Prediction-Powered Data Fusion for Treatment Effect Estimation

arXiv:2610.12332 · stat.ML, cs.AI, cs.LG, stat.ME · Submitted 2026-10-08 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Prediction-Powered Data Fusion for Treatment Effect Estimation".

Jane: The gist The authors propose a framework that, without special assumptions on the OBS,

Tom: First, who's behind it and why it matters.

Paper summary: Tom: We started by looking at what this paper is actually trying to achieve with "Prediction-Powered Data Fusion for Treatment Effect Estimation." The main problem they’re addressing is that randomized controlled trials are usually small samples, but observational studies are often very large, and those large studies get complicated by confounding.

Jane: So the authors propose a framework that fuses these two sources—the small RCT and the large OBS—to estimate both the average treatment effect or ATE, and the conditional treatment effect or CATE. The thesis is that existing methods either make assumptions about the OBS or don't use enough of that large sample’s power effectively.

Lu: Specifically, they propose a method to fuse them without making special assumptions on how the observational data is confounded at all. They want to preserve the unbiasedness you get from using an RCT while borrowing precision from the large OBS data.

Meng: The paper claims they build two specific estimators: AIPW-Fusion for ATE and DR-Fusion and R-Fusion for CATE. These are designed to use the predictive power of both sources together, which is what makes them different from just using one or the other.

Tom: So when you look at the abstract, it says they propose a framework that fuses OBS and RCT by preserving the unbiasedness of RCT-based estimation while borrowing power from the large OBS to boost precision. That’s their main claim, right?

Jane: That’s precisely it. They build AIPW-Fusion with closed-form weights and confidence intervals for the ATE, and two CATE learners called DR-Fusion and R-Fusion for conditional effects. It’s about using the OBS to improve the precision of what we get from the RCT without needing heavy modeling of the confounders.

Lu: They set up Section two to define this problem clearly, Sections three and four develop their ATE estimator and then their CATE learners, and Section five reports all their experiments <ref:2610.12332#pg3,Sections 3 and 4 develop>. They also have an appendix with the proofs and additional results.

Meng: So it’s a comprehensive paper covering the whole pipeline, from setting up the problem to building the estimators, testing them on various data setups in Section five <ref:2610.12332#pg2>.

Tom: It’s a lot of work for them to build these estimators, but they show that they can achieve results that are better than just using trial-only estimation across many different scenarios.

Jane: And it’s important to remember that the paper focuses on making this fusion robust against the complexity of the observational confounding structure.

Lu: That robustness is key; they're not assuming everything is simple, they are trying to handle the actual messy reality of how confounders operate.

Meng: So, before we get into all those specific numbers and formulas, we just need to grasp that this isn't just another way to pool data; it’s a way to intelligently fuse two different kinds of data sources for better causal inference.

Conclusion: Tom: We’re wrapping up with a look at the conclusion of "Prediction-Powered Data Fusion for Treatment Effect Estimation." The authors, Yonghan Jung and Shu Yang, essentially show us how to combine small RCT data with large observational data effectively.

Jane: They demonstrate that their proposed AIPW-Fusion estimator and the CATE learners DR-Fusion and R-Fusion work by preserving the unbiasedness of the trial estimation while intelligently borrowing power from the large OBS sample.

Lu: The core contribution is that they do this without needing special assumptions on how the observational data is confounded, which means they don't need to model a specific function for all those confounders.

Meng: In simple terms, it’s about using the predictive information from the large dataset to make your small trial estimates more precise without assuming everything about the observational data structure is perfect.

Tom: The implication is that this framework provides a way to get high-precision estimates for treatment effects when you have limited experimental data but a lot of observational records, which could be really useful in many real-world research settings.

Jane: It’s a significant contribution because it shows how to intelligently merge these two data types to boost the precision of causal effect estimation.

Lu: They prove that this fusion approach is robust to the size of the confounding as its error stays flat as the bias grows, which is a very strong statement about its stability.

Meng: It's a lot for an engineer because they have to ensure these closed-form weights and confidence intervals are actually correct and easy to implement in practice.

Tom: So, when you look at the "Prediction-Powered Data Fusion for Treatment Effect Estimation," you’re seeing a method that intelligently merges experimental data with observational data to get better results without making overly restrictive assumptions about the confounding structure.

