}}d_N+1}{L^2d_N^2}\\
&\leq\Pr(\Lambda_N>c_{\mathrm{stop}}d_N)+\frac{c_{\mathrm{stop}}+1}{L^2d_N}.
\end{aligned} Pr { ∣ M N , N ∣ > L d N } ≤ Pr { E N ( c stop ) c } + Pr { ∣ M N , N ∣ > L d N , E N ( c stop )} ≤ Pr ( Λ N > c stop d N ) + Pr { ∣ M N , N ∧ τ N ∣ > L d N } ≤ Pr ( Λ N > c stop d N ) + L 2 d N 2 E ( M N , N ∧ τ N 2 ) = Pr ( Λ N > c stop d N ) + L 2 d N 2 E {⟨ M N ⟩ N ∧ τ N } ≤ Pr ( Λ N > c stop d N ) + L 2 d N 2 c stop d N + 1 ≤ Pr ( Λ N > c stop d N ) + L 2 d N c stop + 1 .
The third line is Chebyshev’s inequality. The equality on the fourth line is the martingale isometry for the stopped square-integrable martingale. For any ϵ > 0 \epsilon> 0 ϵ > 0 , Λ N = O p ( d N ) \Lambda_N = O_p(d_N) Λ N = O p ( d N ) permits a fixed c s t o p c_{\mathrm{stop}} c stop for which the first term is below ϵ / 2 \epsilon/2 ϵ /2 for all sufficiently large N N N . With this c s t o p c_{\mathrm{stop}} c stop fixed, choose L L L so that ( c s t o p + 1 ) / L 2 < ϵ / 2 (c_{\mathrm{stop}}+1)/L^2 < \epsilon/2 ( c stop + 1 ) / L 2 < ϵ /2 ; the second term is then below ϵ / 2 \epsilon/2 ϵ /2 because d N ≥ 1 d_N \ge1 d N ≥ 1 . This proves M N , N / d N = O p ( 1 ) M_{N,N}/d_N = O_p(1) M N , N / d N = O p ( 1 ) . Finally, J N = ∑ i = 1 N I i = ∑ i = 1 N p N , i + M N , N = Λ N + M N , N J_N = \sum_{i=1}^{N} I_i = \sum_{i=1}^{N} p_{N,i} + M_{N,N} = \Lambda_N + M_{N,N} J N = ∑ i = 1 N I i = ∑ i = 1 N p N , i + M N , N = Λ N + M N , N , so J N = O p ( d N ) J_N = O_p(d_N) J N = O p ( d N ) . Hence
J N = O p ( d N ) . (15) J_N = O_p(d_N). \tag*{(15)} J N = O p ( d N ) . (15)
The mismatch bound also controls the difference between the averaged kernel coefficients. Let J N = { j : Z j ≠ Z j ∘ } \mathcal{J}_N = \{j : Z_j \ne Z_j^\circ\} J N = { j : Z j = Z j ∘ } , so ∣ J N ∣ = J N |\mathcal{J}_N| = J_N ∣ J N ∣ = J N . For a fixed evaluation participant i i i and arm z z z , the actual and oracle kernel rows use the same nonnegative kernel values. Their supports can differ only at indices in J N \mathcal{J}_N J N . Lemma S3 therefore gives
∥ S z N C ( X i ) − S z N ∘ ( X i ) ∥ 1 ≤ 2 ∑ j ∈ J N { S z N , j C ( X i ) + S z N , j ∘ ( X i ) } . \left\|S^C_{zN}(X_i)-S^\circ_{zN}(X_i)\right\|_1
\le2\sum_{j\in\mathcal{J}_N}\left\{S^C_{zN,j}(X_i)+S^\circ_{zN,j}(X_i)\right\}. S z N C ( X i ) − S z N ∘ ( X i ) 1 ≤ 2 j ∈ J N ∑ { S z N , j C ( X i ) + S z N , j ∘ ( X i ) } .
The coefficient vectors are averages of these rows. The triangle inequality, followed by a change in the order of summation, now gives
∥ α ^ z N C − α ^ z N ∘ ∥ 1 ≤ 1 N ∑ i = 1 N ∥ S z N C ( X i ) − S z N ∘ ( X i ) ∥ 1 ≤ 2 N ∑ j : Z j ≠ Z j ∘ ∑ i = 1 N { S z N , j C ( X i ) + S z N , j ∘ ( X i ) } = O p ( J N / N ) = O p ( d N / N ) , z ∈ { 0 , 1 } . (16) \begin{aligned}
\left\|\widehat{\alpha}^C_{zN}-\widehat{\alpha}^\circ_{zN}\right\|_1
\le\frac{1}{N}\sum_{i=1}^{N}\left\|S^C_{zN}(X_i)-S^\circ_{zN}(X_i)\right\|_1 \\
&\le\frac{2}{N}\sum_{j:Z_j\ne Z_j^\circ}\sum_{i=1}^{N}\left\{S^C_{zN,j}(X_i)+S^\circ_{zN,j}(X_i)\right\} \\
&= O_p(J_N/N)=O_p(d_N/N),\qquad z\in\{0,1\}. \tag*{(16)}
\end{aligned} α z N C − α z N ∘ 1 ≤ N 1 i = 1 ∑ N S z N C ( X i ) − S z N ∘ ( X i ) 1 ≤ N 2 j : Z j = Z j ∘ ∑ i = 1 ∑ N { S z N , j C ( X i ) + S z N , j ∘ ( X i ) } = O p ( J N / N ) = O p ( d N / N ) , z ∈ { 0 , 1 } . (16)
The penultimate equality uses the maximum column-load bound in (S2), and the final equality uses (S6). The same bound holds for the Euclidean norm because ∥ v ∥ 2 ≤ ∥ v ∥ 1 \|v\|_2\le\|v\|_1 ∥ v ∥ 2 ≤ ∥ v ∥ 1 .
To transfer the coefficient bound to the estimators, we control the largest concurrent outcome. For any c t a i l > 0 c_{\mathrm{tail}}>0 c tail > 0 , the union bound and the uniform q q q th-moment bound in (C1) give
Pr { max i , z ∣ Y i ( z ) ∣ > c t a i l N 1 / q } ≤ c t a i l − q N − 1 ∑ i = 1 N ∑ z = 0 1 E ∣ Y i ( z ) ∣ q ≤ 2 C c t a i l − q , \Pr\left\{\max_{i,z}|Y_i(z)|>c_{\mathrm{tail}}N^{1/q}\right\}
\le c_{\mathrm{tail}}^{-q}N^{-1}\sum_{i=1}^{N}\sum_{z=0}^{1}\mathbb{E}|Y_i(z)|^q
\le2C c_{\mathrm{tail}}^{-q}, Pr { i , z max ∣ Y i ( z ) ∣ > c tail N 1/ q } ≤ c tail − q N − 1 i = 1 ∑ N z = 0 ∑ 1 E ∣ Y i ( z ) ∣ q ≤ 2 C c tail − q ,
where the first inequality is Markov’s inequality applied after the union bound, and the second follows by taking expectations in the conditional moment bound. Letting c t a i l c_{\mathrm{tail}} c tail increase proves
max i , z ∣ Y i ( z ) ∣ = O p ( N 1 / q ) . (17) \max_{i,z}|Y_i(z)|=O_p(N^{1/q}). \tag*{(17)} i , z max ∣ Y i ( z ) ∣ = O p ( N 1/ q ) . (17)
Combining (S7) and (S8) yields ∣ Δ ^ N C − Δ ^ N ∘ ∣ = O p ( d N N − 1 + 1 / q ) = o p ( N − 1 / 2 ) |\widehat{\Delta}^C_N-\widehat{\Delta}^\circ_N|=O_p(d_NN^{-1+1/q})=o_p(N^{-1/2}) ∣ Δ N C − Δ N ∘ ∣ = O p ( d N N − 1 + 1/ q ) = o p ( N − 1/2 ) by the rate imposed on d N d_N d N in (S1).
It remains to bound the contribution of terminal historical borrowing. Since D N F ( X i ) ≥ 0 D^F_N(\mathbf{X}_i) \ge0 D N F ( X i ) ≥ 0 , the Stage III definition and W N , i F = ( R N , i F − 1 ) / R N , i F W^F_{N,i} = (R^F_{N,i}-1)/R^F_{N,i} W N , i F = ( R N , i F − 1 ) / R N , i F give
0 ≤ W N , i F ≤ R N , i F − 1 ≤ exp { − D N F ( X i ) } σ ^ 0 , N 2 ( X i ) n 0 , N , i e f f S E 0 , N , i 2 ≤ σ ^ 0 , N 2 ( X i ) n 0 , N , i e f f S E 0 , N , i 2 = G N , i F . 0 \le W^F_{N,i} \le R^F_{N,i}-1 \le\frac{\exp\{-D^F_N(\mathbf{X}_i)\}\widehat{\sigma}^2_{0,N}(\mathbf{X}_i)}{n^{\mathrm{eff}}_{0,N,i}\mathrm{SE}^2_{0,N,i}} \le\frac{\widehat{\sigma}^2_{0,N}(\mathbf{X}_i)}{n^{\mathrm{eff}}_{0,N,i}\mathrm{SE}^2_{0,N,i}} = G^F_{N,i}. 0 ≤ W N , i F ≤ R N , i F − 1 ≤ n 0 , N , i eff SE 0 , N , i 2 exp { − D N F ( X i )} σ 0 , N 2 ( X i ) ≤ n 0 , N , i eff SE 0 , N , i 2 σ 0 , N 2 ( X i ) = G N , i F .
