📊 statistics

NP-LEAP: Nonparametric Latent Exchangeability Prior for Model-Lean Borrowing from Historical Data

The paper proposes NP-LEAP, a nonparametric, outcome-agnostic Bayesian framework that dynamically borrows information from historical data by assessing individual-level exchangeability and averaging over data partitions, thereby avoiding the parametric misspecification risks inherent in existing methods, particularly for time-to-event outcomes.

Original authors: Ethan M. Alt, Miheer Dewaskar, Jacob M. Maronge, Yuelin Lu, Matthew A. Psioda

Published 2026-08-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ethan M. Alt, Miheer Dewaskar, Jacob M. Maronge, Yuelin Lu, Matthew A. Psioda

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of medical research, scientists often face a difficult puzzle: how to make sense of a new drug trial when the group of patients receiving the standard treatment is too small to give a clear answer. This is a common hurdle in cancer studies, where researchers might test a new therapy on a handful of people while relying on a much larger group of patients from a previous study to represent the standard of care. The challenge lies in deciding how much weight to give that older data. If the two groups of patients are too different, mixing them together can lead to false conclusions. If they are too similar, ignoring the older data wastes valuable information that could make the new study more precise. For decades, statisticians have tried to solve this by building rigid mathematical models that assume the patients behave in specific, predictable ways. But real human biology is rarely that simple, and when these rigid models are wrong, the results can be misleading.

A team of researchers has developed a new approach called the nonparametric latent exchangeability prior, or NP-LEAP, designed to navigate this uncertainty without forcing the data into a pre-made box. Instead of assuming the patients fit a specific pattern, this method lets the data speak for itself by examining each individual patient one by one. It asks a simple question for every person in the historical group: does this person's outcome look like it belongs with the current group, or does it stand apart? The method then uses a flexible, computer-driven process to sort the historical patients into two piles: those who are similar enough to the current study to be included, and those who are too different to be trusted. This sorting happens automatically, adjusting the influence of the old data based on how well it actually matches the new data, rather than relying on a fixed rule set by the researcher.

The researchers tested this new tool using a real-world scenario involving patients with non-small cell lung cancer, a common and serious form of the disease. They compared a small, current clinical trial with a much larger historical trial. In the current study, only thirty-four patients received the standard chemotherapy control, a number too small to draw strong conclusions on its own. The historical study, however, included three hundred and forty-three patients on a similar regimen. By applying their new method, the researchers were able to look at the survival times of every single patient in the historical group and decide, individually, whether to include them in the analysis. They found that the method successfully identified which parts of the old data were relevant and which were not, effectively filtering out the noise while keeping the signal.

When they compared the results of this new approach against older methods, the difference was clear. The traditional techniques, which rely on rigid mathematical assumptions, often produced biased results when the two groups of patients were not perfectly alike. In some cases, these older methods were so misled by the mismatch that they performed worse than simply ignoring the historical data entirely. The new method, by contrast, remained robust. It managed to borrow just the right amount of information from the past to sharpen the picture of the present, reducing the uncertainty around the treatment effect without introducing error. In the lung cancer example, this precision was enough to turn an inconclusive result into a clear finding, showing that the new treatment was likely less effective than the standard care at the twelve-month mark.

The power of this approach lies in its ability to handle complexity without breaking down. In many medical studies, the risk of death does not follow a straight line or a simple curve; it can change over time in unpredictable ways. Older methods often struggle with this, forcing the data into a shape that does not fit. The new method does not try to force the data into a shape. Instead, it builds a flexible map of the outcomes, allowing it to adapt to whatever pattern the patients actually show. This flexibility means that even when the historical data is only partially similar to the current data, the method can still find value in it. It does not discard the old information just because it is not a perfect match; it simply weighs the matching parts more heavily and the mismatched parts less.

The researchers also checked to see how sensitive their conclusions were to the amount of borrowing they allowed. They found that the results were stable across a wide range of scenarios. Even when they adjusted the strictness of the matching criteria, the core conclusion remained the same, giving them confidence that the finding was not an artifact of a specific setting. This stability is crucial for medical regulators, who need to know that a study's results are not just a fluke of the statistical method used. The method proved capable of handling the messy reality of clinical trials, where patient populations shift and outcomes vary, without losing its way.

Ultimately, this work offers a new way to think about learning from the past. It suggests that we do not need to choose between ignoring old data or blindly trusting it. Instead, we can use a tool that listens to the data itself, deciding on a case-by-case basis what is useful and what is not. For the field of clinical research, this means more reliable answers from smaller studies, faster development of life-saving treatments, and a deeper understanding of how new therapies truly perform. The method does not promise to solve every problem, but it provides a more honest and adaptable way to combine the wisdom of the past with the questions of the present.

Problem Statement
Bayesian dynamic borrowing (BDB) methods aim to reduce treatment effect uncertainty in current clinical trials by leveraging historical data. However, existing approaches almost universally rely on parametric outcome models. This reliance is problematic because borrowing under a misspecified model can yield biased inference even when populations are genuinely exchangeable. This issue is particularly acute for time-to-event outcomes, where parametric BDB is potentially "triply misspecified": it imposes a parametric baseline hazard, assumes a proportional hazards structure, and assumes blanket exchangeability of all historical observations, whereas in reality, only a subset may be exchangeable. Furthermore, standard group-level exchangeability arguments often require strong informative priors when the number of historical datasets is small and apply "blanket discounting," treating all historical individuals as equally relevant or irrelevant without individual-level assessment.

