Bayesian Design and Analysis of Precision Trials with Partial Borrowing

Motivated by a gastric cancer trial, this paper proposes a Bayesian framework for precision clinical trials that utilizes individually weighted external data to partially borrow information for estimating subgroup effects, supported by a simulation study comparing its performance to dynamic borrowing and demonstrating its application in determining sample sizes and decision boundaries.

The Gist

Imagine you are a chef trying to perfect a new recipe for a very specific group of people: left-handed people who love spicy food.

You have a small group of test subjects in your kitchen (your Clinical Trial). But here's the problem: there are only 5 left-handed, spicy-food lovers in your kitchen. If you try to taste-test the dish with just these five people, your results will be shaky. You might think the dish is amazing just because you got lucky with those five, or you might think it's terrible because one of them had a bad day. You don't have enough data to be sure.

However, you know you have access to two other cookbooks:

1. The "Spicy Food" Cookbook: A huge book of recipes tested on thousands of people who love spicy food, but they are mostly right-handed.
2. The "Left-Handed" Cookbook: A small book of recipes tested on left-handed people, but they mostly eat mild food.

The Challenge:
You want to use the information from these two other books to help you judge your dish for your specific group (left-handed + spicy). But you can't just copy-paste the whole books.

• If you copy the "Spicy Food" book entirely, you might be ignoring the fact that your test group is left-handed.
• If you copy the "Left-Handed" book entirely, you might be ignoring the fact that they don't eat spicy food.

This is exactly the problem the paper solves. It proposes a smart way to "borrow" information from these external sources without getting confused or biased.

The Solution: The "Fit Score" System

The authors propose a method called Individually Weighted Borrowing. Here is how it works, using our cooking analogy:

1. The "Fit Score" (Similarity Measure)
Instead of treating every page in the external cookbooks as equally important, the authors suggest looking at every single person in those books and giving them a "Fit Score."

• Does this person in the external book look like the people in your kitchen? Do they have similar age, health status, and habits?
• If a person in the external book is very similar to your test group, they get a high score (a high weight). Their opinion counts a lot.
• If a person is very different (e.g., a right-handed person who hates spicy food), they get a low score. Their opinion counts very little.

2. The "Partial Borrowing" (The Mix)
You don't throw away the external books. Instead, you create a "smoothie" of information:

• You take a big spoonful of your own kitchen data (the 5 people).
• You take a tiny spoonful of the "Spicy Food" book, but only the parts that look like your left-handed people.
• You take a tiny spoonful of the "Left-Handed" book, but only the parts that look like your spicy-food lovers.

By doing this, you get a much clearer picture of how your dish will taste for your specific group, without being misled by people who are too different.

3. The "Cut-Off" Rule (Truncation)
Sometimes, the external books are massive (like a library of a million recipes). Even if most of those recipes get a low "Fit Score," if you add up a million low scores, they might drown out your 5 kitchen tests.
To fix this, the authors suggest a Cut-Off Rule: If a recipe's "Fit Score" is too low, we throw it in the trash. We only keep the external data that is "close enough" to our own group. This prevents the huge external books from overpowering your small, specific trial.

Why This Matters for Medicine

In the real world, this is about Precision Medicine.

• The Problem: Modern medicine tries to treat patients based on their specific genetics or history (subgroups). But these subgroups are often tiny. A trial might have 1,000 patients, but only 50 have a specific rare mutation. It's impossible to get a reliable answer with just 50 people.
• The Old Way: Doctors might ignore all other data (wasting valuable knowledge) or blindly combine all data (risking wrong conclusions because the groups are different).
• The New Way (This Paper): The authors provide a mathematical "recipe" to mix the new trial data with old, real-world data. They weigh the old data based on how similar the patients are.

The "Design" Part: Planning Ahead

The paper also talks about Design. Imagine you are planning the cooking competition before you start.

• "If I use this smart borrowing method, do I need 100 people in my kitchen, or can I get away with 50?"
• The authors show that by using this smart borrowing method, you can often run a smaller trial and still get a reliable answer. This saves time, money, and spares patients from unnecessary procedures.

The Bottom Line

Think of this paper as a smart filter.
When you have a small, specific group of patients you need to study, you can look at the massive amount of data from the past. But instead of blindly trusting it, you use a filter to let in only the data that looks like your specific group.

• Too different? Filter it out (Low weight).
• Very similar? Let it in (High weight).
• Result? A more accurate, faster, and cheaper way to figure out if a new medicine works for the specific people who need it most.

The authors tested this with computer simulations (like running a thousand virtual cooking competitions) and found that their "Fit Score" method was often more accurate and safer than other fancy, complex methods currently in use. It's a practical, simple tool to make medical trials smarter.

Technical Summary

Here is a detailed technical summary of the paper "Bayesian Design and Analysis of Precision Trials with Partial Borrowing" by Golchi and Morita.

1. Problem Statement

The advancement of precision medicine has increased the demand for clinical trials capable of investigating effect heterogeneity and estimating subgroup-specific effects. However, these trials often face a critical limitation: small sample sizes within specific subgroups (e.g., patients with a specific biomarker or disease status) lead to underpowered studies and imprecise estimates of treatment-covariate interactions.

To address this, researchers often seek to borrow information from external data sources (retrospective studies, early-phase trials, single-arm studies). However, a major challenge arises when these external sources differ from the current trial population in terms of covariate distributions or underlying treatment effects. Naive borrowing can introduce significant bias, while ignoring external data results in low statistical power. Existing methods, such as Power Priors and dynamic borrowing approaches (e.g., LEAP), often struggle to balance the need for precision with the risk of bias when data sources are discordant, particularly in the context of subgroup analysis where data is sparse.

