Causal Inference in Biomedical Imaging via Functional Linear Structural Equation Models

Ting Li · 2026

Read Causal Inference in Biomedical Imaging via Functional Linear Structural Equation Models by Ting Li (2026) 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-2601-20610.txt.

Introduction

Ting Li builds a way to draw cause-and-effect claims from brain scans and other medical images. The aim is to learn how features of an organ, seen in a scan, change health outcomes. That kind of claim is key for research and for care, but it is hard to make with care.

The model she proposes is a functional linear structural equation model. It links three things: health outcomes, the scan data treated as smooth curves, and plain facts about the patient, such as genes, sex, and age. Keeping the scan as a curve keeps where in the organ a signal sits. That care is what lets the model aim at causes, not just at guesses about who will do well.

Catalog record

CREATORTing Li
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE64 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-20610

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Causal Inference in Biomedical Imaging via Functional Linear Structural Equation Models Thanks: Address for correspondence: Hongtu Zhu, Ph.D., E-mail: [email protected]. Ting Li Abstract Understanding the causal effects of organ-specific features from medical imaging on clinical outcomes is essential for biomedical research and patient care. We propose a novel Functional Linear Structural Equation Model (FLSEM) to capture the relationships among clinical outcomes, functional imaging exposures, and scalar covariates like genetics, sex, and age. Traditional methods struggle with the infinite-dimensional nature of exposures and complex covariates. Our FLSEM overcomes these challenges by establishing identifiable conditions using scalar instrumental variables. We develop the Functional Group Support Detection and Root Finding (FGSDAR) algorithm for efficient variable selection, supported by rigorous theoretical guarantees, including selection consistency and accurate parameter estimation. We further propose a test statistic to test the nullity of the functional coefficient, establishing its null limit distribution. Our approach is validated through extensive simulations and applied to UK Biobank data, demonstrating robust performance in detecting causal relationships from medical imaging. Keywords: Causal Effect; Functional Linear Structural Equation Model; Identification; Imaging Genetics; Instrumental Variable. 1 Introduction This paper proposes the Causal-Genetic-Imaging-Clinical (CGIC) framework to uncover causal pathways linking genetic factors, organ-level imaging markers, and clinical outcomes for complex phenotypes such as Alzheimer’s Disease (AD), while accounting for environmental influences and unobserved confounders ( 61 ; 32 ; 47 ) . Imaging modalities like functional Magnetic Resonance Imaging (fMRI) play a central role in capturing anatomical and functional properties of organs such as the brain ( 60 ) . The CGIC framework extends Jack’s influential hypothetical model of AD progression ( 26 ; 25 ) , offering a more comprehensive view of disease etiology (Figure 1 (a)). Large-scale studies such as the Alzheimer’s Disease Neuroimaging Initiative (ADNI) provide critical multimodal data for investigating the interplay among genetics, imaging, and clinical outcomes. A central challenge in establishing CGIC pathways is the presence of unmeasured or poorly controlled confounders ( 49 ) . Mendelian Randomization (MR) offers a powerful approach to infer causal effects by leveraging genetic variants as instrumental variables (IVs), under the assumption that these variants affect the outcome only through the exposure of interest ( 31 ; 4 ; 43 ) . However, applying MR in high-dimensional genetic settings is complex. Genetic variables can play diverse roles—some serve as valid instruments, others act as confounders, precision covariates, or irrelevant noise. Moreover, the number of genetic variants often far exceeds the sample size, posing significant statistical and computational challenges. Aligned with the principles of MR, we leverage genetic variants as potential IVs to investigate causal relationships between functional exposures and clinical outcomes. However, unique methodological challenges arise in this context due to the infinite-dimensional nature of functional exposures (e.g., brain imaging data), the high-dimensional and mixed roles of genetic variants, and the confounding influence of environmental and lifestyle factors on both exposures and outcomes. These complexities call for advanced statistical frameworks capable of rigorously modeling the CGIC pathway. Figure 1: (a) Directed acyclic graph showing the causal genetic-imaging-clinical (CGIC) pathway that links from genetic factors to organ imaging measures to clinical outcomes confounded with environmental factors (e.g., lifestyle factors) and possible unobserved confounders. (b) Four groups of mixed covariates. To enhance the robustness and interpretability of causal inference in this setting, we propose the Functional Linear Structural Equation Model (FLSEM). This model consists of two interconnected components: (i) a structural equation linking a scalar outcome to an endogenous functional exposure and scalar covariates; and (ii) an exposure model describing the functional predictor in terms of scalar instrumental variables. Our main contributions are as follows: (C1) We propose a three-step procedure for selecting the desired instruments and important control variables using L_{0} penalty. First, we identify the relevant variables using the functional on scalar model under the reproducing kernel Hilbert space (RKHS) framework using the L_{0} penalty. Second, we replace the treatment variable with its post-adaptive predicted value and select useful controls. Third, we use the selected controls and the predicted treatment variable to obtain the treatment effect estimator following the partial functional linear model. (C2) Different from previous studies that use the functional instrumental variables, we establish the identifiable conditions for the FLSEM, where both the instrumental covariates and useful controls are scalar. To the best of our knowledge, this is the first identifiable condition for the functional exposure model with a mixed set of scalar instruments. (C3) We propose the Functional Group Support Detection and Root Finding (FGSDAR) algorithm to select relevant instrumental covariates for the functional exposure. (C4) Theoretically, we establish the selection consistency of the true instruments and control, characterize the estimation error of both functional and scalar estimates, and propose a test statistic to test the nullity of the functional coefficient. (C5) We conduct extensive simulation studies and analyze a real dataset from UK Biobank to demonstrate the finite sample performance of the proposed methods. 1.1 Related Work Endogeneity in functional data has attracted growing attention, yet most econometric solutions still hinge on Tikhonov regularization for ill-posed inverse problems. A seminal example is 18 , which establishes IV estimation for functional linear models, convergence rates, and a notion of “instrumental strength.” Subsequent work includes spline‐based two-step estimators with measurement error ( 48 ) , comparable two-step procedures with nullity tests ( 27 ) , and fully Bayesian error correction ( 62 ) . Shifting to scalar instruments, 1 extend Florens et al.’s framework. Additional advances feature a GMM estimator for serially correlated predictors ( 9 ) , an exogeneity test with asymptotic guarantees ( 12 ) , FPCA-based IV methods ( 44 ) , and Bayesian mediation models for spatial imaging data that exploit latent-confounder structure ( 52 ; 53 ) . Yet none of these methods tackles the high-dimensional, mixed sets of scalar instruments and controls typical of modern genetics and neuroimaging studies. Our work closes this gap. Our work advances function-on-scalar regression with high-dimensional scalar covariates—a setting where most methods first expand each functional coefficient in a fixed basis and then impose group penalties to achieve sparsity. Representative examples include group MCP ( 10 ) and group LASSO ( 2 ) . Robust variants handle outliers via exponential-squared loss plus group SCAD ( 7 ) , or by combining least-squares loss with an adaptive LASSO ( 16 ) . Other directions promote simultaneous smoothness and selection through RKHS‐based penalties ( 39 ) or a smoothing elastic net ( 38 ) . Related matrix-response work employs the trace norm for selection ( 29 ) . Departing from these LASSO-type frameworks, we adopt an L_{0}- penalty in the exposure model, which enforces exact sparsity without the shrinkage bias inherent to convex penalties. More importantly, whereas prior studies focus on predictive function-on-scalar models, our goal is causal: we estimate the effect of an endogenous functional exposure on a scalar outcome. To do so, we embed the L_{0}- based selection within a functional-on-scalar exposure model that explicitly corrects for endogeneity, thereby extending the existing literature to high-dimensional causal inference with functional data. The remainder of the paper is organized as follows. Section 2 introduces the FLSEM and its identifiability conditions. Section 3 outlines the estimation procedure and proposes a test statistic to test

Read the complete text