A Hybrid Prior Bayesian Method for Combining Domestic Real-World Data and Overseas Data in Global Drug Development

Jumin Deng; Keer Chen; Nan Li; Pengfei Zhu; Pingyan Chen; Shuping Jiang; Ying Wu; Zengyue Zheng; Zhenyu Wu

A Hybrid Prior Bayesian Method for Combining Domestic Real-World Data and Overseas Data in Global Drug Development

T0 review · 4 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read EQPS-rMAP maintains accuracy while borrowing external real-world data.

desk verdict A plausible incremental Bayesian borrowing method whose headline efficiency claims rest on unverified data-dependent weight selection and thin simulation reporting. read the letter →

arxiv 2505.12308 v1 pith:RTDMOZZB submitted 2025-05-18 stat.ME math.STstat.TH

Keer Chen , Zengyue Zheng , Pengfei Zhu , Shuping Jiang , Nan Li , Jumin Deng , Pingyan Chen , Zhenyu Wu

show 1 more author

Ying Wu

This is my paper · ORCID

classification stat.MEmath.STstat.TH MSC 62F1562P10

keywords hybridclinicaltrialdesignreal-worlddatameta-analyticpredictivepriorpropensityscorestratificationadaptiveborrowingequivalenceprobabilityweightbridgingtrialsrisankizumab

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper aims to establish that a hybrid Bayesian prior, EQPS-rMAP, can combine overseas randomized trial data with domestic real-world data in bridging and multi-regional trials without the bias typical of fixed-borrowing methods. It does this in three steps: stratify all patients by propensity score, build meta-analytic predictive priors inside each stratum, and set the borrowing weight through an equivalence probability that measures conflict between external and current data. The method is designed so that compatible evidence is borrowed heavily, while incompatible evidence is effectively shut off. If the simulations and case analysis are right, this gives trialists an adaptive, pre-specifiable way to reduce sample size in new regions without giving up estimation accuracy.

What carries the argument

The machinery is the data-dependent prior weight $\omega_r^*$ inside a robust MAP mixture. The weight is computed from the equivalence-probability consistency metric $p = P(\theta_{\text{post}}-\delta < \theta_{\text{current}} < \theta_{\text{post}}+\delta)$, with $\omega_r^*$ set to the smallest value satisfying $p \ge \lambda$ and to 1 when no such value exists. Before that, propensity-score stratification trims subjects whose propensity scores fall outside the current trial's range and divides the rest into strata, and stratum-specific MAP priors carry between-source heterogeneity through half-normal variance parameters whose scales are informed by the overlap coefficient of the propensity-score distributions. The weight is updated from the current trial's own data after seeing the current outcomes, which is what lets the borrowing proportion adapt, but also what makes the final operating characteristics depend on the data twice.

What would settle it

Simulate the full procedure repeatedly under the null hypothesis, computing $\omega_r^*$ by Eq. (2-19) on each replicate and testing $P(\theta_T>\theta_C)>0.95$ as the decision rule; if the proportion of false successes exceeds 5% by more than simulation error, the claimed error control does not hold in that scenario.

Watch

Extended reading notes

Core claim

The central claim is that baseline discrepancies and multi-source heterogeneity can be handled together by making the borrowing weight a function of measured data conflict. Within each propensity-score stratum, the external and real-world sources enter as stratum-specific robust MAP priors, and the final EQPS-rMAP posterior is a mixture of that informative component and a vague prior. The weight of the vague component, $\omega_r^*$, is the smallest value such that the mixed posterior's response probability stays within a clinical equivalence margin $\delta$ of the current-trial response distribution with probability at least $\lambda$. If agreement is poor, the weight goes to one, turning off borrowing entirely; if agreement is good, the method uses the maximal safe amount of external information. The paper reports that this keeps bias and mean squared error low across six scenarios and in a risankizumab psoriasis case study, while reducing the sample size the current trial would otherwise need.

