📊 statistics

Bayesian integration G-formula for platform SMART designs allowing for adding new treatments

This paper introduces a novel platform SMART design that allows for the addition of new treatments during the trial and proposes Bayesian integration G-formula (BIG) estimators to effectively handle non-concurrent treatment comparisons, with performance validated through simulations and a real-world application to the SNAP trial.

Original authors: Xinru Wang, Meghna Bose, Bibhas Chakraborty, Robert Mahar

Published 2026-04-29

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Xinru Wang, Meghna Bose, Bibhas Chakraborty, Robert Mahar

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

Imagine you are trying to figure out the best recipe for a complex, multi-course meal to treat a specific illness. In the old way of doing this (called a SMART trial), you would set up a rigid kitchen. You'd pick two main ingredients (Treatments A and B) to start with. If the dish tastes good after the first course, you might switch to a different sauce. If it tastes bad, you stick with the original. You cook this for a long time, taste-test, and finally decide which recipe is the "winner."

The problem is that this process takes a long time. While you are cooking, a brand-new, amazing ingredient (a new drug) might become available in the market. In a traditional kitchen, you'd have to stop, throw away your current setup, and start a whole new experiment to test this new ingredient. That's a waste of time and money.

The Solution: The "Platform Kitchen"
This paper introduces a new way of cooking called a Platform SMART. Imagine a kitchen that is designed to be flexible. You start with your original ingredients, but the kitchen has a special rule: if a new, promising ingredient arrives while you are still cooking, you can slide it into the recipe immediately without stopping the show.

However, this creates a tricky statistical problem. If you mix the data from the "old" cooks (who only knew the old ingredients) with the "new" cooks (who had the new ingredient), you might get a confused result. Maybe the new cooks were better just because the weather changed, or the ingredients were fresher, not because the new drug was actually better. This is called a time effect.

The New Tool: The "Bayesian Integration G-formula" (BIG)
The authors propose a new mathematical tool called BIG (Bayesian Integration G-formula) to solve this confusion. Think of BIG as a very smart, cautious sous-chef who helps you decide how much to trust the old recipes versus the new ones.

Instead of just ignoring the old data (which wastes information) or blindly mixing it all together (which risks bad results), BIG uses a "sliding scale" of trust:

The "Separate" Approach: The sous-chef says, "I only trust the new cooks. Let's ignore the old data completely." This is safe (no confusion), but you miss out on helpful tips from the past.
The "Pooling" Approach: The sous-chef says, "Let's mix all the data together!" This is efficient, but if the weather changed (time effects), the result might be biased and wrong.
The "BIG" Approach: This is the smart middle ground. The sous-chef looks at the old data and the new data.
- If the old cooks and new cooks seem to be making dishes that taste very similar, BIG says, "Great! Let's borrow heavily from the old data to make our new estimate stronger."
- If the old cooks and new cooks seem to be making very different dishes (maybe the new cooks are using a different technique or the ingredients changed), BIG says, "Okay, let's be careful. We'll only borrow a little bit of information from the old data, or maybe none at all, to avoid ruining the new result."

What the Paper Found
The authors tested this "smart sous-chef" (BIG) using computer simulations (virtual cooking trials) with different scenarios:

When nothing changes over time: If the kitchen conditions stay the same, simply mixing all the data (Pooling) works well. But BIG is also very good at this.
When things change over time: If the kitchen conditions change (like a new standard of care or a shift in patient types), the "mix everything" approach fails and gives wrong answers. The "ignore the old data" approach is safe but less precise.
The Winner: The BIG approach consistently found the "best recipe" (the optimal treatment strategy) more often than the others. It managed to be precise (using old data when safe) without being biased (ignoring old data when risky).

Real-World Example
The paper demonstrates this method using the SNAP trial, a real-world study for treating Staphylococcus aureus blood infections. In this trial, doctors can switch patients to oral antibiotics if they respond well, or keep them on IV antibiotics if they don't. The trial is designed to add new antibiotics as they become available. The authors show how their BIG method could be used to analyze this trial, ensuring that the addition of new drugs doesn't mess up the comparison of which treatment strategy is truly the best.

In Summary
This paper presents a new way to analyze clinical trials that allow new treatments to be added mid-stream. It offers a "smart" statistical method (BIG) that knows when to trust past data and when to be skeptical, helping doctors find the best treatment sequences faster and more accurately than before.

1. Problem Statement

Sequential Multiple Assignment Randomized Trials (SMARTs) are the gold standard for evaluating Dynamic Treatment Regimes (DTRs), which are sequences of decision rules tailored to a patient's evolving disease status. However, traditional SMARTs have two major limitations:

Duration: They often require significantly more time to complete than single-stage Randomized Controlled Trials (RCTs) due to multiple randomization stages.
Rigidity: New candidate treatments may become available (e.g., via regulatory approval) during the trial but cannot be incorporated without restarting or running a separate study.

Platform trials address the second issue by allowing new treatments to be added to an ongoing study under a master protocol. However, adding treatments creates non-concurrent cohorts (participants randomized before vs. after the addition). This introduces statistical challenges:

Time Effects: Changes in standard of care, population characteristics, or randomization probabilities over time can confound treatment effect estimates.
The Trade-off:
- Separate Approach: Using only data from the post-adaptation cohort ensures unbiasedness but wastes valuable data from the pre-adaptation cohort, reducing efficiency.
- Pooling Approach: Combining data from all cohorts increases efficiency but risks significant bias if time effects exist (violating exchangeability).