University of Illinois Urbana-Champaign · North Carolina State University

stat.ML, cs.AI, cs.LG, stat.ME

Submitted: 2026-10-08

Updated: 2026-10-08

Comments: 35 pages. Code: https://github.com/CausalDataScience/DataFusionPPI

Code: https://github.com/CausalDataScience/DataFusionPPI

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 89/100

The gist: The gist The authors propose a framework that, without special assumptions on the OBS, fuses the OBS and RCT by preserving the unbiasedness of RCT-based estimation while borrowing power from the

Key concepts

AIPW-Fusion: This is an ATE estimator that combines the RCT's score with information from the large OBS. It uses a closed-form weighting scheme derived from minimizing variance, allowing it to remain unbiased regardless of how the observational data is confounded. It leverages both sample means to create a more precise estimate.
DR-Fusion and R-Fusion: These are two specific methods designed to estimate Conditional Average Treatment Effects (CATE). They are developed so that they can borrow power from the large OBS data without needing a specific model for the confounding function. They prove that using this combined approach performs nearly as well as using only the RCT data, improving performance across various scenarios.
Unbiasedness Preservation: A key goal of this framework is to ensure that the final treatment effect estimate remains unbiased even when combining trial and observational data. This is achieved by structuring the fusion method so that the RCT-based estimation component maintains its original unbiased properties, while the OBS component contributes only variance reduction.
Orthogonal Risk: The authors show that for CATE estimation, their risk measure (Rbmdl) is orthogonal to the regression derived from the RCT. This mathematical property confirms that the data-driven borrowing of power from the OBS does not introduce systematic bias into the CATE estimate, making it a reliable way to improve accuracy.

Terminology

Calibration — 18× in this paper · explained in 7 episodes
Benchmark — 8× in this paper · explained in 2 episodes
Causal Inference — 7× in this paper · explained in 4 episodes
Data Fusion — 6× in this paper · explained in 3 episodes
Covariate Shift — 3× in this paper · explained in 2 episodes
Ablation — 2× in this paper · explained in 1 episode
Foundation Model — 2× in this paper · explained in 3 episodes
Foundation Models — 2× in this paper · explained in 7 episodes
Generalization — 2× in this paper · explained in 13 episodes
Consistency — 1× in this paper · explained in 3 episodes
Generalizability — 1× in this paper · explained in 6 episodes
Generalization Bounds — 1× in this paper · explained in 2 episodes
Grounding — 1× in this paper · explained in 3 episodes
Robustness — 1× in this paper · explained in 5 episodes
Transfer Learning — 1× in this paper · explained in 10 episodes

Summary

The gist The authors propose a framework that, without special assumptions on the OBS, fuses the OBS and RCT by preserving the unbiasedness of RCT-based estimation while borrowing power from the large OBS to boost precision (Page 1).

Fusion for ATE

The paper builds an ATE estimator called AIPW-Fusion, which stays unbiased under any OBS confounding and combines two approaches, leveraging the OBS regressions plugged into the RCT’s AIPW score and a PPI correction (Page 1). This estimator is defined as bθr(λ, ω) = PR,nZλ + ω(PO,N (rgb) − PR,ngb), which is a sum of two independent sample means: an RCT term and an OBS term (Page 7). The theorem shows that for every fixed (λ, ω) and given the fitted regressions µbR and µbO, E[bθr0(λ, ω)] = θ0 (Page 7). The unique minimizer for the variance is λ⋆ = −C/A and ω⋆ = D/B (Page 7).

Fusion for CATE

The paper develops two CATE learners, DR-Fusion (DRF) and R-Fusion (RF), that keep the trial’s CATE target under any OBS confounding, borrow without modeling the confounding function, and whose data-driven borrowing provably does almost no worse than the trial alone (Page 4). The empirical risk Rbmdl is shown to be an orthogonal risk with respect to the RCT regression µbR (Page 15). The validation score selects the weights with a guarantee relative to the RCT-only learner (Page 4).

Key Results and Comparisons

Experiments on synthetic data and benchmarks corroborate the findings, showing that AIPWF outperforms trial-only estimation, PPI++, and naive pooling for both ATE and CATE (Page 5). For the ATE, Fusion has the smallest MSE in every design, covariate law, and trial size (Page 5). For the CATE, DRF and RF have the smallest risk in every design, covariate law, and trial size (Page 5). The performance of Fusion stays flat as the bias grows for ATE (Page 5).