Load-bearing premise

The claimed bias and sample-size advantages assume that selecting the vague-prior weight from the current trial's own data and then using that same data in the final posterior preserves the nominal frequentist type I error, even though the weight itself is random and data-dependent.

Editorial extensions

If this is right

In bridging and multi-regional trials, investigators can pre-specify $\lambda$ and $\delta$ and let the data decide how much foreign or real-world information to borrow, instead of committing to a fixed proportion.
A new region that wants to run a smaller trial can quantify how much sample size it saves under EQPS-rMAP when external data are compatible, because the posterior precision rises with the borrowed information.
The propensity-score stratification means baseline differences between domestic real-world patients and overseas trial patients are removed before the prior is built, so the method does not require exchangeability across sources.
The risankizumab case shows the posterior estimate staying near the true treatment effect as the proportion of borrowed data varies, whereas standard MAP and PS-MAP estimates drift toward the external data.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Because the same current data both selects $\omega_r^*$ and enters the final posterior, the effective type I error could diverge from 5% in settings not covered by the reported scenarios; a calibration study sweeping sample size, endpoint type, and $\lambda$/ $\delta$ would be the natural next check.
The weighting scheme could be ported to non-binary endpoints by replacing the beta-binomial mixture with conjugate normal or gamma mixtures; the paper states this as future work but does not implement it.
Regulatory use would likely require pre-specifying $\lambda$ and $\delta$ before unblinding; the paper demonstrates the trade-off in simulations but does not supply a default combination that guarantees operating characteristics.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The paper does something real: it combines propensity-score stratification with stratum-specific robust MAP priors and an equivalence-probability rule for setting the vague-prior weight. That combination is not in the cited literature, and it targets a genuine regulatory need for bridging trials and MRCTs where local RWD and overseas trial data differ in baseline and heterogeneity. The method is clearly described, and the authors cite the relevant building blocks honestly.

What the paper does well: the three-stage structure is sensible, the overlap-coefficient idea for stratum-specific similarity is reasonable, and the authors are upfront in Section 5 that lambda and delta need systematic calibration and that the framework currently only handles qualitative efficacy comparisons. The case study is transparent about using simulated current data.

The soft spots are in proportion. The big one is the circularity the stress-test flags: omega_r is chosen from the current trial data via Eqs 2-18 and 2-19, and the same data then enter the final posterior in Eq 2-20. No calibration argument or theorem shows this double use preserves frequentist type I error. The paper itself punts on operationalizing lambda and delta, saying it will be examined in subsequent simulation trials. That is an honest admission, but it undercuts the abstract's claim that the method 'reduces sample size demands' while controlling error. Sample-size savings are exactly where data-dependent borrowing can inflate type I error, and six simulated scenarios with user-chosen thresholds and no Monte Carlo standard errors do not establish the operating characteristics.

Second, the simulation reporting is too thin. Results are presented only through figures, with no numeric tables, no MCSE, and no confidence intervals for bias, MSE, or type I error. Third, the data accessibility statement says R code is available for Section 4, but no repository link appears in the text. Fourth, the abstract and conclusion call the case study a 'retrospective case analysis' when it is actually a synthetic illustration using real external data and simulated current data. That overstatement should be fixed.

None of this makes the method invalid. The central idea is plausible, and the paper identifies a real gap. But the evidence presented does not support the strength of the claims. I would not cite this as a validated method yet, and I would not use it in a regulatory setting without a calibration study.

Who is this for? Statisticians in pharma or regulatory science working on Bayesian borrowing methods. They will get a clear description of a workable extension and a list of open questions worth addressing.

Recommendation: send it to peer review, but require major revision: add a code repository, report numeric simulation results with Monte Carlo errors, provide a calibration rule or sensitivity analysis for lambda and delta, re-label the case study honestly, and address the data-dependent weight selection directly. As it stands, the paper is a promising methods note rather than a validated method.