Methodology: NP-LEAP
The authors propose the Nonparametric Latent Exchangeability Prior (NP-LEAP), an outcome-agnostic, assumption-lean framework that borrows information directly in the space of outcome densities via Bayesian nonparametric modeling.

Model Structure: The NP-LEAP models the current data density $f$ and the historical data as a two-part mixture $\gamma f + (1-\gamma)g$ , where $g$ represents the non-exchangeable component. Crucially, both $f$ and $g$ are specified as Dirichlet Process Mixture Models (DPMMs).
Individual-Level Exchangeability: Unlike group-level methods, NP-LEAP introduces individual-level latent Bernoulli indicators $\epsilon_{0j}$ for each historical observation $j$ . If $\epsilon_{0j}=1$ , the observation is exchangeable with the current data (drawn from $f$ ); if $\epsilon_{0j}=0$ , it is non-exchangeable (drawn from $g$ ).
Bayesian Model Averaging (BMA): The framework induces Bayesian model averaging over all $2^{n_0}$ possible partitions of the historical data into exchangeable and non-exchangeable subsets. The weights of these partitions are driven by marginal likelihoods rather than subjective tuning parameters.
Clustering Scheme: The method utilizes a collapsed Gibbs sampler. Historical observations are assigned to clusters via a "rich-get-richer" mechanism. Exchangeable historical individuals join clusters shared with current data (increasing the attractiveness of those clusters), while non-exchangeable individuals are assigned to separate non-exchangeable clusters.
Covariate Adjustment: For time-to-event data, the authors employ an ANOVA Dependent Dirichlet Process (DDP) model, allowing the kernel parameters to depend on treatment arm and covariates.
Theoretical Guarantees: The paper establishes posterior consistency (Theorem 1) under mild regularity conditions. Specifically, if the historical sample size grows at the same rate as the current sample size ( $n_0/n \to \beta$ ), the posterior error in estimating the true current data density $f_0$ vanishes asymptotically, even when historical data are only partially exchangeable. The authors note that while the model is not fully identifiable regarding the non-exchangeable density $g$ or the exchangeability proportion $\gamma$ , the primary goal of estimating $f_0$ remains consistent.

Key Contributions

Nonparametric Flexibility: By using DPMMs, the NP-LEAP removes the need to specify a parametric family for the outcome density or a fixed number of mixture components. The Dirichlet process adapts automatically to the data complexity.
Individual-Level Borrowing: The framework performs individualized discounting, classifying each historical observation as exchangeable or not, rather than applying blanket discounting to an entire dataset.
Estimand-Agnostic Inference: By modeling the full outcome density, the method allows for the estimation of various functional summaries (e.g., median survival, restricted mean survival time, milestone probabilities) within a single unified model without re-specifying the model for each estimand.
Robustness to Misspecification: The approach avoids the "triply misspecified" pitfalls of parametric time-to-event borrowing by not imposing proportional hazards or parametric baseline hazards.

Results

Simulation Studies: The authors conducted simulations based on a non-small cell lung cancer (NSCLC) trial setting, generating data with time-varying treatment effects and varying degrees of exchangeability.
- Robustness: Parametric borrowing methods (e.g., Power Priors, Commensurate Priors) exhibited substantially elevated bias and Mean Squared Error (MSE) when exchangeability was partial or absent. In several scenarios, the MSE of parametric borrowing exceeded that of non-borrowing methods, confirming that borrowing under a misspecified model can be actively harmful.
- Efficiency: NP-LEAP largely avoided this pathology. Its bias was comparable to the frequentist Cox model, and its MSE was generally as good as or better than non-borrowing Bayesian nonparametric (BNP) methods, except in the most extreme non-exchangeability scenarios.
- Precision Gains: Under partial or full exchangeability, NP-LEAP achieved meaningful efficiency gains, yielding lower MSE and narrower credible intervals than non-borrowing approaches while maintaining approximately 95% coverage.
Real Data Application: The method was applied to augment the control arm of a Phase 2 NSCLC trial using data from a larger Phase 3 historical trial.
- Covariate Balance: Propensity score matching was used to balance baseline covariates.
- Survival Analysis: The NP-LEAP narrowed the 95% credible interval for the control arm survival probabilities and the treatment effect difference ( $\Delta(t)$ ) compared to the non-borrowing BNP approach.
- Inferential Impact: Without borrowing, the analysis yielded no statistically meaningful conclusion (the credible interval for the 12-month survival difference straddled zero). With NP-LEAP, the upper bound of the 95% credible interval fell below zero from 4.2 months onward, tipping the analysis from inconclusive to actionable.
- Sensitivity: A tipping point analysis showed that the conclusion (statistical significance) was robust for high degrees of assumed exchangeability ( $\gamma \ge 0.9$ ), while the lower bound of the credible interval remained stable across varying $\gamma$ .

Significance and Claims
The paper claims that NP-LEAP provides a flexible and principled foundation for information sharing across studies, demonstrating that dynamic borrowing need not be confined to parametric outcome models. By integrating nonparametric modeling with individualized discounting, the framework offers a solution to the tension between parametric borrowing and semiparametric practice, particularly in time-to-event settings. The authors highlight the method's computational tractability (analyzing a dataset in approximately five seconds) and its ability to resolve inferential uncertainty that current data alone cannot, as illustrated in the NSCLC application. The work addresses regulatory challenges identified in recent FDA guidance regarding the need for robustness to prior-data conflict and the specification of similarity measures, as NP-LEAP adapts automatically to the degree of exchangeability without requiring pre-specified discounting rates or mixture components.

More like this