Identifying Treatment Effect Heterogeneity with Bayesian Hierarchical Adjustable Random Partition in Adaptive Enrichment Trials

This paper introduces the Bayesian Hierarchical Adjustable Random Partition (BHARP) model, a self-contained framework that utilizes a finite mixture model and reversible-jump Markov chain Monte Carlo sampling to automatically identify treatment effect heterogeneity and adjust information borrowing in adaptive enrichment trials, thereby outperforming existing methods in accuracy and precision.

The Gist

Imagine you are a doctor trying to figure out which patients respond best to a new exercise program. You have a group of 10 different types of people (subgroups) based on their health history, weight, and relationships. Some might love the program, some might hate it, and some might not care at all.

The big challenge is: How do you know who belongs in which group without guessing?

If you treat everyone as exactly the same, you miss the nuances. If you treat everyone as completely different, you don't have enough data to make good guesses for the smaller groups. This is the problem of Treatment Effect Heterogeneity (TEH).

This paper introduces a new tool called BHARP (Bayesian Hierarchical Adjustable Random Partition) to solve this puzzle. Here is how it works, explained with simple analogies.

1. The Old Way: The "One-Size-Fits-All" vs. "Guess the Group" Problem

• The "All Together" Approach: Imagine putting all 10 subgroups into one big pot and stirring them together. You get an average result. This is efficient, but if Group A loves the exercise and Group B hates it, the average tells you nothing useful. It's like saying "the average temperature in the room is 70°F," when one person is freezing and another is sweating.
• The "Pick One Group" Approach: Some older methods try to sort the patients into clusters (e.g., "Group 1" and "Group 2") and then pick the single best way to sort them. The problem? There are thousands of ways to sort 10 people. If you pick just one way, you ignore the uncertainty. It's like a detective picking one suspect and ignoring all other possibilities, even if the evidence is shaky.

2. The BHARP Solution: The "Smart, Shapeshifting Organizer"

The BHARP model is like a super-intelligent, shapeshifting organizer who doesn't just pick one way to sort the patients. Instead, it tries many different ways to sort them simultaneously and weighs the results.

Here is the magic trick:

• The "Mixture" Concept: Imagine you have a bag of marbles of different colors (patients). You don't know how many distinct colors there are. BHARP asks: "What if there is 1 color? What if there are 2? What if there are 5?"
• The "Random Partition": Instead of forcing the marbles into a fixed number of boxes, BHARP lets the marbles float into boxes dynamically. It uses a special computer algorithm (called rjMCMC) that acts like a dance floor.
  • Sometimes, two groups of dancers merge into one big circle (because they are similar).
  • Sometimes, one big circle splits into two smaller circles (because they are different).
  • The algorithm dances back and forth, trying different formations, and keeps track of which formations make the most sense based on the data.

3. How It Handles Uncertainty (The "Voting System")

In the old methods, the computer would say, "I think there are 3 groups," and stop there.

BHARP says, "I think there's a 10% chance there are 2 groups, a 60% chance there are 3 groups, and a 30% chance there are 4 groups."

It then averages all these possibilities.

• Analogy: Imagine you are trying to guess the winner of a race.
  • Old Method: You pick one expert, ask them who will win, and bet on that.
  • BHARP: You ask 1,000 experts. Some say Runner A, some say Runner B. You don't just pick one; you look at the pattern of their votes. If 600 experts say Runner A is in a "fast group" and 400 say Runner B is in a "slow group," you get a much clearer, more reliable picture of the race dynamics.

4. Why This Matters for Medicine (The "Adaptive Enrichment" Trial)

The paper tests this in a "Adaptive Enrichment Trial." Imagine a clinical trial that is a live, changing experiment.