Referee Report

4 major / 4 minor

Summary. The manuscript proposes EQPS-rMAP, a three-stage Bayesian hybrid method that combines domestic real-world data and overseas external trial data with a new-region randomized controlled trial. Stage 1 uses propensity score trimming and stratification to address baseline covariate imbalance; Stage 2 builds stratum-specific robust meta-analytic predictive (MAP) priors with heterogeneity variances scaled by propensity-score overlap; Stage 3 introduces an equivalence-probability weight omega_r, defined as the smallest vague-prior weight such that the probability that the hybrid posterior lies within a margin delta of the current-data response distribution reaches a threshold lambda. The authors claim that this adaptive borrowing preserves estimation robustness under heterogeneity and reduces required sample sizes. The method is evaluated with six simulation scenarios comparing EQPS-rMAP with MAP, PS-MAP, and EB-rMAP, and with an illustrative risankizumab psoriasis case analysis in which external and real-world data are real but the current-trial data are simulated. The central claim is that EQPS-rMAP resolves baseline-heterogeneity conflicts while maintaining type I error control and estimation accuracy.

Significance. If the claims were fully supported, the method would be a practically useful contribution to hybrid trial design, bridging studies, and multi-regional clinical trials, where regulatory interest in Bayesian borrowing from external data is high. The paper deserves credit for a clear conceptual structure, for combining propensity-score stratification with stratum-specific MAP priors, and for attempting a realistic case analysis; the authors also state that R code for Section 4 is available. However, the validation evidence is currently insufficient in two load-bearing respects: the borrowing weight is chosen from the current outcome data that are then used again in the final posterior, and the simulation results are reported only graphically, without numeric summaries or Monte Carlo standard errors. The significance of the method can be assessed only after these operating characteristics are quantified.

major comments (4)

[Section 2.3, Eqs (2-18)-(2-20)] The vague-prior weight omega_r is a function of the current trial data: Eq (2-18) computes p by comparing the hybrid posterior with the Beta posterior of the current data, and Eq (2-19) selects omega_r as the smallest weight with p >= lambda. The same current data are then used again in the final posterior, Eq (2-20). No theorem, calibration argument, or extensive simulation demonstrates that this double use of the data preserves the nominal 5% type I error. This issue is load-bearing because the claimed sample-size savings are achieved precisely by allowing omega_r < 1; the paper's own statement in Section 2.3 that operationalization of lambda and delta 'will be systematically examined in subsequent simulation trials' does not resolve the concern.
[Section 3.2, Figures 7 and 8] The central performance comparison is presented only graphically, with no numeric table of absolute bias, mean squared error, or type I error and no Monte Carlo standard errors. The abstract's claim that EQPS-rMAP 'maintains estimation robustness under significant heterogeneity' and the conclusion that it 'effectively manages Type I error' cannot be quantitatively assessed from the current figures. The authors should report point estimates with Monte Carlo standard errors for all six scenarios, and ideally across the full 54-scenario grid described in Section 3.1.
[Section 3.1 and Section 5] The parameters lambda and delta are user-specified tuning parameters, and the comparisons in Section 3.1 use a single pair (lambda = 0.8, delta = 0.1) without a sensitivity analysis. The paper itself states in Section 5 that 'systematic simulations are required to identify optimal parameter combinations.' Until a calibration rule or sensitivity results are provided, the claimed superiority over EB-rMAP, MAP, and PS-MAP remains conditional on unexamined choices of lambda and delta.
[Section 4] The 'current trial data' in the illustrative example are simulated with prespecified response rates (40% control, 65% treatment) and are generated using baseline characteristics from the external trial data. The case analysis therefore does not validate the method on observed current-trial outcomes, and the statement in the conclusion that 'case analyses confirm superior external bias control and accuracy' overstates what this example can show. The example should be described as a feasibility illustration based on a simulated current trial.

