GEAR: On optimal decision making with auxiliary data

Hengrui Cai, Rui Song, Wenbin Lu

Abstract

Personalized optimal decision making, finding the optimal decision rule (ODR) based on individual characteristics, has attracted increasing attention recently in many fields, such as education, economics, and medicine. Current ODR methods usually require the primary outcome of interest in samples for assessing treatment effects, namely the experimental sample. However, in many studies, treatments may have a long-term effect, and as such the primary outcome of interest cannot be observed in the experimental sample due to the limited duration of experiments, which makes the estimation of ODR impossible. This paper is inspired to address this challenge by making use of an auxiliary sample to facilitate the estimation of ODR in the experimental sample. We propose an auGmented inverse propensity weighted Experimental and Auxiliary sample-based decision Rule (GEAR) by maximizing the augmented inverse propensity weighted value estimator over a class of decision rules using the experimental sample, with the primary outcome being imputed based on the auxiliary sample. The asymptotic properties of the proposed GEAR estimators and their associated value estimators are established. Simulation studies are conducted to demonstrate its empirical validity with a real AIDS application.

Introduction

In fields like medicine, economics, and education, researchers often seek to identify an Optimal Decision Rule (ODR): a personalized strategy that assigns a treatment ($A$) based on individual characteristics ($X$) to maximize a specific long-term outcome ($Y$).

The "gold standard" for finding an ODR is an experimental sample, such as a randomized controlled trial. However, these trials are often short-lived. While researchers can observe baseline characteristics and intermediate "surrogate" markers ($M$)—like blood pressure or CD4 counts—the primary outcome of interest (e.g., 5-year survival) may not manifest before the trial concludes. Conversely, "auxiliary" datasets (like administrative records or electronic health records) often contain these long-term outcomes but lack rigorous treatment information or randomization.

GEAR (auGmented inverse propensity weighted Experimental and Auxiliary sample-based decision Rule) addresses this gap by combining a trial that has treatments but no long-term outcomes with an auxiliary dataset that has outcomes but no treatment info.

The GEAR Framework and Assumptions

GEAR is a semi-supervised learning approach for causal inference that relies on two key "bridge" assumptions to link the experimental and auxiliary samples:

  1. Comparability: The relationship between covariates, surrogates, and the final outcome ($Y | X, M$) is the same across both the trial and the auxiliary data.
  2. Surrogacy: Once baseline characteristics and intermediate markers are accounted for, the treatment itself provides no additional information about the long-term outcome. In other words, the treatment effect is fully mediated by the surrogates.
Causal diagram illustrating the relationships between covariates (X), treatment (A), intermediate markers (M), and the primary outcome (Y).

As illustrated in the causal diagram above, the surrogacy assumption is represented by the absence of a direct arrow from treatment $A$ to outcome $Y$. This structure allows the model to "impute" the missing outcomes in the trial data by leveraging the relationships learned from the auxiliary data.

The GEAR Algorithm

The GEAR framework employs a multi-step process to derive the optimal rule:

  • Imputation: It uses the auxiliary sample to build a model that predicts the long-term outcome ($Y$) based on covariates ($X$) and intermediate markers ($M$).
  • AIPW Estimator: It employs an Augmented Inverse Propensity Weighted (AIPW) estimator. This is a "doubly robust" technique that models both the probability of receiving a treatment (propensity score) and the outcome itself. If either of these models is correctly specified, the resulting decision rule remains consistent.
  • Optimization: The rule is found by maximizing this value estimator over a class of possible decision rules, such as linear rules or flexible B-splines.

Empirical Results

The authors evaluated GEAR using synthetic scenarios and a real-world application involving the ACTG 175 (AIDS Clinical Trials Group) dataset. This study followed HIV-infected patients randomized to different antiretroviral regimens.

  • Simulation Success: In non-linear scenarios, a flexible GEAR-Bspline approach significantly outperformed standard linear models, maintaining high coverage probabilities (90%+) for the estimated value. The results showed that as the auxiliary sample size increases, the variance of the estimator decreases, effectively "borrowing strength" from the unlabeled data.
  • Real-World Application: On the AIDS data, GEAR identified a personalized rule for assigning patients to treatment regimens. The GEAR-Bspline method achieved a predicted mean CD4 count of 346.4, which is higher than the results achieved by simply assigning every patient the same "best" average treatment.

Conclusion and Future Work

GEAR provides a robust framework for decision-making when primary outcomes are delayed or missing from experimental data. By leveraging auxiliary data through a doubly robust estimator, it offers a path toward more accurate personalized medicine and policy.

Future research will look to extend GEAR beyond binary treatments to multiple or continuous options, and address "dynamic" treatment regimes where decisions are made sequentially over time. While the surrogacy and comparability assumptions are mathematically necessary, practitioners must continue to use domain expertise to validate these structures in real-world applications.

Summary updated

Abstract

Personalized optimal decision making, finding the optimal decision rule (ODR) based on individual characteristics, has attracted increasing attention recently in many fields, such as education, economics, and medicine. Current ODR methods usually require the primary outcome of interest in samples for assessing treatment effects, namely, the experimental sample. However, in many studies, treatments may have a long‐term effect, and as such, the primary outcome of interest cannot be observed in the experimental sample due to the limited duration of experiments, which makes the estimation of ODR impossible. This paper is inspired to address this challenge by making use of an auxiliary sample to facilitate the estimation of ODR in the experimental sample. We propose an auGmented inverse propensity weighted Experimental and Auxiliary sample‐based decision Rule (GEAR) by maximizing the augmented inverse propensity weighted value estimator over a class of decision rules using the experimental sample, with the primary outcome being imputed based on the auxiliary sample. The asymptotic properties of the proposed GEAR estimators and their associated value estimators are established. Simulation studies are conducted to demonstrate its empirical validity with a real AIDS application.
BibTeX
@inproceedings{cai2021gear,
  author = {Hengrui Cai and Rui Song and Wenbin Lu},
  title = {{GEAR: On optimal decision making with auxiliary data}},
  booktitle = {Stat},
  year = {2021},
  doi = {10.1002/sta4.399},
}
Figure