Chair: Taylor Krajewski (Duke University)
Abstract: This session highlights recent Bayesian methodological innovations designed to improve the efficiency, robustness, and interpretability of clinical trial analyses. The speakers will present advances in Bayesian dynamic borrowing of external data, including unified frameworks for comparing borrowing strategies, robust covariate-adjusted borrowing via latent exchangeability priors and Gaussian processes, and inverse probability weighted mixture priors for estimating marginal treatment effects under population heterogeneity. The session also features flexible Bayesian approaches for handling missing covariate data using nonparametric Dirichlet process models. Together, these talks demonstrate how modern Bayesian methods can address key challenges in clinical trials, such as limited sample sizes, external data integration, covariate imbalance, and missingness, while supporting principled uncertainty quantification and reliable decision-making in complex trial settings.
Speaker: Xinxin Chen, PhD Candidate (UNC – Chapel Hill)
Title: A Practical Introduction to Bayesian Dynamic Borrowing with External Data
Abstract: Bayesian methods provide a flexible framework for integrating information from various sources, including expert opinion and historical studies. Dynamic borrowing techniques, such as the power prior, commensurate prior, and meta-analytic approach, offer principled ways to leverage external evidence while adapting to differences between data sources. Although these methods have been well studied, practical guidance on their implementations remains comparatively sparse. In particular, analysts who wish to compare multiple borrowing strategies often find that each approach requires a distinct modeling specification and software workflow, making direct comparison challenging. In this talk, we will review several widely used Bayesian information borrowing approaches and demonstrate how they can be implemented in a unified and user-friendly way using the R package hdbayes, which supports a wide range of information borrowing priors for both generalized linear models and survival models. Through real clinical trial examples, we illustrate how dynamic borrowing can improve efficiency and enhance robustness when incorporating external data.
Speaker: Claire Zhu, PhD Candidate ( UNC – Chapel Hill)
Title: Bayesian Dynamic Borrowing with Robust Covariate Adjustment via the Latent Exchangeability Prior and Gaussian Processes
Abstract: In randomized controlled trials (RCTs), two popular approaches for increasing efficiency are via (i) robust covariate adjustment and (ii) informative prior elicitation. While the recent FDA guidance recommends the use of prognostic baseline covariates to improve statistical efficiency for estimating and testing treatment effects, it fails to mention how to do so while also leveraging historical data. On the other hand, traditional historical data borrowing approaches typically assume that the current and historical data models are correctly specified, which is unlikely to occur in practice. We propose a hierarchical Bayesian model that uses a finite mixture of Gaussian processes (GP) to both account for heterogeneity and adjust for covariates. Using a latent exchangeability prior (LEAP), our approach averages over all possible partitions of the historical data into exchangeable and nonexchangeable groups. Meanwhile, the Gaussian process flexibly provides covariate adjustments for the outcome of interest and can be particularly useful in cases where the relationship between predictors and outcomes is non-linear. By addressing heterogeneity and accounting for complex covariate-outcome relationships, our method is particularly suited for trials in rare diseases and other challenging settings with limited sample sizes.
Speaker: Anil Anderson, PhD Candidate ( UNC – Chapel Hill)
Title: A Dirichlet Process Model for Missing at Random Covariates
Abstract: Missing covariate values can occur in clinical trials for several reasons. Measurements may be difficult or expensive to obtain, subjects may miss visits at the clinic, or respondents may decline to answer items in a questionnaire. The default approach to handle missing data is a complete-case analysis, which is appealing in its simplicity but typically yields biased and inefficient inference. Parametric modeling of the missing covariates—for example, assuming continuous covariates follow a multivariate normal distribution—can improve bias and efficiency when correctly specified, but it can produce worse inference than a complete-case analysis when the assumptions are violated. We introduce a Dirichlet process model for missing-at-random covariates that relaxes these parametric assumptions, resulting in more robust estimation of the regression coefficients. In our method, the response conditional on the covariates is modeled as a generalized linear model, and the covariates are modeled marginally using a Dirichlet process with a parametric distribution as the base measure. We thus maintain interpretability in the response model while gaining flexibility in the covariate model. We show that the proposed method improves estimation of the regression coefficients relative to a complete-case analysis and misspecified parametric approaches, and we apply the methodology to a real clinical trial with missing covariates.
Speaker: Matthew Psioda, PhD (Glaxo Smith Kline)
Title: Inverse Probability Weighted Bayesian Dynamic Borrowing for Estimation of Marginal Treatment Effects
Abstract: We propose an approach for constructing and evaluating the performance of inverse probability weighted robust mixture priors (IPW-RMP) which are applied to the parameters in treatment group-specific marginal outcome models – which may be non-parametric. Our framework allows practitioners to systematically study the robustness of Bayesian dynamic borrowing using the IPW-RMP to enhance the efficiency of inferences on marginal treatment effects (e.g., marginal risk difference) in a target study being planned. A key assumption motivating our work is that the data generation processes for the target study and external data source (e.g., historical study) will not be the same, at minimum having different distributions for key prognostic factors and possibly different outcome distributions even for individuals who have identical prognostic factors (e.g., different outcome model parameters). We demonstrate the approach using simulation studies based on both binary and time-to-event outcomes, and via a case study based on actual clinical trial data for a solid tumor cancer program. Our simulation results show that when the distribution of risk factors does in fact differ, the IPW-RMP provides improved performance compared to a standard RMP (e.g., increased power and reduced bias of the posterior mean point estimator) with essentially no loss of performance when the risk factor distributions do not differ. Thus, the IPW-RMP can safely be used in any situation where a standard RMP is appropriate.
Co-authors: Matthew A. Psiodaa, Nathan W. Beana, Brielle A. Wright, Angel Lu, Alejandro Mantero and Antara Majumdar