Therefore W ‾ N F : = N − 1 ∑ i = 1 N W N , i F = O p ( r N ) \overline{W}^F_N := N^{-1}\sum_{i=1}^N W^F_{N,i} = O_p(r_N) W N F := N − 1 ∑ i = 1 N W N , i F = O p ( r N ) . The historical fit θ ( x ) \theta(\mathbf{x}) θ ( x ) and concurrent-control fit Y ‾ 0 , N ( x ) \overline{Y}_{0,N}(\mathbf{x}) Y 0 , N ( x ) are the kernel fits used at final analysis. The historical sample is fixed after conditioning on D H D_H D H . Both fits are convex combinations, and hence
max 1 ≤ i ≤ N ∣ θ ( X i ) − Y ‾ 0 , N ( X i ) ∣ ≤ max 1 ≤ j ≤ N H ∣ Y j H ∣ + max i , z ∣ Y i ( z ) ∣ = O p ( N 1 / q ) . \max_{1\le i\le N} \left|\theta(\mathbf{X}_i)-\overline{Y}_{0,N}(\mathbf{X}_i)\right| \le\max_{1\le j\le N_H}|Y^H_j|+\max_{i,z}|Y_i(z)|=O_p(N^{1/q}). 1 ≤ i ≤ N max θ ( X i ) − Y 0 , N ( X i ) ≤ 1 ≤ j ≤ N H max ∣ Y j H ∣ + i , z max ∣ Y i ( z ) ∣ = O p ( N 1/ q ) .
The exact identity
Δ ^ N − Δ ^ N C = − 1 N ∑ i = 1 N W N , i F { θ ( X i ) − Y ‾ 0 , N ( X i ) } \widehat{\Delta}_N-\widehat{\Delta}^C_N=-\frac{1}{N}\sum_{i=1}^N W^F_{N,i}\{\theta(\mathbf{X}_i)-\overline{Y}_{0,N}(\mathbf{X}_i)\} Δ N − Δ N C = − N 1 i = 1 ∑ N W N , i F { θ ( X i ) − Y 0 , N ( X i )}
and (S8) consequently give ∣ Δ ^ N − Δ ^ N C ∣ = O p ( r N N 1 / q ) = o p ( N − 1 / 2 ) |\widehat{\Delta}_N-\widehat{\Delta}^C_N|=O_p(r_NN^{1/q})=o_p(N^{-1/2}) ∣ Δ N − Δ N C ∣ = O p ( r N N 1/ q ) = o p ( N − 1/2 ) . Nonnegativity and unit row sums also give
∥ α ^ H N ∥ 1 = W ‾ N F , ∥ α ^ 0 N − α ^ 0 N C ∥ 1 = W ‾ N F , α ^ 1 N = α ^ 1 N C . \|\widehat{\boldsymbol{\alpha}}_{HN}\|_1=\overline{W}^F_N,\qquad\|\widehat{\boldsymbol{\alpha}}_{0N}-\widehat{\boldsymbol{\alpha}}^C_{0N}\|_1=\overline{W}^F_N,\qquad\widehat{\boldsymbol{\alpha}}_{1N}=\widehat{\boldsymbol{\alpha}}^C_{1N}. ∥ α H N ∥ 1 = W N F , ∥ α 0 N − α 0 N C ∥ 1 = W N F , α 1 N = α 1 N C .
The rate for r N r_N r N in (S1) implies r N = o ( N − 1 / 2 ) r_N=o(N^{-1/2}) r N = o ( N − 1/2 ) , so each displayed ℓ 1 \ell_1 ℓ 1 norm, and therefore each corresponding ℓ 2 \ell_2 ℓ 2 norm, is o p ( N − 1 / 2 ) o_p(N^{-1/2}) o p ( N − 1/2 ) . Combining the allocation and terminal-borrowing bounds by the triangle inequality proves (S4). Combining (S7) with the terminal coefficient bounds proves (S5), because
d N / N N − 1 / 2 = d N N − 1 / 2 = { d N N − 1 / 2 + 1 / q } N − 1 / q → 0 , r N = o ( N − 1 / 2 ) . \frac{d_N/N}{N^{-1/2}}=d_NN^{-1/2}=\{d_NN^{-1/2+1/q}\}N^{-1/q}\to0,\qquad r_N=o(N^{-1/2}). N − 1/2 d N / N = d N N − 1/2 = { d N N − 1/2 + 1/ q } N − 1/ q → 0 , r N = o ( N − 1/2 ) .
Proof of Proposition S1. For the assignment limit, (S6) gives J N = ∑ i = 1 N 1 ( Z i ≠ Z i o ) = O p ( d N ) J_N=\sum_{i=1}^N1(Z_i\ne Z_i^o)=O_p(d_N) J N = ∑ i = 1 N 1 ( Z i = Z i o ) = O p ( d N ) . Moreover, d N / N = { d N N − 1 / 2 + 1 / q } N − 1 / 2 − 1 / q → 0 d_N/N=\{d_NN^{-1/2+1/q}\}N^{-1/2-1/q}\to0 d N / N = { d N N − 1/2 + 1/ q } N − 1/2 − 1/ q → 0 by (S1). Hence J N / N = O p ( d N / N ) → p 0 J_N/N=O_p(d_N/N)\overset{p}{\to}0 J N / N = O p ( d N / N ) → p 0 . For the borrowing limit, the terminal-weight bound in the proof of Lemma S4 gives
N − 1 ∑ i = 1 N W N F ( X i ) = W ‾ N F = O p ( r N ) . N^{-1}\sum_{i=1}^N W^F_N(\mathbf{X}_i)=\overline{W}^F_N=O_p(r_N). N − 1 i = 1 ∑ N W N F ( X i ) = W N F = O p ( r N ) .
Because r N → 0 r_N\to0 r N → 0 , this average converges in probability to zero. Finally, (S4) is exactly the asserted first-order estimator equivalence.
Oracle Central Limit Theorem
Lemma S5 (Conditional weighted central limit theorem). Under (C1), (C2), and (C5),
Δ ^ N ∘ − Δ N − B N ∘ V N ∘ → d N ( 0 , 1 ) . \frac{\widehat{\Delta}^{\circ}_N-\Delta_N-B^{\circ}_N}{\sqrt{V^{\circ}_N}}\xrightarrow{d}\mathcal{N}(0,1). V N ∘ Δ N ∘ − Δ N − B N ∘ d N ( 0 , 1 ) .
Proof. Conditioning on the auxiliary randomization variables yields a sum of independent errors. For participant i i i , define ψ N , i = α 1 N , i ∘ ε i ( 1 ) − α 0 N , i ∘ ε i ( 0 ) \psi_{N,i} = \alpha^\circ_{1N,i}\varepsilon_i(1) - \alpha^\circ_{0N,i}\varepsilon_i(0) ψ N , i = α 1 N , i ∘ ε i ( 1 ) − α 0 N , i ∘ ε i ( 0 ) . The two oracle coefficient vectors have disjoint supports, so at most one term in ψ N , i \psi_{N,i} ψ N , i is nonzero. Conditioning on G N \mathcal{G}_N G N fixes the enrolled covariates, the oracle assignments, and the oracle coefficients, while leaving the potential-outcome errors unobserved because the auxiliary uniforms are independent of those errors under (C2). The cross-participant independence and conditional mean-zero assumptions in (C1) therefore give E ( ψ N , i ∣ G N ) = 0 \mathbb{E}(\psi_{N,i} \mid\mathcal{G}_N) = 0 E ( ψ N , i ∣ G N ) = 0 , with ψ N , 1 , … , ψ N , N \psi_{N,1},\ldots,\psi_{N,N} ψ N , 1 , … , ψ N , N conditionally independent. Sourcewise homoskedasticity and disjoint support then imply
∑ i = 1 N Var ( ψ N , i ∣ G N ) = σ 1 2 ∑ i = 1 N ( α 1 N , i ∘ ) 2 + σ 0 2 ∑ i = 1 N ( α 0 N , i ∘ ) 2 = V N ∘ . \sum_{i=1}^{N}\operatorname{Var}(\psi_{N,i} \mid\mathcal{G}_N) = \sigma_1^2\sum_{i=1}^{N}(\alpha^\circ_{1N,i})^2 + \sigma_0^2\sum_{i=1}^{N}(\alpha^\circ_{0N,i})^2
= V_N^\circ. i = 1 ∑ N Var ( ψ N , i ∣ G N ) = σ 1 2 i = 1 ∑ N ( α 1 N , i ∘ ) 2 + σ 0 2 i = 1 ∑ N ( α 0 N , i ∘ ) 2 = V N ∘ .
