Conditional Front-door Adjustment for Heterogeneous Treatment Assignment Effect Estimation Under Non-adherence
Winston Chen, Trenton Chang, Jenna Wiens · 2025
Read Conditional Front-door Adjustment for Heterogeneous Treatment Assignment Effect Estimation Under Non-adherence by Winston Chen, Trenton Chang, Jenna Wiens (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-05677.txt.
Introduction
Winston Chen, Trenton Chang, and Jenna Wiens study how to estimate treatment effects when patients stray from their assigned care, in a 2025 paper. Such estimates guide real treatment choices, and straying is common in practice. Two classic corrections can both recover honest averages, but their precision may differ.
The authors prove, and then confirm in tests, that the front-door correction gives tighter estimates when the assigned treatment changes outcomes only slightly. Because that method needs several helper models, they build one shared neural network, LobsterNet, to fit them jointly. Across semi-real and real datasets it cut estimation error versus rival methods. The work helps analysts pick the steadier tool when adherence is shaky.
Catalog record
| CREATOR | Winston Chen, Trenton Chang, Jenna Wiens |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 172 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2505-05677
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Conditional Front-door Adjustment for Heterogeneous Treatment Assignment Effect Estimation Under Non-adherence Winston Chen, Trenton Chang, Jenna Wiens Abstract Estimates of heterogeneous treatment assignment effects can inform treatment decisions. Under the presence of non-adherence ( e.g. , patients do not adhere to their assigned treatment), both the standard backdoor adjustment (SBD) and the conditional front-door adjustment (CFD) can recover unbiased estimates of the treatment assignment effects. However, the estimation variance of these approaches may vary widely across settings, which remains underexplored in the literature. In this work, we demonstrate theoretically and empirically that CFD yields lower-variance estimates than SBD when the true effect of treatment assignment is small ( i.e. , assigning an intervention leads to small changes in patients’ future outcome). Additionally, since CFD requires estimating multiple nuisance parameters, we introduce LobsterNet, a multi-task neural network that implements CFD with joint modeling of the nuisance parameters. Empirically, LobsterNet reduces estimation error across several semi-synthetic and real-world datasets compared to baselines, even when the true effect of treatment assignment is large. Our findings suggest CFD with shared nuisance parameter modeling can improve treatment assignment effect estimation under non-adherence. † † volume: 287 † † year: 2025 † † workshop: Conference on Health, Inference, and Learning (CHIL) 2025 Data and Code Availability This paper uses IHDP Hill (2011) and AMR-UTI Oberst et al. (2020) datasets. Our training and evaluation codes are available at https://github.com/MLD3/LobsterNet. Institutional Review Board (IRB) This work is not regulated as human subjects research since the data are de-identified and publicly available. 1 Introduction We study the estimation of heterogeneous treatment assignment effects under treatment non-adherence from observational data. Treatment non-adherence arises when some individuals’ treatment intake does not match their treatment assignment ( e.g. , patients fail to take their prescribed medication). This can occur in many real-world applications, such as public health ( Tressler and Bhandari, 2019 ) and online advertising ( Gordon et al., 2019 ) . In many settings, decision-makers ( e.g. , clinicians) can only intervene by assigning treatment and cannot enforce adherence. To facilitate decisions about individual treatment assignments, decision-makers must estimate the heterogeneous treatment assignment effect, which accounts for both the adherence probability and standard treatment effect. We refer to this quantity as the conditional average treatment effect of the assignment (CATEA) , which corresponds to the effect of treatment assignment on the outcome of interest ( e.g. , disease progression) 1 1 1 Our estimand differs from the standard conditional average treatment effect (CATE), i.e. , the effect of adhering to the assigned treatment. . To estimate CATEAs, one can use the standard backdoor adjustment (SBD) ( Montori and Guyatt, 2001 ; Gupta, 2011 ) , which estimates potential outcomes while controlling for confounders. However, SBD ignores treatment intake information, which is often available. We