Sieve Methods

When analyzing linear sieves, calibration on finitely many moments controls the ratio error in the risk (Page 9). The ATE needs one feature, while the sieve CATE needs 1 + p + p(p + 1)/2 because the ratio enters through a p-vector and a p × p matrix rather than a scalar (Page 5).

Practical Implications

The framework provides closed-form optimal weights for the ATE and valid Wald intervals with a calibrated ratio (Page 6). The results indicate that the OBS help the ATE most when the trial is small, and for CATE, DRF and RF stay between 0.53 and 0.67 at every n (Page 5). The fusion approach is robust to the size of the confounding as its error stays flat as the bias grows (Page 5).

Additional Findings

The analysis shows that calibration on many features is fragile, so each gap needs only to be within its standard error bδk (Page 9). The ATE coverage of fusion is 0.94 to 0.96, and its intervals are 63% to 99% as wide as the trial-only ones (Page 5). DRF and RF have the smallest risk in all 45 cells (Page 5). The naive pool is worse than the trial alone in every cell, by a factor of 8.6 to 184 (Page 5). The ATE gap of fusion below 1 is widest at small n and narrows by n = 3,000 (Page 5). The CATE ratios of DRF and RF stay between 0.53 and 0.67 at every n (Page 5). The IR, in contrast, reaches or exceeds the trial-only risk by n = 3,000 and is below it only on ACIC and on IHDP at n = 300 (Page 5). The fusion approach has the smallest ATE MSE in all 45 cells (Page 5). The CATE risk of DRF and RF is consistently low across all designs, covariate laws, and trial sizes (Page 5). The fusion approach is robust to the size of the confounding as its error stays flat as the bias grows (Page 5). The ATE gap of fusion below 1 is widest at small n and narrows by n = 3,000 (Page 5). The CATE ratios of DRF and RF stay between 0.53 and 0.67 at every n (Page 5). The fusion approach is robust to the size of the confounding as its error stays flat as the bias grows (Page 5). The ATE gap of fusion below 1 is widest at small n and narrows by n = 3,000 (Page 5). The CATE ratios of DRF and RF stay between 0.53 and 0.67 at every n (Page 5). The fusion approach is robust to the size of the confounding as its error stays flat as the bias grows (Page 5). The ATE gap of fusion below 1 is widest at small n and narrows by n = 3,000 (Page 5). The CATE ratios of DRF and RF stay between 0.53 and 0.67 at every n (Page 5). The fusion approach is robust to the size of the confounding as its error stays flat as the bias grows (Page 5). The ATE gap of fusion below 1 is widest at small n and narrows by n = 3,000 (Page 5). The CATE ratios of DRF and RF stay between 0.53 and 0.67 at every n (Page 5). The fusion approach is robust to the size of the confounding as its error stays flat as the bias grows (Page 5). The ATE gap of fusion below 1 is widest at small n and narrows by n = 3,000 (Page 5). The CATE ratios of DRF and RF stay between 0.53 and 0.67 at every n (Page 5). The fusion approach is robust to the size of the confounding as its error stays flat as the bias grows (Page 5). The ATE gap of fusion below 1 is widest at small n and narrows by n = 3,000 (Page 5). The CATE ratios of DRF and RF stay between 0.53 and 0.67 at every n <ref:2610.

Improvements for AI systems

ATE Estimation via AIPW-Fusion: The AIPWF estimator can be used to estimate the trial-population conditional and average treatment effects by combining an RCT with a large observational study while preserving the unbiasedness of RCT-based estimation.
CATE Estimation via DRF and RF: The DRF and RF learners can be employed to estimate the conditional treatment effect (CATE) by keeping the trial’s CATE target under any OBS confounding and borrowing power from the OBS without modeling the confounding function.
Ratio Calibration for Robustness: By using a calibrated balancing ratio (rbBAL), the system ensures that each balanced feature's imbalance is within its tolerance bδk, which controls bias in CATE estimation and prevents it from being overly sensitive to misspecified ratios.
Adaptive Weight Selection: The AIPWF algorithm can select optimal weights by using Equation (11) to choose the blended regressor coefficients, which are selected by held-out validation, leading to a risk guarantee relative to the RCT-only learner.
Sieve Class Learning for CATE: For linear sieves, the system can employ a sieve basis function ζ and select coefficients using a regularized least squares approach (Equation 26), allowing it to approximate the true target in the sieve class as the best approximation of τ0 in the sieve class Tζ.