For the Lyapunov condition, let σ 2 = min ( σ 0 2 , σ 1 2 ) > 0 \sigma^2 = \min(\sigma_0^2,\sigma_1^2) > 0 σ 2 = min ( σ 0 2 , σ 1 2 ) > 0 and a N = max z , i ∣ α z N , i ∘ ∣ / V N ∘ a_N = \max_{z,i}|\alpha^\circ_{zN,i}|/\sqrt{V_N^\circ} a N = max z , i ∣ α z N , i ∘ ∣/ V N ∘ . The conditional q q q th-moment bound in (C1) gives a finite constant C ε C_\varepsilon C ε such that E ( ∣ ψ N , i ∣ q ∣ G N ) ≤ C ε ∑ z = 0 1 ∣ α z N , i ∘ ∣ q \mathbb{E}(|\psi_{N,i}|^q \mid\mathcal{G}_N) \le C_\varepsilon\sum_{z=0}^{1}|\alpha^\circ_{zN,i}|^q E ( ∣ ψ N , i ∣ q ∣ G N ) ≤ C ε ∑ z = 0 1 ∣ α z N , i ∘ ∣ q . After summing over i i i and factoring each coefficient power into a maximum factor of order q − 2 q-2 q − 2 and a squared coefficient, we obtain
∑ i = 1 N E ( ∣ ψ N , i ∣ q ∣ G N ) ( V N ∘ ) q / 2 ≤ C ε a N q − 2 ∑ z = 0 1 ∑ i = 1 N ( α z N , i ∘ ) 2 V N ∘ ≤ C ε σ 2 a N q − 2 = o p ( 1 ) , \frac{\sum_{i=1}^{N}\mathbb{E}(|\psi_{N,i}|^q \mid\mathcal{G}_N)}{(V_N^\circ)^{q/2}}
\le C_\varepsilon a_N^{q-2}\frac{\sum_{z=0}^{1}\sum_{i=1}^{N}(\alpha^\circ_{zN,i})^2}{V_N^\circ}
\le\frac{C_\varepsilon}{\sigma^2}a_N^{q-2}=o_p(1), ( V N ∘ ) q /2 ∑ i = 1 N E ( ∣ ψ N , i ∣ q ∣ G N ) ≤ C ε a N q − 2 V N ∘ ∑ z = 0 1 ∑ i = 1 N ( α z N , i ∘ ) 2 ≤ σ 2 C ε a N q − 2 = o p ( 1 ) ,
because V N ∘ ≥ σ 2 ∑ z , i ( α z N , i ∘ ) 2 V_N^\circ\ge\sigma^2\sum_{z,i}(\alpha^\circ_{zN,i})^2 V N ∘ ≥ σ 2 ∑ z , i ( α z N , i ∘ ) 2 and a N ⟶ p 0 a_N \mathrel{\stackrel{p}{\longrightarrow}} 0 a N ⟶ p 0 by (C5). This is the conditional Lyapunov ratio. To average over the auxiliary randomization variables while retaining the base conditioning, take an arbitrary subsequence of trial sizes. Convergence in probability of the Lyapunov ratio yields a further subsequence on which that ratio converges to zero almost surely. Along this further subsequence, the conditional Lyapunov theorem applies to the corresponding triangular array and gives, for almost every realization of the conditioning sequence,
E [ exp { i t ∑ i = 1 N ψ N , i V N ∘ } ∣ G N ] ⟶ exp ( − t 2 / 2 ) \mathbb{E}\left[\left.\exp\left\{it\frac{\sum_{i=1}^{N}\psi_{N,i}}{\sqrt{V_N^\circ}}\right\}\right|\mathcal{G}_N\right]\longrightarrow\exp(-t^2/2) E [ exp { i t V N ∘ ∑ i = 1 N ψ N , i } G N ] ⟶ exp ( − t 2 /2 )
for every fixed t ∈ R t \in\mathbb{R} t ∈ R . The conditional characteristic functions have absolute value at most one. Dominated convergence, after averaging over the auxiliary uniforms under the fixed base conditioning, yields the desired conditional characteristic-function convergence. Every subsequence has a further subsequence with this same limit, so the subsequence principle gives
∑ i = 1 N ψ N , i V N ∘ → d N ( 0 , 1 ) \frac{\sum_{i=1}^{N}\psi_{N,i}}{\sqrt{V_N^\circ}} \xrightarrow{d} \mathcal{N}(0,1) V N ∘ ∑ i = 1 N ψ N , i d N ( 0 , 1 )
for the original sequence. Because the statistic remains centered at the conditioned target Δ N \Delta_N Δ N , the limiting variance remains conditional on the enrolled covariates. Finally, expanding the oracle estimator and using the definition of B N ∘ B_N^\circ B N ∘ gives Δ ^ N ∘ − Δ N − B N ∘ = ∑ i ψ N , i \widehat{\Delta}_N^\circ- \Delta_N - B_N^\circ= \sum_i \psi_{N,i} Δ N ∘ − Δ N − B N ∘ = ∑ i ψ N , i .
Consistency of the Working Variance
Lemma S6 (Consistency of the concurrent residual variance estimators). Under (C1), (C2), (C4), and (C6),
σ ^ z 2 → p σ z 2 , z ∈ { 0 , 1 } , \hat{\sigma}_z^2 \xrightarrow{p} \sigma_z^2,\qquad z \in\{0,1\}, σ ^ z 2 p σ z 2 , z ∈ { 0 , 1 } , and σ ^ H 2 = O p ( 1 ) \hat{\sigma}_H^2 = O_p(1) σ ^ H 2 = O p ( 1 ) .
Proof. Fix z ∈ { 0 , 1 } z \in\{0,1\} z ∈ { 0 , 1 } . Define the full-length vectors Y N ( z ) = { Y 1 ( z ) , … , Y N ( z ) } T \mathbf{Y}_N(z) = \{Y_1(z),\ldots,Y_N(z)\}^{\mathsf{T}} Y N ( z ) = { Y 1 ( z ) , … , Y N ( z ) } T and ε N ( z ) = { ε 1 ( z ) , … , ε N ( z ) } T \boldsymbol{\varepsilon}_N(z) = \{\varepsilon_1(z),\ldots,\varepsilon_N(z)\}^{\mathsf{T}} ε N ( z ) = { ε 1 ( z ) , … , ε N ( z ) } T .
To compare the actual and oracle residual operators, let A z N C \mathbf{A}_{zN}^{C} A z N C be diagonal with i i i th entry 1 ( Z i = z ) 1(Z_i=z) 1 ( Z i = z ) , and let S z N s e l f , C \mathbf{S}_{zN}^{\mathrm{self},C} S z N self , C be the full-length actual self-smoothing matrix defined in the same way as its oracle counterpart in (C6). Set R z N C = A z N C − S z N s e l f , C \mathbf{R}_{zN}^{C} = \mathbf{A}_{zN}^{C} - \mathbf{S}_{zN}^{\mathrm{self},C} R z N C = A z N C − S z N self , C and R z N ∘ = A z N ∘ − S z N s e l f , ∘ \mathbf{R}_{zN}^{\circ} = \mathbf{A}_{zN}^{\circ} - \mathbf{S}_{zN}^{\mathrm{self},\circ} R z N ∘ = A z N ∘ − S z N self , ∘ . Equation (S6) shows that the number of assignment mismatches satisfies J N = O p ( d N ) J_N = O_p(d_N) J N = O p ( d N ) . On a row whose arm indicators agree, Lemma S3 bounds the difference between the two normalized kernel rows. On a mismatched row, the ℓ 1 \ell_1 ℓ 1 distance between the two self-smoothing rows is at most two. Let J N = { j : Z j ≠ Z j ∘ } \mathcal{J}_N = \{j: Z_j \ne Z_j^{\circ}\} J N = { j : Z j = Z j ∘ } and let L N L_N L N denote the maximum column load on the left side of (S2). Thus L N = O p ( 1 ) L_N = O_p(1) L N = O p ( 1 ) . Summing first over rows and then over the mismatched columns gives
∑ i = 1 N ∥ S z N , i s e l f , C − S z N , i s e l f , ∘ ∥ 1 ≤ 2 J N + 2 ∑ j ∈ J N ∑ i = 1 N { S z N , j C ( X i ) + S z N , j ∘ ( X i ) } \sum_{i=1}^{N}\left\|\mathbf{S}_{zN,i}^{\mathrm{self},C}-\mathbf{S}_{zN,i}^{\mathrm{self},\circ}\right\|_1
\le2J_N+2\sum_{j\in\mathcal{J}_N}\sum_{i=1}^{N}\left\{S_{zN,j}^{C}(X_i)+S_{zN,j}^{\circ}(X_i)\right\} i = 1 ∑ N S z N , i self , C − S z N , i self , ∘ 1 ≤ 2 J N + 2 j ∈ J N ∑ i = 1 ∑ N { S z N , j C ( X i ) + S z N , j ∘ ( X i ) }
≤ 2 J N ( 1 + L N ) = O p ( d N ) . (18) \le2J_N(1+L_N)=O_p(d_N). \tag*{(18)} ≤ 2 J N ( 1 + L N ) = O p ( d N ) . (18)
Here the matched rows contribute at most twice the total kernel mass assigned to columns in J N \mathcal{J}_N J N , and the at most J N J_N J N mismatched rows each contribute at most two. Every row difference has ℓ 1 \ell_1 ℓ 1 norm at most two. Therefore
∥ S z N s e l f , C − S z N s e l f , ∘ ∥ F 2 = ∑ i = 1 N ∥ S z N , i s e l f , C − S z N , i s e l f , ∘ ∥ 2 2 \left\|\mathbf{S}_{zN}^{\mathrm{self},C}-\mathbf{S}_{zN}^{\mathrm{self},\circ}\right\|_F^2
=\sum_{i=1}^{N}\left\|\mathbf{S}_{zN,i}^{\mathrm{self},C}-\mathbf{S}_{zN,i}^{\mathrm{self},\circ}\right\|_2^2 S z N self , C − S z N self , ∘ F 2 = i = 1 ∑ N S z N , i self , C − S z N , i self , ∘ 2 2
≤ 2 ∑ i = 1 N ∥ S z N , i s e l f , C − S z N , i s e l f , ∘ ∥ 1 = O p ( d N ) . \le2\sum_{i=1}^{N}\left\|\mathbf{S}_{zN,i}^{\mathrm{self},C}-\mathbf{S}_{zN,i}^{\mathrm{self},\circ}\right\|_1
=O_p(d_N). ≤ 2 i = 1 ∑ N S z N , i self , C − S z N , i self , ∘ 1 = O p ( d N ) .
Let Y N max = max i , z ∣ Y i ( z ) ∣ Y_N^{\max}=\max_{i,z}|Y_i(z)| Y N m a x = max i , z ∣ Y i ( z ) ∣ . Equation (S8) gives Y N max = O p ( N 1 / q ) Y_N^{\max}=O_p(N^{1/q}) Y N m a x = O p ( N 1/ q ) . The selector matrices differ on at most J N J_N J N diagonal entries, so ∥ ( A z N C − A z N ∘ ) Y N ( z ) ∥ 2 2 ≤ J N ( Y N max ) 2 \|(\mathbf{A}_{zN}^{C}-\mathbf{A}_{zN}^{\circ})\mathbf{Y}_N(z)\|_2^2\le J_N(Y_N^{\max})^2 ∥ ( A z N C − A z N ∘ ) Y N ( z ) ∥ 2 2 ≤ J N ( Y N m a x ) 2 . For the smoothing part, the rowwise bound ∣ v T Y N ( z ) ∣ ≤ Y N max ∥ v ∥ 1 |\mathbf{v}^{\mathsf{T}}\mathbf{Y}_N(z)|\le Y_N^{\max}\|\mathbf{v}\|_1 ∣ v T Y N ( z ) ∣ ≤ Y N m a x ∥ v ∥ 1 and (S9) give
∥ ( S z N s e l f , C − S z N s e l f , ∘ ) Y N ( z ) ∥ 2 2 ≤ 2 ( Y N max ) 2 ∑ i = 1 N ∥ S z N , i s e l f , C − S z N , i s e l f , ∘ ∥ 1 = O p { ( Y N max ) 2 d N } . \|(\mathbf{S}_{zN}^{\mathrm{self},C}-\mathbf{S}_{zN}^{\mathrm{self},\circ})\mathbf{Y}_N(z)\|_2^2
\le2(Y_N^{\max})^2\sum_{i=1}^{N}\left\|\mathbf{S}_{zN,i}^{\mathrm{self},C}-\mathbf{S}_{zN,i}^{\mathrm{self},\circ}\right\|_1
=O_p\{(Y_N^{\max})^2d_N\}. ∥ ( S z N self , C − S z N self , ∘ ) Y N ( z ) ∥ 2 2 ≤ 2 ( Y N m a x ) 2 i = 1 ∑ N S z N , i self , C − S z N , i self , ∘ 1 = O p {( Y N m a x ) 2 d N } .