2. Methodology

The authors propose a comprehensive framework comprising an individually weighted analysis model and a Bayesian design procedure.

A. Individually Weighted Analysis Model

Instead of applying a single global weight to an entire external dataset (study-level discounting), the authors propose an individual-level weighting scheme within a Power Prior framework.

• Model Structure: The analysis uses a generalized linear model (GLM) where the outcome depends on prognostic covariates and treatment-covariate interactions.
• Weighting Mechanism: Each patient in the external dataset is assigned a weight ($\omega_n$) based on their similarity to the target trial population.
  • Similarity Measure: The weights are derived using a posterior predictive similarity function. This function calculates the probability that an external patient's covariate vector fits the distribution of the internal trial data, using a similarity model $q$ (e.g., kernel density for continuous variables, multinomial for categorical).
  • Covariate Selection: Weights are calculated based on prognostic-only covariates (excluding effect modifiers) to avoid down-weighting external data that belongs to the sparse subgroups of interest.
• Truncation: To prevent large external datasets from dominating the inference (even with small weights), the authors propose truncating the weight distribution. A cutoff is applied such that the effective sample size (sum of weights) of the external data does not exceed that of the internal trial data.
• Prior Formulation: The analysis prior is constructed as:
  $$\pi_a(\theta) \propto \pi_0(\theta) \prod_{n=1}^{N_E} f(\theta; d_n)^{\omega_n}$$
  where $\pi_0$ is a weakly informative default prior and $f$ is the likelihood of the external data.

B. Bayesian Design Framework

The paper extends the methodology to the trial design phase (sample size and decision boundaries).

• Design Priors: The authors utilize external data to construct design priors (distinct from analysis priors) to reflect beliefs at the design stage.
• Handling Partial Information: Since external data may only inform subsets of model parameters (e.g., a single-arm study informs the control arm but not the interaction), the design priors are defined conditionally on the external data ($\pi(\theta | D_E)$) constrained by the null ($H_0$) and alternative ($H_1$) hypotheses regarding the marginal treatment effect ($\Gamma$).
• Operating Characteristics: The framework uses Bayesian Type I error and Power to determine sample sizes, employing efficient Monte Carlo integration and predictive modeling to handle computational complexity.

3. Key Contributions

1. Individualized Partial Borrowing: The proposal of an individually weighted Power Prior that allows for granular adjustment of external data contribution based on covariate fit, rather than a binary "borrow or not" decision.
2. Robustness to Discordance: The introduction of a truncation mechanism to mitigate bias when external datasets are large or significantly discordant with the trial population.
3. Integrated Design and Analysis: A unified framework that uses external data not just for analysis, but to construct design priors for sample size determination, specifically addressing scenarios where external data provides incomplete information (e.g., missing interaction terms).
4. Practical Implementation: The use of posterior predictive similarity functions (extending Page & Quintana, 2018) as a flexible, covariate-based similarity metric that avoids the complexity of fully dynamic Bayesian models while maintaining robustness.

4. Results

The authors evaluated the framework through simulation studies and a real-world application.

Simulation Study

• Scenarios: Simulations covered various levels of discordance between internal and external data (covariate distribution shifts and parameter shifts) and varying sample sizes.
• Comparison: The proposed Individually Weighted (IW) model was compared against Full Borrowing (FB), No Borrowing (NB), and the Latent Exchangeability Prior (LEAP) (a dynamic borrowing method).
• Findings:
  • Bias vs. Variance: Full borrowing minimized variance but introduced systematic bias under discordance. No borrowing was unbiased but had high variance.
  • Performance of IW: The IW model (and its truncated version, IW.t) achieved a superior balance, maintaining unbiased estimates (or minimal bias) in scenarios with covariate discordance while achieving lower variance than LEAP and No Borrowing.
  • Type I Error: Under the null hypothesis, IW and IW.t maintained Type I error rates closer to nominal levels compared to LEAP and Full Borrowing, which showed inflation when discordance existed.
  • Truncation: Truncation was shown to be critical when external sample sizes were large, preventing the external data from overwhelming the internal signal.

Application: XParTS-II Gastric Cancer Trial

• Context: A Phase II trial comparing S-1+Cisplatin vs. Capecitabine+Cisplatin in gastric cancer. The subgroup of interest (recurrent disease) constituted only 23% of the trial, leading to low power.
• Analysis: Re-analysis using the IW model with external data (from a Phase II trial and a retrospective study) resulted in a posterior distribution that was a "compromise" between the sparse internal data and the external data. It reduced uncertainty slightly but prevented the bias that would occur from naive full borrowing.
• Design: The framework demonstrated how external data could be used to construct design priors. While the actual external data was too small to drastically reduce the required sample size for the specific subgroup, the methodology showed that with larger synthetic external data, the required sample size could be reduced by approximately 50% to achieve the same power.

5. Significance

This paper provides a practical, robust, and computationally feasible solution for integrating external data into precision medicine trials.

• Addressing the "Sparse Subgroup" Problem: It offers a viable path to power subgroup analyses without requiring prohibitively large recruitment numbers, which is crucial for rare diseases and biomarker-driven trials.
• Bias Control: By moving from study-level to individual-level weighting and introducing truncation, the method effectively guards against the bias that plagues many external borrowing techniques when populations are not perfectly aligned.
• Simplicity vs. Flexibility: While acknowledging that fully dynamic Bayesian methods (like LEAP) are more elegant, the authors argue that their static, covariate-based approach offers a simpler implementation that performs comparably or better in terms of bias and Type I error control in realistic discordant scenarios.
• Design Utility: The integration of external data into the design phase (not just analysis) allows for more efficient trial planning, potentially saving resources and time in the development of precision therapies.