propose to leverage conditional front-door adjustment (CFD) ( Pearl, 2009 ; Xu et al., 2024 ) as an alternative for estimating the CATEA that leverages treatment intake. CFD works by combining estimates of three nuisance parameters: (1) the probability of treatment assignment, (2) the probability of treatment intake given assignment, and (3) the outcome given treatment assignment and intake via an adjustment formula to form CATEA estimates. The error associated with CATEA estimates decomposes into bias and variance ( Domingos, 2000 ) . In this work, we first prove that given sufficient identification assumptions, both CFD and SBD produce unbiased CATEA estimates (Section 3 ). Thus, one should choose the estimator with the lower variance. However, whether CFD or SBD has lower variance is underexplored in the literature (see Section 2 ). Intuitively, using treatment intake information in CFD should reduce variance. However, since CFD estimates more nuisance parameters, variance from each estimate could inflate the overall variance. Due to this tension, it is unclear when CFD will yield lower-variance estimates than SBD. Thus, we compare the asymptotic variance of CFD to that of SBD and show that CFD has lower variance than SBD when the true CATEA is small. Small CATEA is common in real-world treatment assignments ( Ioannidis, 2005 ; Gu et al., 2016 ; Vivalt, 2020 ) , highlighting the practical benefits of CFD. Building on our theoretical analysis, we propose LobsterNet, a multi-task neural network architecture that jointly estimates nuisance parameters required by CFD. In particular, LobsterNet leverages shared representations between nuisance parameters to improve CFD’s performance in finite sample settings. In summary, our contributions are as follows: • to the best of our knowledge, we are the first to propose CFD for CATEA estimation and prove its unbiasedness under sufficient assumptions, • we compare the asymptotic variance of CFD to SBD in CATEA estimation, showing that CFD improves variance under small CATEA, • we propose LobsterNet, a novel CFD implementation with joint nuisance parameter modeling, improving the empirical performance of CFD compared to existing implementations. 2 Related Work Under non-adherence, the effect of treatment intake on the outcome, known as the conditional average treatment effect (CATE), is often estimated to understand the impact of treatment intake ( e.g. , a drug’s effectiveness) ( Marcus and Gibbons, 2001 ; Nagelkerke et al., 2000 ; Bargagli-Stoffi et al., 2022 ; Miladinovic et al., 2011 ) . While CATE estimation is important, estimating the treatment assignment effect ( i.e. , CATEA) could be independently useful in scenarios where decision-makers can only assign treatments but cannot enforce treatment intake ( e.g. , encouragement designs ( Hirano et al., 2000 ) ). In this work, we are interested in using observational data to estimate CATEAs ( Rahier et al., 2021 ) , which account for both the individual’s adherence probability and the effect of the treatment. Knowledge of CATEA can help optimize decisions when assigning treatments. For example, if an individual is unlikely to adhere to a prescription, the overall effect of assigning the drug would remain small even for a highly effective drug. To estimate the CATEA, prior work has applied SBD ( Montori and Guyatt, 2001 ; Gupta, 2011 ) . However, such approaches ignore the treatment intake, an intermediate variable that can improve the precision of effect estimates. Such variables have been used outside the treatment non-adherence setting ( Fisher, 1928 ; Cox, 1960 ) . Most recently, Rahier et al. (2021) demonstrated that combining CATE estimates with estimated intake probability results in better CATEA estimates than SBD. However, their approach does not leverage CFD and only applies to one-sided non-adherence (defined in Section 3.2 ) with randomized control trial (RCT) data. In contrast, we leverage the CFD formula to enable CATEA estimation with both one-sided and two-sided non-adherence using observational data. The CFD formula provides a flexible approach to combine an intermediate variable in causal effect estimation ( Pearl, 2009 ; Xu et al., 2024 ) . It has previously been proposed as an alternative to SBD in scenarios outside of treatment non-adherence ( Hill et al., 2015 ; Bellemare et al., 2020 ) . However, the relative performance of SBD and CFD is poorly understood. Glynn and Kashin (2018) studied their estimation bias subject to unobserved confounding and demonstrated that CFD has a lower bias in a broad class of applications than SBD. Similar to our work, Gupta et al. (2021)