Applying ∥ a + b ∥ 2 2 ≤ 2 ∥ a ∥ 2 2 + 2 ∥ b ∥ 2 2 \lVert a+b\rVert_2^2\leq2\lVert a\rVert_2^2+2\lVert b\rVert_2^2 ∥ a + b ∥ 2 2 ≤ 2 ∥ a ∥ 2 2 + 2 ∥ b ∥ 2 2 to the selector and smoothing parts yields
1 N ∥ ( R z N C − R z N ∘ ) Y N ( z ) ∥ 2 2 = O p { ( Y N max ) 2 d N N } = O p ( d N N − 1 + 2 / q ) = o p ( 1 ) . \frac{1}{N}\left\lVert\left(R_{zN}^{C}-R_{zN}^{\circ}\right)Y_N(z)\right\rVert_2^2=O_p\left\{\frac{\left(Y_N^{\max}\right)^2d_N}{N}\right\}
=O_p\left(d_NN^{-1+2/q}\right)=o_p(1). N 1 ( R z N C − R z N ∘ ) Y N ( z ) 2 2 = O p { N ( Y N m a x ) 2 d N } = O p ( d N N − 1 + 2/ q ) = o p ( 1 ) .
For the last equality, write the middle rate as { d N N − 1 / 2 + 1 / q } N − 1 / 2 + 1 / q \{d_NN^{-1/2+1/q}\}N^{-1/2+1/q} { d N N − 1/2 + 1/ q } N − 1/2 + 1/ q and use q > 2 q>2 q > 2 together with (S1).
For the oracle residual sum of squares, write Y N ( z ) = μ z N + ε N ( z ) Y_N(z)=\mu_{zN}+\varepsilon_N(z) Y N ( z ) = μ z N + ε N ( z ) . Conditional on G N \mathcal{G}_N G N , the oracle selector and smoother are fixed and independent of the potential-outcome errors. Let p = min ( q / 2 , 2 ) > 1 p=\min(q/2,2)>1 p = min ( q /2 , 2 ) > 1 and define ζ N , i = 1 ( Z i ∘ = z ) { ε i ( z ) 2 − σ z 2 } \zeta_{N,i}=\mathbf{1}(Z_i^\circ=z)\{\varepsilon_i(z)^2-\sigma_z^2\} ζ N , i = 1 ( Z i ∘ = z ) { ε i ( z ) 2 − σ z 2 } . The variables ζ N , i \zeta_{N,i} ζ N , i are conditionally independent and centered. Since 2 p ≤ q 2p\leq q 2 p ≤ q , (C1) also gives sup i E ( ∣ ζ N , i ∣ p ∣ G N ) = O ( 1 ) \sup_i\mathbb{E}(|\zeta_{N,i}|^p\mid\mathcal{G}_N)=O(1) sup i E ( ∣ ζ N , i ∣ p ∣ G N ) = O ( 1 ) almost surely. The von Bahr–Esseen inequality [7 ] , with the ordinary variance identity when p = 2 p=2 p = 2 , therefore yields
E ( ∣ 1 N ∑ i = 1 N ζ N , i ∣ p ∣ G N ) ≤ C p N p ∑ i = 1 N E ( ∣ ζ N , i ∣ p ∣ G N ) = O ( N 1 − p ) ⟶ 0. \mathbb{E}\left(\left.\left\lvert\frac{1}{N}\sum_{i=1}^{N}\zeta_{N,i}\right\rvert^p\right|\mathcal{G}_N\right)
\leq\frac{C_p}{N^p}\sum_{i=1}^{N}\mathbb{E}(|\zeta_{N,i}|^p\mid\mathcal{G}_N)
=O(N^{1-p})\longrightarrow0. E ( N 1 i = 1 ∑ N ζ N , i p G N ) ≤ N p C p i = 1 ∑ N E ( ∣ ζ N , i ∣ p ∣ G N ) = O ( N 1 − p ) ⟶ 0.
Markov’s inequality, followed by averaging over G N \mathcal{G}_N G N , proves N − 1 ∑ i ζ N , i ⟶ p 0 N^{-1}\sum_i\zeta_{N,i}\mathrel{\mathop{\longrightarrow}^{p}}0 N − 1 ∑ i ζ N , i ⟶ p 0 . The sum of squared errors for participants assigned to arm z z z equals N z ∘ σ z 2 + ∑ i ζ N , i N_z^\circ\sigma_z^2+\sum_i\zeta_{N,i} N z ∘ σ z 2 + ∑ i ζ N , i . Therefore, the arm-fraction condition in (C6) gives N − 1 ∥ A z N ∘ ε N ( z ) ∥ 2 2 ⟶ p π z σ z 2 N^{-1}\lVert A_{zN}^\circ\varepsilon_N(z)\rVert_2^2\mathrel{\mathop{\longrightarrow}^{p}}\pi_z\sigma_z^2 N − 1 ∥ A z N ∘ ε N ( z ) ∥ 2 2 ⟶ p π z σ z 2 . Next, conditional homoskedasticity and independence give
E ( ∥ S z N s e l f , ∘ ε N ( z ) ∥ 2 2 ∣ G N ) = σ z 2 ∥ S z N s e l f , ∘ ∥ F 2 = o p ( N ) . \mathbb{E}\left(\left.\lVert S_{zN}^{\mathrm{self},\circ}\varepsilon_N(z)\rVert_2^2\right|\mathcal{G}_N\right)
=\sigma_z^2\lVert S_{zN}^{\mathrm{self},\circ}\rVert_F^2=o_p(N). E ( ∥ S z N self , ∘ ε N ( z ) ∥ 2 2 G N ) = σ z 2 ∥ S z N self , ∘ ∥ F 2 = o p ( N ) .
For the conditional Markov argument, set Q N = N − 1 ∥ S z N s e l f , ∘ ε N ( z ) ∥ 2 2 Q_N=N^{-1}\lVert S_{zN}^{\mathrm{self},\circ}\varepsilon_N(z)\rVert_2^2 Q N = N − 1 ∥ S z N self , ∘ ε N ( z ) ∥ 2 2 . For any ϵ , ϵ 0 > 0 \epsilon,\epsilon_0>0 ϵ , ϵ 0 > 0 , Pr ( Q N > ϵ ) ≤ Pr { E ( Q N ∣ G N ) > ϵ 0 } + ϵ 0 / ϵ \Pr(Q_N>\epsilon)\leq\Pr\{\mathbb{E}(Q_N\mid\mathcal{G}_N)>\epsilon_0\}+\epsilon_0/\epsilon Pr ( Q N > ϵ ) ≤ Pr { E ( Q N ∣ G N ) > ϵ 0 } + ϵ 0 / ϵ . The first term tends to zero by (C6). Letting ϵ 0 \epsilon_0 ϵ 0 tend to zero after taking the limit shows that Q N ⟶ p 0 Q_N\mathrel{\mathop{\longrightarrow}^{p}}0 Q N ⟶ p 0 .
Since R z N ∘ = A z N ∘ − S z N s e l f , ∘ R_{zN}^\circ=A_{zN}^\circ-S_{zN}^{\mathrm{self},\circ} R z N ∘ = A z N ∘ − S z N self , ∘ , the difference-of-squares identity and Cauchy–Schwarz give
∣ 1 N ∥ R z N ∘ ε N ( z ) ∥ 2 2 − 1 N ∥ A z N ∘ ε N ( z ) ∥ 2 2 ∣ ≤ Q N + 2 Q N 1 / 2 { 1 N ∥ A z N ∘ ε N ( z ) ∥ 2 2 } 1 / 2 = o p ( 1 ) . \left|\frac{1}{N}\lVert R_{zN}^\circ\varepsilon_N(z)\rVert_2^2-\frac{1}{N}\lVert A_{zN}^\circ\varepsilon_N(z)\rVert_2^2\right|
\leq Q_N+2Q_N^{1/2}\left\{\frac{1}{N}\lVert A_{zN}^\circ\varepsilon_N(z)\rVert_2^2\right\}^{1/2}=o_p(1). N 1 ∥ R z N ∘ ε N ( z ) ∥ 2 2 − N 1 ∥ A z N ∘ ε N ( z ) ∥ 2 2 ≤ Q N + 2 Q N 1/2 { N 1 ∥ A z N ∘ ε N ( z ) ∥ 2 2 } 1/2 = o p ( 1 ) .
It follows that N − 1 ∥ R z N ∘ ε N ( z ) ∥ 2 2 ⟶ p π z σ z 2 N^{-1}\lVert R_{zN}^\circ\varepsilon_N(z)\rVert_2^2\mathrel{\mathop{\longrightarrow}^{p}}\pi_z\sigma_z^2 N − 1 ∥ R z N ∘ ε N ( z ) ∥ 2 2 ⟶ p π z σ z 2 . The mean-approximation requirement in (S3) gives N − 1 ∥ R z N ∘ μ z N ∥ 2 2 = o p ( 1 ) N^{-1}\lVert R_{zN}^\circ\mu_{zN}\rVert_2^2=o_p(1) N − 1 ∥ R z N ∘ μ z N ∥ 2 2 = o p ( 1 ) . Applying the same difference-of-squares bound to R z N ∘ Y N ( z ) = R z N ∘ μ z N + R z N ∘ ε N ( z ) R_{zN}^\circ Y_N(z)=R_{zN}^\circ\mu_{zN}+R_{zN}^\circ\varepsilon_N(z) R z N ∘ Y N ( z ) = R z N ∘ μ z N + R z N ∘ ε N ( z ) shows that the mean term and its cross product are o p ( 1 ) o_p(1) o p ( 1 ) after division by N N N . Therefore
1 N ∥ R z N ∘ Y N ( z ) ∥ 2 2 ⟶ p π z σ z 2 . (19) \frac{1}{N}\lVert R_{zN}^\circ Y_N(z)\rVert_2^2\mathrel{\mathop{\longrightarrow}^{p}}\pi_z\sigma_z^2. \tag*{(19)} N 1 ∥ R z N ∘ Y N ( z ) ∥ 2 2 ⟶ p π z σ z 2 . (19)
The residual-operator bound transfers this limit to the actual RADISH path. Define
a N = R z N ∘ Y N ( z ) , b N = ( R z N C − R z N ∘ ) Y N ( z ) . a_N=R_{zN}^\circ Y_N(z),\qquad b_N=(R_{zN}^{C}-R_{zN}^\circ)Y_N(z). a N = R z N ∘ Y N ( z ) , b N = ( R z N C − R z N ∘ ) Y N ( z ) .