The paper addresses the gap in analytical methods for Platform SMARTs (SMARTs that allow adding new treatments), specifically focusing on how to estimate DTR means while balancing bias and efficiency.

2. Methodology: Bayesian Integration G-formula (BIG)

The authors propose a novel class of estimators called Bayesian Integration G-formula (BIG) estimators. These utilize Bayesian methods to "borrow" information from non-concurrent cohorts (pre-adaptation data) in a controlled manner.

A. Study Design Setup

Scenario: A two-stage SMART where an additional first-stage treatment ( $a_{13}$ ) is added mid-trial.
Cohorts:
- $C=c_1$ : Pre-adaptation cohort (original treatments only).
- $C=c_2$ : Post-adaptation cohort (original + new treatment).
Goal: Estimate the mean outcome ( $\mu_{jk}$ ) for embedded DTRs in the post-adaptation cohort ( $c_2$ ) while leveraging data from $c_1$ .

B. The BIG Estimator Framework

The method models the response rate ( $R$ ) and the final outcome ( $Y$ ) using Generalized Linear Models (GLMs).

Model Specification:
- Response model: $\text{logit}(\pi) = f_R(A_1; \beta)$ .
- Outcome model: $Y \sim f_Y(A_1, R, A_2; \psi) + \epsilon$ .
Parameter Partitioning:
- $\theta_s$ : Coefficients shared between cohorts (e.g., effects of original treatments). These can borrow information from $c_1$ .
- $\theta_{ns}$ : Coefficients specific to the new treatment ( $a_{13}$ ). These have no prior data from $c_1$ .
Prior Distributions:
The core innovation lies in the priors assigned to $\theta_s$ $θ_{s}$ based on data from $c_1$ $c_{1}$ :
- BIGweak: Uses weakly informative priors (e.g., $N(0, 100)$ ) for all parameters. Equivalent to ignoring $c_1$ data for borrowing.
- BIGlogdis (Log-distance Prior): Sets the prior mean based on the Maximum Likelihood Estimate (MLE) from $c_1$ . The prior variance is adaptive, defined by the maximum of the squared difference between cohort estimates and the variance of the $c_1$ estimate. This automatically reduces borrowing if the cohorts differ significantly.
- BIGcomP (Commensurate Prior): Introduces a commensurability parameter $\tau$ . High $\tau$ implies high similarity (more borrowing); low $\tau$ implies dissimilarity (less borrowing).
- BIGcommP (Mixed Commensurate Prior): Uses a mixture of fixed $\tau$ values (e.g., one low, one high) with equal weights to allow for "partial borrowing" without needing to estimate $\tau$ directly.

C. Estimation Algorithm

The authors use a Bayesian G-formula algorithm:

Generate posterior draws of model parameters ( $\theta$ ) using data from $c_2$ and the specified priors derived from $c_1$ .
Simulate a large population ( $N$ ) under specific DTRs using these posterior draws.
Calculate the mean outcome for each draw to approximate the posterior distribution of the DTR mean ( $\mu_{jk}$ ).
The final estimate is the mean of these posterior draws.

3. Key Contributions

Novel Design Integration: First study to propose analytical methods specifically for Platform SMARTs, combining the adaptive nature of platform trials with the DTR evaluation goals of SMARTs.
BIG Estimators: Development of a flexible Bayesian framework that allows for dynamic borrowing of non-concurrent data. Unlike standard pooling, it accounts for potential time effects through adaptive priors.
Prior Flexibility: Comparison of multiple prior strategies (Log-distance, Commensurate, Mixed Commensurate) to handle varying degrees of heterogeneity between cohorts.
Real-World Application: Demonstration of the method using the S. aureus Network Adaptive Platform (SNAP) trial, a complex adaptive platform trial for bloodstream infections.

4. Simulation Results

The authors conducted extensive simulations across five scenarios (varying time effects, optimal DTRs, and sample sizes) comparing:

Separate: Uses only $c_2$ data.
Pooling: Pools $c_1$ and $c_2$ data.
BIG variants: Using different priors.

Key Findings:

No Time Effects (Scenarios 1–3): The Pooling approach performed best (lowest variance/MSE). However, BIGlogdis, BIGcomP, and BIGcommP performed nearly as well as pooling while maintaining robustness.
With Time Effects (Scenarios 4–5):
- Pooling failed significantly, producing high bias and low coverage rates due to the violation of exchangeability.
- Separate and BIGweak remained unbiased but had higher variance (lower efficiency).
- BIG (Log-distance/Commensurate) successfully balanced the trade-off: they achieved lower bias than Pooling and lower variance than Separate/BigWeak. They effectively "down-weighted" the pre-adaptation data when time effects were detected.
Optimal DTR Identification: The BIG approaches consistently outperformed the Separate and BIGweak approaches in identifying the true optimal DTR across all scenarios.

5. Significance and Conclusion

Efficiency vs. Validity: The BIG estimators provide a "sweet spot" between the inefficiency of ignoring historical data and the bias of blindly pooling it. This is crucial for platform trials where time effects are inevitable.
Clinical Impact: By enabling the addition of new treatments without restarting trials and providing statistically valid methods to analyze the resulting complex data, this methodology accelerates the development of optimal treatment sequences for chronic and complex diseases.
Future Directions: The authors note that while they focused on a single added treatment and linear models, future work should extend this to multiple added treatments at different stages and more flexible modeling approaches (e.g., machine learning) to handle model misspecification.

In summary, the paper establishes a rigorous statistical framework for Platform SMARTs, proving that Bayesian integration via the G-formula allows researchers to leverage non-concurrent data efficiently without compromising the validity of Dynamic Treatment Regime estimation.