• The Scenario: You start with 10 subgroups. As the trial runs, you check the data every few months.
• The BHARP Advantage: Because BHARP is so good at figuring out who is similar to whom, it can tell the trial organizers:
  • "Stop giving the drug to Group 7; it's not working for them." (Saving money and protecting patients).
  • "Focus all your energy on Group 3; they are responding amazingly!" (Finding the cure faster).
  • "Group 4 and Group 5 are actually the same; let's treat them as one big group to get better data."

5. The "Speed" Factor

Usually, doing this kind of complex math takes forever. The authors built a custom computer engine (written in C++) that makes BHARP incredibly fast.

• Analogy: Other methods are like a librarian trying to sort books by hand, checking every single shelf one by one. BHARP is like a robot that can scan the whole library, sort the books, and re-sort them instantly to find the perfect arrangement, all while you wait for your coffee.

Summary: The Big Takeaway

BHARP is a smart, flexible tool that helps doctors figure out which treatments work for which specific types of people.

Instead of forcing patients into rigid boxes or guessing the best grouping, BHARP explores all possible groupings at once. It admits, "We aren't 100% sure how many groups there are, but here is the most likely pattern." This leads to:

1. More accurate results (less guessing).
2. Faster trials (stopping bad treatments early).
3. Better personalized medicine (finding the right treatment for the right person).

It turns a messy, confusing pile of patient data into a clear, actionable map for doctors.

Technical Summary

Here is a detailed technical summary of the paper "Identifying Treatment Effect Heterogeneity with Bayesian Hierarchical Adjustable Random Partition (BHARP) Model in Adaptive Enrichment Trials."

1. Problem Statement

In adaptive enrichment clinical trials, a primary challenge is identifying Treatment Effect Heterogeneity (TEH) across predefined subgroups. While Bayesian Hierarchical Models (BHMs) improve estimation efficiency by borrowing information across all subgroups, they often assume full exchangeability. This assumption can lead to excessive shrinkage, biasing subgroup-specific estimates and masking true heterogeneity.

Existing partition-based methods attempt to group subgroups with similar responses to enable borrowing within clusters while preserving heterogeneity between clusters. However, these methods typically:

• Select a single "optimal" partition based on a predefined criterion (e.g., DIC), ignoring model uncertainty regarding the borrowing structure.
• Fail to propagate the uncertainty of the partition choice to the final subgroup-level estimates.
• Struggle with the combinatorial explosion of possible partitions as the number of subgroups increases, making exhaustive enumeration infeasible.

2. Methodology: The BHARP Model

The authors propose the Bayesian Hierarchical Adjustable Random Partition (BHARP) model, a self-contained framework that treats the partition structure as a random quantity rather than a fixed input.

Core Model Formulation

• Finite Mixture Model (FMM): The model assumes subgroup-specific treatment effects ($\theta_k$) arise from a finite mixture with an unknown number of components ($q$).
• Latent Partition: A vector $z$ assigns each of the $K$ subgroups to one of $q$ mixture components. Subgroups assigned to the same component are treated as exchangeable (sharing information), while those in different components are not.
• Hierarchical Structure:
  • Within-component variance ($\sigma$): Modeled with an informative Inverse-Gamma prior to ensure that subgroups with clinically meaningful differences are assigned to different clusters.
  • Between-component precision ($\tau$): A weakly informative Gamma prior regulates the scale of component means without inducing excessive shrinkage.
  • Number of components ($q$): Assigned a prior $P(q) \propto q^\alpha$ (with $\alpha=2$ recommended), allowing the model capacity to adapt to the data.

Computational Approach: Reversible-Jump MCMC (rjMCMC)

To handle the varying dimensionality of the parameter space (as $q$ changes), the authors implement a custom reversible-jump Markov chain Monte Carlo (rjMCMC) algorithm:

• Transdimensional Moves: The sampler performs "split" and "merge" moves to traverse models with different numbers of components.
• Split-Merge Framework: Based on the Nobile and Green framework, proposals are constructed deterministically to maintain detailed balance.
• Efficiency: Implemented in C++ via Rcpp, the algorithm jointly updates the partition structure and model parameters, capturing uncertainty in both simultaneously.