The residual-operator bound above and (S10) give, respectively,
N − 1 / 2 ∥ b N ∥ 2 = o p ( 1 ) , N − 1 / 2 ∥ a N ∥ 2 = O p ( 1 ) . N^{-1/2}\|b_N\|_2=o_p(1), \qquad N^{-1/2}\|a_N\|_2=O_p(1). N − 1/2 ∥ b N ∥ 2 = o p ( 1 ) , N − 1/2 ∥ a N ∥ 2 = O p ( 1 ) .
Since R z N C Y N ( z ) = a N + b N R_{zN}^{C}Y_N(z)=a_N+b_N R z N C Y N ( z ) = a N + b N ,
∣ 1 N ∥ a N + b N ∥ 2 2 − 1 N ∥ a N ∥ 2 2 ∣ ≤ 2 N ∥ a N ∥ 2 ∥ b N ∥ 2 + 1 N ∥ b N ∥ 2 2 = o p ( 1 ) . \left|\frac{1}{N}\|a_N+b_N\|_2^2-\frac{1}{N}\|a_N\|_2^2\right|\leq\frac{2}{N}\|a_N\|_2\|b_N\|_2+\frac{1}{N}\|b_N\|_2^2=o_p(1). N 1 ∥ a N + b N ∥ 2 2 − N 1 ∥ a N ∥ 2 2 ≤ N 2 ∥ a N ∥ 2 ∥ b N ∥ 2 + N 1 ∥ b N ∥ 2 2 = o p ( 1 ) .
Combining this bound with (S10) yields
1 N ∥ R z N C Y N ( z ) ∥ 2 2 → p π z σ z 2 . (20) \frac{1}{N}\|R_{zN}^{C}Y_N(z)\|_2^2\xrightarrow{p}\pi_z\sigma_z^2. \tag*{(20)} N 1 ∥ R z N C Y N ( z ) ∥ 2 2 p π z σ z 2 . (20)
The numerator on the left is exactly the within-arm residual sum of squares in § 5.4.3 , written on the common participant index.
To obtain the variance limit, we also normalize the residual degrees of freedom. Because ∣ N z − N z ∘ ∣ ≤ J N |N_z-N_z^{\circ}|\leq J_N ∣ N z − N z ∘ ∣ ≤ J N and d N / N → 0 d_N/N\to0 d N / N → 0 , N z / N → p π z N_z/N\xrightarrow{p}\pi_z N z / N p π z . The Frobenius bound established above and (S3) give, by the squared triangle inequality,
1 N ∥ S z N s e l f , C ∥ F 2 ≤ 2 N ∥ S z N s e l f , ∘ ∥ F 2 + 2 N ∥ S z N s e l f , C − S z N s e l f , ∘ ∥ F 2 → p 0. \frac{1}{N}\|S_{zN}^{\mathrm{self},C}\|_F^2\leq\frac{2}{N}\|S_{zN}^{\mathrm{self},\circ}\|_F^2+\frac{2}{N}\|S_{zN}^{\mathrm{self},C}-S_{zN}^{\mathrm{self},\circ}\|_F^2\xrightarrow{p}0. N 1 ∥ S z N self , C ∥ F 2 ≤ N 2 ∥ S z N self , ∘ ∥ F 2 + N 2 ∥ S z N self , C − S z N self , ∘ ∥ F 2 p 0.
Also, ∣ tr ( S z N s e l f , C ) ∣ ≤ N ∥ S z N s e l f , C ∥ F = o p ( N ) |\operatorname{tr}(S_{zN}^{\mathrm{self},C})|\leq\sqrt{N}\|S_{zN}^{\mathrm{self},C}\|_F=o_p(N) ∣ tr ( S z N self , C ) ∣ ≤ N ∥ S z N self , C ∥ F = o p ( N ) . Let dof z r a w \operatorname{dof}_z^{\mathrm{raw}} dof z raw denote the unfloored residual degrees of freedom. Writing it on the common participant index gives
dof z r a w N = N z N − 2 N tr ( S z N s e l f , C ) + 1 N ∥ S z N s e l f , C ∥ F 2 → p π z . \frac{\operatorname{dof}_z^{\mathrm{raw}}}{N}=\frac{N_z}{N}-\frac{2}{N}\operatorname{tr}(S_{zN}^{\mathrm{self},C})+\frac{1}{N}\|S_{zN}^{\mathrm{self},C}\|_F^2\xrightarrow{p}\pi_z. N dof z raw = N N z − N 2 tr ( S z N self , C ) + N 1 ∥ S z N self , C ∥ F 2 p π z .
Because π z > 0 \pi_z>0 π z > 0 , Pr { dof z r a w > 1 } → 1 \Pr\{\operatorname{dof}_z^{\mathrm{raw}}>1\}\to1 Pr { dof z raw > 1 } → 1 . Hence dof z = max { dof z r a w , 1 } \operatorname{dof}_z=\max\{\operatorname{dof}_z^{\mathrm{raw}},1\} dof z = max { dof z raw , 1 } has the same normalized limit. Dividing (S11) by this positive limit proves σ ^ z 2 → p σ z 2 \widehat{\sigma}_z^2\xrightarrow{p}\sigma_z^2 σ z 2 p σ z 2 . Finally, because each historical smoothing row is convex and the residual degrees of freedom are bounded below by one, the historical residual sum of squares is bounded by 4 N H max j ∣ Y j H ∣ 2 4N_H\max_j|Y_j^H|^2 4 N H max j ∣ Y j H ∣ 2 . Conditional on the fixed historical dataset, this gives σ ^ H 2 = O p ( 1 ) \widehat{\sigma}_H^2=O_p(1) σ H 2 = O p ( 1 ) even when the bandwidth varies along the trial sequence.
Lemma S7 (Working-variance consistency). Under (C1), (C2), (C4), (C5), and (C6),
V ^ l i n V N ∘ → p 1 , V ^ C V N ∘ → p 1 , V ^ N V N ∘ → p 1. (21) \frac{\widehat{V}_{\mathrm{lin}}}{V_N^\circ}\xrightarrow{p}1,\qquad
\frac{\widehat{V}_{C}}{V_N^\circ}\xrightarrow{p}1,\qquad
\frac{\widehat{V}_{N}}{V_N^\circ}\xrightarrow{p}1. \tag*{(21)} V N ∘ V lin p 1 , V N ∘ V C p 1 , V N ∘ V N p 1. (21)
Proof. Set α H N ∘ = 0 \alpha_{HN}^{\circ}=0 α H N ∘ = 0 and d z N = α ^ z N − α z N ∘ d_{zN}=\widehat{\alpha}_{zN}-\alpha_{zN}^{\circ} d z N = α z N − α z N ∘ for z ∈ { 1 , 0 , H } z\in\{1,0,H\} z ∈ { 1 , 0 , H } . By (S5), ∥ d z N ∥ 2 = o p ( N − 1 / 2 ) \|d_{zN}\|_2=o_p(N^{-1/2}) ∥ d z N ∥ 2 = o p ( N − 1/2 ) . Condition (C5) also gives ∥ α z N ∘ ∥ 2 = O p ( N − 1 / 2 ) \|\alpha_{zN}^{\circ}\|_2=O_p(N^{-1/2}) ∥ α z N ∘ ∥ 2 = O p ( N − 1/2 ) for z ∈ { 0 , 1 } z\in\{0,1\} z ∈ { 0 , 1 } . The identity ∥ a + d ∥ 2 2 − ∥ a ∥ 2 2 = 2 a T d + ∥ d ∥ 2 2 \|a+d\|_2^2-\|a\|_2^2=2a^\mathsf{T}d+\|d\|_2^2 ∥ a + d ∥ 2 2 − ∥ a ∥ 2 2 = 2 a T d + ∥ d ∥ 2 2 and Cauchy–Schwarz imply
∥ α ^ z N ∥ 2 2 − ∥ α z N ∘ ∥ 2 2 ≤ 2 ∥ α z N ∘ ∥ 2 ∥ d z N ∥ 2 + ∥ d z N ∥ 2 2 = o p ( N − 1 ) , z ∈ { 0 , 1 } , \begin{aligned}
\|\widehat{\alpha}_{zN}\|_2^2-\|\alpha_{zN}^{\circ}\|_2^2\leq2\|\alpha_{zN}^{\circ}\|_2\|d_{zN}\|_2+\|d_{zN}\|_2^2 \\
&=o_p(N^{-1}),\qquad z\in\{0,1\},
\end{aligned} ∥ α z N ∥ 2 2 − ∥ α z N ∘ ∥ 2 2 ≤ 2∥ α z N ∘ ∥ 2 ∥ d z N ∥ 2 + ∥ d z N ∥ 2 2 = o p ( N − 1 ) , z ∈ { 0 , 1 } ,
and ∥ α ^ H N ∥ 2 2 = o p ( N − 1 ) \lVert\widehat{\alpha}_{HN}\rVert_2^2=o_p(N^{-1}) ∥ α H N ∥ 2 2 = o p ( N − 1 ) . The triangle inequality also gives ∥ α ^ z N ∥ 2 ≤ ∥ α z N ∘ ∥ 2 + ∥ d z N ∥ 2 = O p ( N − 1 / 2 ) \lVert\widehat{\alpha}_{zN}\rVert_2\leq\lVert\alpha^\circ_{zN}\rVert_2+\lVert d_{zN}\rVert_2=O_p(N^{-1/2}) ∥ α z N ∥ 2 ≤ ∥ α z N ∘ ∥ 2 + ∥ d z N ∥ 2 = O p ( N − 1/2 ) , and hence ∥ α ^ z N ∥ 2 2 = O p ( N − 1 ) \lVert\widehat{\alpha}_{zN}\rVert_2^2=O_p(N^{-1}) ∥ α z N ∥ 2 2 = O p ( N − 1 ) . The same coupling and terminal-weight condition give ∥ α ^ 0 N C − α 0 N ∘ ∥ 2 = o p ( N − 1 / 2 ) \lVert\widehat{\alpha}^{C}_{0N}-\alpha^\circ_{0N}\rVert_2=o_p(N^{-1/2}) ∥ α 0 N C − α 0 N ∘ ∥ 2 = o p ( N − 1/2 ) , so the same squared-norm argument applies to the full-length concurrent-only coefficient used in V ^ C \widehat{V}_C V C . Zero-padding preserves the norm of the arm-compressed vector α 0 C \alpha^C_0 α 0 C .