minor comments (4)

[Section 3.1] The text refers to 'EQPS-MAP (lambda = 0.8, delta = 0.1)' while the abstract and the rest of the paper define the method as EQPS-rMAP; the terminology should be standardized.
[Throughout] There are repeated typographical errors, including 'External trail data' instead of 'External trial data' and 'Jefferys' prior' instead of 'Jeffreys' prior'; a careful proofread is needed.
[Equation (2-14)] Equation (2-14) is typeset in a way that makes the integral and the density arguments difficult to parse; it should be rewritten with standard integral notation and clearly defined variables.
[Figures 3-6] The figures would be much easier to evaluate if they included numeric axis labels and a legend identifying the curves for the different parameter values; currently several figure descriptions refer to lines and columns that are not explicitly labeled.

Circularity Check

1 steps flagged · score 6.0 of 10

Data-dependent selection of omega_r from current trial data (Eqs. 2-18/2-19) before using that same data in the final posterior (Eq. 2-20) makes the claimed robustness and sample-size savings reduce by construction.

fitted input called prediction [Section 2.3, Step 3 (Eqs. 2-18, 2-19) and Step 4 (Eq. 2-20)]
"the response probability distribution for the Current trial data is modeled as: theta_current ~ Beta(r_current, n_current - r_current). ... To assess the agreement between the hybrid posterior and the current data, we define the following probability: p = Pr(theta_mix - delta < theta_current < theta_mix + delta) #(2-18) ... The optimal vague prior weight omega_r is then defined as: omega_r = min{omega: p >= lambda}, if exists omega s.t. p >= lambda; 1, otherwise #(2-19) ... For binary endpoints, the posterior distribution is ..."

The vague-prior weight omega_r is selected from the same current-trial data that are then used in the final posterior. Eq. (2-18) computes p by comparing the current-trial-only distribution theta_current ~ Beta(r_current, n_current - r_current) with theta_mix, the hybrid posterior that has already been updated with r_current and n_current. Eq. (2-19) then sets omega_r to the smallest weight making that agreement reach lambda. Eq. (2-20) reuses r_current and n_current in the final posterior, so the borrowing proportion is a function of the outcome being estimated. The headline 'reducing sample size demands' (Figures 5-6) is therefore a restatement of this data-adaptive fit rather than an independent prediction.

full rationale

The non-circular components of the paper are standard and externally grounded: propensity-score stratification, the MAP/rMAP mixture construction, and the simulation comparisons to MAP, PS-MAP, and EB-rMAP do not by themselves reduce to the paper's conclusions. The circularity is confined to the novel weight-selection step. Eq. (2-19) makes omega_r depend on the current trial outcome through p in Eq. (2-18), and Eq. (2-20) then uses that same outcome in the final posterior. Consequently, any claim that EQPS-rMAP 'reduces sample size demands' is essentially a consequence of the algorithm's permission to shrink omega_r when current and external data agree; the claimed type I error control is not derived from first principles but only illustrated for six simulated scenarios with user-chosen lambda and delta. This is partial circularity: the borrowing rule and the headline efficiency gain are the same fitted quantity, although the method still contains independent content in its propensity-score stratification and MAP prior construction. Score 6 reflects this construction-level circularity without alleging that the entire method is empty.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The framework rests on a chain of modeling choices: correctly specified propensity scores, within-stratum exchangeability, independent Beta approximations for conflict assessment, and user-set thresholds. These are all standard in the Bayesian borrowing literature, but none is derived or independently verified in the paper. No new physical entities are introduced.