Adaptive Trial Integration

The model is embedded within an adaptive enrichment trial design:

• Interim Analyses: Decisions to stop recruitment for futile subgroups or declare efficacy are based on posterior probabilities ($P(\theta_{ik} \le x_F | D_\ell)$ and $P(\theta_{ik} > x_E | D_\ell)$).
• Dynamic Allocation: The framework allows for the early termination of arms or subgroups, reallocating resources to promising arms with complex TEH structures.

3. Key Contributions

1. Formal Accounting of Model Uncertainty: Unlike methods that condition on a single partition, BHARP averages over the partition space via posterior sampling, formally integrating uncertainty about the borrowing structure into subgroup estimates.
2. Automated Partition Discovery: The model eliminates the need for manual calibration or pre-specification of the number of clusters. It automatically learns the optimal borrowing structure from the data.
3. Computational Efficiency: Despite the complexity of rjMCMC, the custom implementation is highly efficient, outperforming fixed-component FMMs (like BLAST) and non-parametric methods (like BART) in runtime, even for large-scale multi-arm trials.
4. Clinical Interpretability: The use of a Finite Mixture Model (rather than a Dirichlet Process) provides a more transparent balance between flexibility and interpretability, with hyperparameters that can be elicited based on clinically meaningful effect size differences.

4. Simulation Results

The authors evaluated BHARP against four comparators: Independent (IND), standard BHM, Fixed-component FMM (BLAST), and Bayesian Additive Regression Trees (BART) across 12 diverse TEH scenarios.

• Partition Recovery: BHARP successfully identified the true number of clusters in all scenarios. The posterior mode of $q$ closely matched the true number of clusters, whereas BLAST (fixed at 3 components) failed to adapt to scenarios with 1 or 2 clusters.
• Estimation Accuracy: BHARP achieved low Root Mean Squared Error (RMSE) and low Interquartile Range (IQR) across all scenarios.
  • In homogeneous scenarios, BHARP shrank towards a single cluster ($q=1$), outperforming BLAST which unnecessarily split the data.
  • In heterogeneous scenarios, BHARP maintained precision comparable to or better than BLAST.
• Co-clustering Probabilities: The model produced sharp block patterns in co-clustering probability heatmaps, accurately recovering true subgroup groupings even with small sample sizes or "border" subgroups.
• Computational Speed: BHARP was the fastest method among those capable of partitioning, requiring significantly less time than BLAST (which relies on DIC-based model selection) and BART.

5. Application: Partner Step T2D Trial

The model was applied to a hypothetical adaptive enrichment design based on the "Partner Step T2D" study (physical activity interventions for Type 2 diabetes dyads).

• Scenario: A 3-arm trial with 6 subgroups defined by relationship quality and weight concordance.
• Performance:
  • TEH Recovery: BHARP correctly identified a homogeneous cluster in Arm 1, a complex cross-paired structure in Arm 2, and two distinct clusters in Arm 3.
  • Decision Making: BHARP demonstrated superior generalized power at the subgroup level (0.44) compared to BHM (0.18) and IND (0.20), meaning it was much better at identifying effective interventions for specific subpopulations.
  • Resource Allocation: The model facilitated faster termination of futile arms (Arm 1) and effective arms (Arm 3), allowing more samples to be allocated to the complex Arm 2.
  • Uncertainty: BHARP consistently achieved the lowest IQR, indicating more precise posterior estimates even with smaller effective sample sizes due to early stopping.

6. Significance

The BHARP model represents a significant advancement in precision medicine and adaptive clinical trial design. By treating the information-borrowing structure as a latent variable to be inferred rather than a fixed assumption, it provides a robust, automated, and computationally efficient solution for identifying treatment effect heterogeneity. Its ability to handle model uncertainty and scale to multi-arm platform trials makes it particularly valuable for modern drug development where subgroup-specific efficacy is critical.