Lemma S6 gives σ ^ z 2 − σ z 2 = o p ( 1 ) \widehat{\sigma}_z^2-\sigma_z^2=o_p(1) σ z 2 − σ z 2 = o p ( 1 ) and σ ^ H 2 = O p ( 1 ) \widehat{\sigma}_H^2=O_p(1) σ H 2 = O p ( 1 ) . Together with the coefficient bounds, adding and subtracting ∑ z = 0 1 σ z 2 ∥ α ^ z N ∥ 2 2 \sum_{z=0}^{1}\sigma_z^2\lVert\widehat{\alpha}_{zN}\rVert_2^2 ∑ z = 0 1 σ z 2 ∥ α z N ∥ 2 2 gives
V ^ l i n − V N ∘ = ∑ z = 0 1 ( σ ^ z 2 − σ z 2 ) ∥ α ^ z N ∥ 2 2 + ∑ z = 0 1 σ z 2 ( ∥ α ^ z N ∥ 2 2 − ∥ α z N ∘ ∥ 2 2 ) + σ ^ H 2 ∥ α ^ H N ∥ 2 2 = o p ( N − 1 ) . \begin{aligned}
\widehat{V}_{\mathrm{lin}}-V_N^\circ
&=\sum_{z=0}^{1}(\widehat{\sigma}_z^2-\sigma_z^2)\lVert\widehat{\alpha}_{zN}\rVert_2^2\\
&\quad+\sum_{z=0}^{1}\sigma_z^2\left(\lVert\widehat{\alpha}_{zN}\rVert_2^2-\lVert\alpha^\circ_{zN}\rVert_2^2\right)+\widehat{\sigma}_H^2\lVert\widehat{\alpha}_{HN}\rVert_2^2\\
&=o_p(N^{-1}).
\end{aligned} V lin − V N ∘ = z = 0 ∑ 1 ( σ z 2 − σ z 2 ) ∥ α z N ∥ 2 2 + z = 0 ∑ 1 σ z 2 ( ∥ α z N ∥ 2 2 − ∥ α z N ∘ ∥ 2 2 ) + σ H 2 ∥ α H N ∥ 2 2 = o p ( N − 1 ) .
The first sum is o p ( N − 1 ) o_p(N^{-1}) o p ( N − 1 ) because the two factors in each product have orders o p ( 1 ) o_p(1) o p ( 1 ) and O p ( N − 1 ) O_p(N^{-1}) O p ( N − 1 ) . The second sum is o p ( N − 1 ) o_p(N^{-1}) o p ( N − 1 ) by the squared-norm bounds above. The historical term is o p ( N − 1 ) o_p(N^{-1}) o p ( N − 1 ) because its two factors have orders O p ( 1 ) O_p(1) O p ( 1 ) and o p ( N − 1 ) o_p(N^{-1}) o p ( N − 1 ) .
Dividing by the oracle variance gives the required ratios. The preceding display shows N ( V ^ l i n − V N ∘ ) → p 0 N(\widehat{V}_{\mathrm{lin}}-V_N^\circ)\overset{p}{\to}0 N ( V lin − V N ∘ ) → p 0 , whereas N V N ∘ → p v > 0 NV_N^\circ\overset{p}{\to}v>0 N V N ∘ → p v > 0 by (C5). Therefore V ^ l i n / V N ∘ − 1 = N ( V ^ l i n − V N ∘ ) / ( N V N ∘ ) → p 0 \widehat{V}_{\mathrm{lin}}/V_N^\circ-1=N(\widehat{V}_{\mathrm{lin}}-V_N^\circ)/(NV_N^\circ)\overset{p}{\to}0 V lin / V N ∘ − 1 = N ( V lin − V N ∘ ) / ( N V N ∘ ) → p 0 . Replacing α ^ 0 N \widehat{\alpha}_{0N} α 0 N by α ^ 0 N C \widehat{\alpha}^{C}_{0N} α 0 N C in the same calculation gives V ^ C / V N ∘ → p 1 \widehat{V}_C/V_N^\circ\overset{p}{\to}1 V C / V N ∘ → p 1 . Since the maximum is a continuous map, V ^ N = max { V ^ l i n , V ^ C } \widehat{V}_N=\max\{\widehat{V}_{\mathrm{lin}},\widehat{V}_C\} V N = max { V lin , V C } also satisfies V ^ N / V N ∘ → p 1 \widehat{V}_N/V_N^\circ\overset{p}{\to}1 V N / V N ∘ → p 1 , proving (S12).
Proof of Studentized Asymptotic Normality
Proof of Theorem 1. Lemma S5 gives ( Δ ^ N ∘ − Δ N − B N ∘ ) / V N ∘ → d N ( 0 , 1 ) (\widehat{\Delta}^{\circ}_N-\Delta_N-B_N^\circ)/\sqrt{V_N^\circ}\overset{d}{\to}\mathcal{N}(0,1) ( Δ N ∘ − Δ N − B N ∘ ) / V N ∘ → d N ( 0 , 1 ) . Lemma S4 then gives Δ ^ N − Δ ^ N ∘ = o p ( N − 1 / 2 ) \widehat{\Delta}_N-\widehat{\Delta}^{\circ}_N=o_p(N^{-1/2}) Δ N − Δ N ∘ = o p ( N − 1/2 ) . Because N V N ∘ → p v > 0 NV_N^\circ\overset{p}{\to}v>0 N V N ∘ → p v > 0 , it follows that ( Δ ^ N − Δ ^ N ∘ ) / V N ∘ → p 0 (\widehat{\Delta}_N-\widehat{\Delta}^{\circ}_N)/\sqrt{V_N^\circ}\overset{p}{\to}0 ( Δ N − Δ N ∘ ) / V N ∘ → p 0 . Finally, Lemma S7 gives V ^ N / V N ∘ → p 1 \widehat{V}_N/V_N^\circ\overset{p}{\to}1 V N / V N ∘ → p 1 .
The remaining term is the integrated smoothing bias. Condition (C5) and the variance consistency established above imply N V ^ N → p v N\widehat{V}_N\overset{p}{\to}v N V N → p v , so 1 / V ^ N = O p ( N ) 1/\sqrt{\widehat{V}_N}=O_p(\sqrt{N}) 1/ V N = O p ( N ) . Condition (C3) therefore implies B N ∘ / V ^ N = o p ( 1 ) B_N^\circ/\sqrt{\widehat{V}_N}=o_p(1) B N ∘ / V N = o p ( 1 ) . The exact decomposition
Δ ^ N − Δ N V ^ N = Δ ^ N ∘ − Δ N − B N ∘ V N ∘ V N ∘ V ^ N + Δ ^ N − Δ ^ N ∘ V N ∘ V N ∘ V ^ N + B N ∘ V ^ N \frac{\widehat{\Delta}_N-\Delta_N}{\sqrt{\widehat{V}_N}}
=
\frac{\widehat{\Delta}^{\circ}_N-\Delta_N-B_N^\circ}{\sqrt{V_N^\circ}}
\sqrt{\frac{V_N^\circ}{\widehat{V}_N}}
+
\frac{\widehat{\Delta}_N-\widehat{\Delta}^{\circ}_N}{\sqrt{V_N^\circ}}
\sqrt{\frac{V_N^\circ}{\widehat{V}_N}}
+
\frac{B_N^\circ}{\sqrt{\widehat{V}_N}} V N Δ N − Δ N = V N ∘ Δ N ∘ − Δ N − B N ∘ V N V N ∘ + V N ∘ Δ N − Δ N ∘ V N V N ∘ + V N B N ∘
has a first term converging in distribution to a standard normal variable by Lemma S5 and variance ratio consistency. The second term vanishes by Lemma S4, and the third vanishes by the bias bound above. Slutsky’s theorem therefore proves the convergence asserted in Theorem 1.
Proof of Corollary S1. Let T N = ( Δ ^ N − Δ N ) / V ^ N T_N = (\widehat{\Delta}_N - \Delta_N)/\sqrt{\widehat{V}_N} T N = ( Δ N − Δ N ) / V N . Theorem 1 gives T N → d N ( 0 , 1 ) T_N \mathrel{\overset{d}{\to}} \mathcal{N}(0,1) T N → d N ( 0 , 1 ) . Because the standard normal distribution is continuous at ± z 1 − α / 2 \pm z_{1-\alpha/2} ± z 1 − α /2 , Pr { − z 1 − α / 2 ≤ T N ≤ z 1 − α / 2 } → 1 − α \Pr\{-z_{1-\alpha/2} \le T_N \le z_{1-\alpha/2}\} \to1-\alpha Pr { − z 1 − α /2 ≤ T N ≤ z 1 − α /2 } → 1 − α . The event in this probability is equivalent to ∣ Δ ^ N − Δ N ∣ ≤ z 1 − α / 2 V ^ N |\widehat{\Delta}_N-\Delta_N| \le z_{1-\alpha/2}\sqrt{\widehat{V}_N} ∣ Δ N − Δ N ∣ ≤ z 1 − α /2 V N , which is exactly the event that the interval in Equation (8) of the main text contains Δ N \Delta_N Δ N . This proves the coverage statement. Under H 0 : Δ N = 0 H_0 : \Delta_N = 0 H 0 : Δ N = 0 , the test rejects precisely on the complementary event ∣ T N ∣ > z 1 − α / 2 |T_N| > z_{1-\alpha/2} ∣ T N ∣ > z 1 − α /2 . Its probability therefore converges to α \alpha α .
Algorithms
Algorithm S1: Stage I: local historical reference construction.