free parameters (4)

lambda (consistency threshold) = 0.8 (used in the main simulation comparison)
Controls when external data borrowing is activated in Eq 2-19. User-specified, with no principled calibration rule given.
delta (equivalence margin) = 0.1 (used in the main simulation comparison)
Defines the acceptable distance between the hybrid posterior and the current data in Eq 2-18. User-specified and context-dependent.
Number of propensity score strata (S) = 5 (typical)
Stratification granularity affects bias-variance behavior. The paper uses quartiles and says the number is typically five, but does not provide a rule for choosing it.
Half-normal prior scale for heterogeneity variances = not specified
The hyperparameters for tau_ext,s and tau_rwd,s are described as reflecting stratum-specific similarity but their numerical values are not given in the manuscript.

assumptions (5)

domain assumption The propensity score model is correctly specified and includes all confounding covariates.
Step 1 (Section 2.3) uses propensity score trimming and stratification to remove baseline discrepancies; unmeasured confounders would leave residual imbalance and invalidate the comparability of external and current data.
domain assumption Within each propensity-score stratum, the log-odds of response for external, real-world, and current data are exchangeable under the hierarchical model in Eq 2-11.
The stratum-specific MAP prior assumes a common distribution for stratum means across data sources; if exchangeability fails, the prior is misspecified and the borrowing weights are not valid.
domain assumption The consistency probability p in Eq 2-18, computed using independent Beta approximations for theta_ext,s, theta_rwd,s, and theta_cur,s, is a valid measure of prior-data conflict.
The independence assumption ignores correlations among parameters within the hierarchical model; the paper uses this approximation to select the borrowing weight.
domain assumption A 95% posterior probability threshold provides frequentist type I error control at 5%.
The paper states this equivalence but only checks it by simulation; no formal calibration proof is given, and the data-dependent choice of omega_r could alter the operating characteristics.
domain assumption Binomial outcome with a logit link adequately represents response rates in each stratum.
The method is developed for binary endpoints in Section 2.3, and the paper explicitly says extension to survival and other complex endpoints is future work.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Hybrid Prior Bayesian Method for Combining Domestic Real-World Data and Overseas Data in Global Drug Development." pith.science (2026). https://pith.science/paper/RTDMOZZB

@misc{pith2026250512308,
  author       = {Pith},
  title        = {Pith review of: A Hybrid Prior Bayesian Method for Combining Domestic Real-World Data and Overseas Data in Global Drug Development},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RTDMOZZB}},
  note         = {Machine review of arXiv:2505.12308}
}

read the original abstract

Background Hybrid clinical trial design integrates randomized controlled trials (RCTs) with real-world data (RWD) to enhance efficiency through dynamic incorporation of external data. Existing methods like the Meta-Analytic Predictive Prior (MAP) inadequately control data heterogeneity, adjust baseline discrepancies, or optimize dynamic borrowing proportions, introducing bias and limiting applications in bridging trials and multi-regional clinical trials (MRCTs). Objective This study proposes a novel hybrid Bayesian framework (EQPS-rMAP) to address heterogeneity and bias in multi-source data integration, validated through simulations and retrospective case analyses of risankizumab's efficacy in moderate-to-severe plaque psoriasis. Design and Methods EQPS-rMAP eliminates baseline covariate discrepancies via propensity score stratification, constructs stratum-specific MAP priors to dynamically adjust external data weights, and introduces equivalence probability weights to quantify data conflict risks. Performance was evaluated across six simulated scenarios (heterogeneity differences, baseline shifts) and real-world case analyses, comparing it with traditional methods (MAP, PSMAP, EBMAP) on estimation bias, type I error control, and sample size requirements. Results Simulations show EQPS-rMAP maintains estimation robustness under significant heterogeneity while reducing sample size demands and enhancing trial efficiency. Case analyses confirm superior external bias control and accuracy compared to conventional approaches. Conclusion and Significance EQPS-rMAP provides empirical evidence for hybrid clinical designs. By resolving baseline-heterogeneity conflicts through adaptive mechanisms, it enables reliable integration of external and real-world data in bridging trials, MRCTs, and post-marketing studies, broadening applicability without compromising rigor.

Discussion (0). Continue with ORCID to comment.

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.