Input: Historical control data D H = { ( X j H , Y j H ) } j = 1 N H D_H = \{(X_j^H,Y_j^H)\}_{j=1}^{N_H} D H = {( X j H , Y j H ) } j = 1 N H ; kernel bandwidth h h h ; variance floor ϵ v = 10 − 12 \epsilon_v = 10^{-12} ϵ v = 1 0 − 12 .
Output: Local reference center θ ( x ) \theta(x) θ ( x ) and undiscounted reference variance τ ^ H 2 ( x ) \widehat{\tau}_H^2(x) τ H 2 ( x ) for each query point x x x .
for each query covariate x x x do
w j H ( x ) ← K h ( X j H − x ) , j = 1 , … , N H w_j^H(x) \leftarrow K_h(X_j^H-x),\quad j=1,\ldots,N_H w j H ( x ) ← K h ( X j H − x ) , j = 1 , … , N H ;
θ ( x ) ← [ ∑ j w j H ( x ) Y j H ] / [ ∑ j w j H ( x ) ] \theta(x) \leftarrow[\sum_j w_j^H(x)Y_j^H]/[\sum_j w_j^H(x)] θ ( x ) ← [ ∑ j w j H ( x ) Y j H ] / [ ∑ j w j H ( x )] ;
σ ^ H 2 ( x ) ← max { [ ∑ j w j H ( x ) { Y j H − θ ( x ) } 2 ] / [ ∑ j w j H ( x ) ] , ϵ v } \widehat{\sigma}_H^2(x) \leftarrow\max\{[\sum_j w_j^H(x)\{Y_j^H-\theta(x)\}^2]/[\sum_j w_j^H(x)],\epsilon_v\} σ H 2 ( x ) ← max {[ ∑ j w j H ( x ) { Y j H − θ ( x ) } 2 ] / [ ∑ j w j H ( x )] , ϵ v } ;
n H e f f ( x ) ← [ ∑ j w j H ( x ) ] 2 / [ ∑ j { w j H ( x ) } 2 ] n_H^{\mathrm{eff}}(x) \leftarrow[\sum_j w_j^H(x)]^2/[\sum_j\{w_j^H(x)\}^2] n H eff ( x ) ← [ ∑ j w j H ( x ) ] 2 / [ ∑ j { w j H ( x ) } 2 ] ;
τ ^ H 2 ( x ) ← max { σ ^ H 2 ( x ) / n H e f f ( x ) , ϵ v } \widehat{\tau}_H^2(x) \leftarrow\max\{\widehat{\sigma}_H^2(x)/n_H^{\mathrm{eff}}(x),\epsilon_v\} τ H 2 ( x ) ← max { σ H 2 ( x ) / n H eff ( x ) , ϵ v } ;
end
return { θ ( x ) , τ ^ H 2 ( x ) } \{\theta(x),\widehat{\tau}_H^2(x)\} { θ ( x ) , τ H 2 ( x )} .
Algorithm S2: Stage II: sequential information-ratio randomization.
Input: Stage I outputs { θ ( x ) , τ ^ H 2 ( x ) } \{\theta(x),\widehat{\tau}_H^2(x)\} { θ ( x ) , τ H 2 ( x )} ; planned trial size N N N ; kernel function K h K_h K h and bandwidth h h h ; safeguards n i n i t = 20 n_{\mathrm{init}}=20 n init = 20 , n min = 5 n_{\min}=5 n m i n = 5 , ϵ c = ϵ v = 10 − 12 \epsilon_c=\epsilon_v=10^{-12} ϵ c = ϵ v = 1 0 − 12 .
Output: Treatment assignments { Z 1 , … , Z N } \{Z_1,\ldots,Z_N\} { Z 1 , … , Z N } and allocation-time information ratios { R n A ( X n + 1 ) } n = 0 N − 1 \{R_n^A(X_{n+1})\}_{n=0}^{N-1} { R n A ( X n + 1 ) } n = 0 N − 1 .
for n = 0 , 1 , … , N − 1 n=0,1,\ldots,N-1 n = 0 , 1 , … , N − 1 do
Observe new covariate x ≡ x n + 1 x \equiv x_{n+1} x ≡ x n + 1 ; evaluate θ ( x ) \theta(x) θ ( x ) , τ ^ H 2 ( x ) \widehat{\tau}_H^2(x) τ H 2 ( x ) using the Stage I reference;
if n < n i n i t n<n_{\mathrm{init}} n < n init then
Set R n A ( x ) ← 1 R_n^A(x)\leftarrow1 R n A ( x ) ← 1 ; draw Z n + 1 ∼ Bernoulli ( 1 / 2 ) Z_{n+1}\sim\operatorname{Bernoulli}(1/2) Z n + 1 ∼ Bernoulli ( 1/2 ) , observe Y n + 1 Y_{n+1} Y n + 1 , append the record, and continue;
Compute M 0 , n ( x ) M_{0,n}(x) M 0 , n ( x ) , n ~ 0 , n e f f ( x ) \widetilde{n}_{0,n}^{\mathrm{eff}}(x) n 0 , n eff ( x ) , Y ‾ 0 , n ( x ) \overline{Y}_{0,n}(x) Y 0 , n ( x ) , and σ ^ 0 , n 2 ( x ) \widehat{\sigma}_{0,n}^2(x) σ 0 , n 2 ( x ) via K h K_h K h ;
n ~ 0 , n e f f ( x ) ← max { n ~ 0 , n e f f ( x ) , n min } \widetilde{n}_{0,n}^{\mathrm{eff}}(x)\leftarrow\max\{\widetilde{n}_{0,n}^{\mathrm{eff}}(x),n_{\min}\} n 0 , n eff ( x ) ← max { n 0 , n eff ( x ) , n m i n } ; σ ^ 0 , n 2 ( x ) ← max { σ ^ 0 , n 2 ( x ) , ϵ v } \widehat{\sigma}_{0,n}^2(x)\leftarrow\max\{\widehat{\sigma}_{0,n}^2(x),\epsilon_v\} σ 0 , n 2 ( x ) ← max { σ 0 , n 2 ( x ) , ϵ v } ;
SE n 2 ( x ) ← max { τ ^ H 2 ( x ) + σ ^ 0 , n 2 ( x ) / n ~ 0 , n e f f ( x ) , ϵ v } \operatorname{SE}_n^2(x)\leftarrow\max\{\widehat{\tau}_H^2(x)+\widehat{\sigma}_{0,n}^2(x)/\widetilde{n}_{0,n}^{\mathrm{eff}}(x),\epsilon_v\} SE n 2 ( x ) ← max { τ H 2 ( x ) + σ 0 , n 2 ( x ) / n 0 , n eff ( x ) , ϵ v } ;
D n ( x ) ← [ − log max { 2 Φ ( − ∣ { Y ‾ 0 , n ( x ) − θ ( x ) } / SE n ( x ) ∣ ) , ϵ c } − 1 ] D_n(x)\leftarrow[-\log\max\{2\Phi(-|\{\overline{Y}_{0,n}(x)-\theta(x)\}/\operatorname{SE}_n(x)|),\epsilon_c\}-1] D n ( x ) ← [ − log max { 2Φ ( − ∣ { Y 0 , n ( x ) − θ ( x )} / SE n ( x ) ∣ ) , ϵ c } − 1 ] ;
R n A ( x ) ← 1 + σ ^ 0 , n 2 ( x ) exp { − D n ( x ) } / [ n ~ 0 , n e f f ( x ) SE n 2 ( x ) ] R_n^A(x)\leftarrow1+\widehat{\sigma}_{0,n}^2(x)\exp\{-D_n(x)\}/[\widetilde{n}_{0,n}^{\mathrm{eff}}(x)\operatorname{SE}_n^2(x)] R n A ( x ) ← 1 + σ 0 , n 2 ( x ) exp { − D n ( x )} / [ n 0 , n eff ( x ) SE n 2 ( x )] ;
Compute the treatment-arm kernel mass M 1 , n ( x ) M_{1,n}(x) M 1 , n ( x ) via K h K_h K h ;
ϕ n ( x ) ← { R n A ( x ) M 0 , n ( x ) } 2 / [ { R n A ( x ) M 0 , n ( x ) } 2 + { M 1 , n ( x ) } 2 ] \phi_n(x)\leftarrow\{R_n^A(x)M_{0,n}(x)\}^2/[\{R_n^A(x)M_{0,n}(x)\}^2+\{M_{1,n}(x)\}^2] ϕ n ( x ) ← { R n A ( x ) M 0 , n ( x ) } 2 / [{ R n A ( x ) M 0 , n ( x ) } 2 + { M 1 , n ( x ) } 2 ] ;
Draw Z n + 1 ∼ Bernoulli { ϕ n ( x ) } Z_{n+1}\sim\operatorname{Bernoulli}\{\phi_n(x)\} Z n + 1 ∼ Bernoulli { ϕ n ( x )} , observe Y n + 1 Y_{n+1} Y n + 1 ;
end
return { Z i } i = 1 N \{Z_i\}_{i=1}^N { Z i } i = 1 N and { R n A ( X n + 1 ) } n = 0 N − 1 \{R_n^A(X_{n+1})\}_{n=0}^{N-1} { R n A ( X n + 1 ) } n = 0 N − 1 .
The algorithm uses the abbreviations D N , i = D N ( X i ) D_{N,i}=D_N(X_i) D N , i = D N ( X i ) , D N , i F = D N F ( X i ) D^{F}_{N,i}=D^{F}_N(X_i) D N , i F = D N F ( X i ) , S E N , i 2 = S E N 2 ( X i ) \mathrm{SE}^{2}_{N,i}=\mathrm{SE}^{2}_N(X_i) SE N , i 2 = SE N 2 ( X i ) , R N , i F = R N F ( X i ) R^{F}_{N,i}=R^{F}_N(X_i) R N , i F = R N F ( X i ) , and W N , i F = W N F ( X i ) W^{F}_{N,i}=W^{F}_N(X_i) W N , i F = W N F ( X i ) when evaluating these quantities at X i X_i X i .
Algorithm S3: Stage III: Final estimation and inference for the treatment effect Δ N \Delta_N Δ N .
Input: Trial data { ( X i , Y i , Z i ) } i = 1 N \{(X_i,Y_i,Z_i)\}_{i=1}^{N} {( X i , Y i , Z i ) } i = 1 N ; historical control data D H \mathcal{D}_H D H ; Stage I outputs { θ ( x ) , τ ^ H 2 ( x ) } \{\theta(x),\widehat{\tau}_H^2(x)\} { θ ( x ) , τ H 2 ( x )} ; kernel function K h K_h K h with bandwidth h h h ; safeguards n min = 5 n_{\min}=5 n m i n = 5 , ϵ c = ϵ v = 10 − 12 \epsilon_c=\epsilon_v=10^{-12} ϵ c = ϵ v = 1 0 − 12 ; significance level α = 0.05 \alpha=0.05 α = 0.05 .
Output: Point estimate Δ ^ N \widehat{\Delta}_N Δ N , variance estimate V ^ N \widehat{V}_N V N , 100 ( 1 − α ) % 100(1-\alpha)\% 100 ( 1 − α ) % Wald interval.
Compute the marginal score D N G D_N^G D N G from the completed concurrent and historical control means and variances;
for i = 1 , … , N i=1,\ldots,N i = 1 , … , N do
Form kernel rows S 1 ( X i ) , S 0 ( X i ) , S H ( X i ) S_1(X_i),S_0(X_i),S_H(X_i) S 1 ( X i ) , S 0 ( X i ) , S H ( X i ) and row-normalize;
Compute final n 0 , N e f f ( X i ) n^{\mathrm{eff}}_{0,N}(X_i) n 0 , N eff ( X i ) , Y ‾ 0 , N ( X i ) \overline{Y}_{0,N}(X_i) Y 0 , N ( X i ) , and σ ^ 0 , N 2 ( X i ) \widehat{\sigma}^{2}_{0,N}(X_i) σ 0 , N 2 ( X i ) from all concurrent controls;
n ~ 0 , N e f f ( X i ) ← max { n 0 , N e f f ( X i ) , n min } \widetilde{n}^{\mathrm{eff}}_{0,N}(X_i)\leftarrow\max\{n^{\mathrm{eff}}_{0,N}(X_i),n_{\min}\} n 0 , N eff ( X i ) ← max { n 0 , N eff ( X i ) , n m i n } ; σ ^ 0 , N 2 ( X i ) ← max { σ ^ 0 , N 2 ( X i ) , ϵ v } \widehat{\sigma}^{2}_{0,N}(X_i)\leftarrow\max\{\widehat{\sigma}^{2}_{0,N}(X_i),\epsilon_v\} σ 0 , N 2 ( X i ) ← max { σ 0 , N 2 ( X i ) , ϵ v } ;
S E N , i 2 ← max { τ ^ H 2 ( X i ) + σ ^ 0 , N 2 ( X i ) / n ~ 0 , N e f f ( X i ) , ϵ v } \mathrm{SE}^{2}_{N,i}\leftarrow\max\left\{\widehat{\tau}^{2}_H(X_i)+\widehat{\sigma}^{2}_{0,N}(X_i)/\widetilde{n}^{\mathrm{eff}}_{0,N}(X_i),\epsilon_v\right\} SE N , i 2 ← max { τ H 2 ( X i ) + σ 0 , N 2 ( X i ) / n 0 , N eff ( X i ) , ϵ v } ;
D N , i ← [ − log max { 2 Φ ( − ∣ { Y ‾ 0 , N ( X i ) − θ ( X i ) } / S E N , i ∣ ) , ϵ c } − 1 ] + D_{N,i}\leftarrow\left[-\log\max\left\{2\Phi\left(-\left|\{\overline{Y}_{0,N}(X_i)-\theta(X_i)\}/\mathrm{SE}_{N,i}\right|\right),\epsilon_c\right\}-1\right]_+ D N , i ← [ − log max { 2Φ ( − { Y 0 , N ( X i ) − θ ( X i )} / SE N , i ) , ϵ c } − 1 ] + ;
D N , i F ← D N , i + { 1 − exp [ − D N , i ] } D N G D^F_{N,i}\leftarrow D_{N,i}+\{1-\exp[-D_{N,i}]\}D_N^G D N , i F ← D N , i + { 1 − exp [ − D N , i ]} D N G ;
R N , i F ← 1 + exp ( − D N , i F ) σ ^ 0 , N 2 ( X i ) / [ n ~ 0 , N e f f ( X i ) S E N , i 2 ] R^F_{N,i}\leftarrow1+\exp(-D^F_{N,i})\widehat{\sigma}^{2}_{0,N}(X_i)/[\widetilde{n}^{\mathrm{eff}}_{0,N}(X_i)\mathrm{SE}^{2}_{N,i}] R N , i F ← 1 + exp ( − D N , i F ) σ 0 , N 2 ( X i ) / [ n 0 , N eff ( X i ) SE N , i 2 ] ; W N , i F ← ( R N , i F − 1 ) / R N , i F W^F_{N,i}\leftarrow(R^F_{N,i}-1)/R^F_{N,i} W N , i F ← ( R N , i F − 1 ) / R N , i F ;
S ~ 0 ( X i ) ← { 1 − W N , i F } S 0 ( X i ) \widetilde{S}_0(X_i)\leftarrow\{1-W^F_{N,i}\}S_0(X_i) S 0 ( X i ) ← { 1 − W N , i F } S 0 ( X i ) ; S ~ H ( X i ) ← W N , i F S H ( X i ) \widetilde{S}_H(X_i)\leftarrow W^F_{N,i}S_H(X_i) S H ( X i ) ← W N , i F S H ( X i ) ;
end
a 1 ← N − 1 ∑ i S 1 ( X i ) a_1\leftarrow N^{-1}\sum_i S_1(X_i) a 1 ← N − 1 ∑ i S 1 ( X i ) ; a 0 ← N − 1 ∑ i S ~ 0 ( X i ) a_0\leftarrow N^{-1}\sum_i\widetilde{S}_0(X_i) a 0 ← N − 1 ∑ i S 0 ( X i ) ; a H ← N − 1 ∑ i S ~ H ( X i ) a_H\leftarrow N^{-1}\sum_i\widetilde{S}_H(X_i) a H ← N − 1 ∑ i S H ( X i ) ;
Δ ^ N ← a 1 T Y ‾ 1 − a 0 T Y ‾ 0 − a H T Y ‾ H \widehat{\Delta}_N\leftarrow a_1^\mathsf{T}\overline{Y}_1-a_0^\mathsf{T}\overline{Y}_0-a_H^\mathsf{T}\overline{Y}_H Δ N ← a 1 T Y 1 − a 0 T Y 0 − a H T Y H ;
For z = 0 , 1 z=0,1 z = 0 , 1 , set n z ← N z n_z\leftarrow N_z n z ← N z ; for z = H z=H z = H , set ( n z , X z , Y z ) ← ( N H , X H , Y H ) (n_z,X_z,Y_z)\leftarrow(N_H,X_H,Y_H) ( n z , X z , Y z ) ← ( N H , X H , Y H ) ;
for z ∈ { 0 , 1 , H } z\in\{0,1,H\} z ∈ { 0 , 1 , H } do
S z z ← normRows [ K h ( X z , X z ) ] S_{zz}\leftarrow\operatorname{normRows}[K_h(X_z,X_z)] S z z ← normRows [ K h ( X z , X z )] ;
dof z ← max { n z − 2 tr ( S z z ) + tr ( S z z T S z z ) , 1 } \operatorname{dof}_z\leftarrow\max\{n_z-2\operatorname{tr}(S_{zz})+\operatorname{tr}(S_{zz}^{\mathsf{T}}S_{zz}),1\} dof z ← max { n z − 2 tr ( S z z ) + tr ( S z z T S z z ) , 1 } ;
σ ^ z 2 ← max { ∥ Y z − S z z Y z ∥ 2 2 / dof z , ϵ v } \widehat{\sigma}^{2}_z\leftarrow\max\{\lVert Y_z-S_{zz}Y_z\rVert_2^2/\operatorname{dof}_z,\epsilon_v\} σ z 2 ← max {∥ Y z − S z z Y z ∥ 2 2 / dof z , ϵ v } ;
end
V ^ l i n ← σ ^ 1 2 ∥ a 1 ∥ 2 2 + σ ^ 0 2 ∥ a 0 ∥ 2 2 + σ ^ H 2 ∥ a H ∥ 2 2 \widehat{V}_{\mathrm{lin}}\leftarrow\widehat{\sigma}^{2}_1\lVert a_1\rVert_2^2+\widehat{\sigma}^{2}_0\lVert a_0\rVert_2^2+\widehat{\sigma}^{2}_H\lVert a_H\rVert_2^2 V lin ← σ 1 2 ∥ a 1 ∥ 2 2 + σ 0 2 ∥ a 0 ∥ 2 2 + σ H 2 ∥ a H ∥ 2 2 ;
a 0 C ← N − 1 ∑ i S 0 ( X i ) a^C_0\leftarrow N^{-1}\sum_i S_0(X_i) a 0 C ← N − 1 ∑ i S 0 ( X i ) ; V ^ C ← σ ^ 1 2 ∥ a 1 ∥ 2 2 + σ ^ 0 2 ∥ a 0 C ∥ 2 2 \widehat{V}_C\leftarrow\widehat{\sigma}^{2}_1\lVert a_1\rVert_2^2+\widehat{\sigma}^{2}_0\lVert a^C_0\rVert_2^2 V C ← σ 1 2 ∥ a 1 ∥ 2 2 + σ 0 2 ∥ a 0 C ∥ 2 2 ;
Floor V ^ l i n \widehat{V}_{\mathrm{lin}} V lin and V ^ C \widehat{V}_C V C at ϵ v \epsilon_v ϵ v ;
V ^ N ← max { V ^ l i n , V ^ C } \widehat{V}_N\leftarrow\max\{\widehat{V}_{\mathrm{lin}},\widehat{V}_C\} V N ← max { V lin , V C } ;
s e ^ N ← V ^ N \widehat{\mathrm{se}}_N\leftarrow\sqrt{\widehat{V}_N} se N ← V N ;
return Δ ^ N , V ^ N , [ Δ ^ N ± z 1 − α / 2 s e ^ N ] \widehat{\Delta}_N,\widehat{V}_N,\left[\widehat{\Delta}_N\pm z_{1-\alpha/2}\widehat{\mathrm{se}}_N\right] Δ N , V N , [ Δ N ± z 1 − α /2 se N ] .