Learning the Optimal Composite Mediator:
Closed-Form Solution and Inference · 2026
Read Learning the Optimal Composite Mediator: by Closed-Form Solution and Inference (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-2603-13762.txt.
Introduction
Zihuai He studies how to learn the optimal composite mediator between an exposure and an outcome. How effects pass through high-dimensional intermediaries is central to observational research. In a linear model, the aim is to maximize the indirect effect, yet the objective is non-convex in the weights. A geometric argument still yields a closed-form global solution: the best weight bisects the angle between two path vectors. His MaxIE method runs at least squares cost, with a dual form when mediators outnumber observations.
Catalog record
| CREATOR | Closed-Form Solution and Inference |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 72046 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.13762
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Learning the Optimal Composite Mediator: Closed-Form Solution and Inference Abstract Abstract Understanding how an exposure transmits its effect through high-dimensional intermediaries is a central problem in observational research. We study the problem of finding a composite mediator that maximises the indirect effect of an exposure on an outcome in a linear structural equation model. Although the objective is non-convex in the weight vector, a geometric argument yields a closed-form global solution: the optimal weight bisects the angle between two computable path vectors in a weighted inner product space, recovered via two linear solves. The resulting algorithm, MaxIE , runs at the same cost as ordinary least squares—orders of magnitude lower than numerical optimisation—with a dual formulation for settings where mediators outnumber observations. The same path vectors yield a test for the global null that no composite mediator exists, with t(p-1) in the classical and t(n-2) in the dual regime. Power is characterised analytically as a function of the population path angle; simulations confirm size control and the power characterisation. Applied to a UK Biobank proteomics dataset ( n=38{,}383 , p=2{,}916 ), the method rejects the global null ( p\text{-value}=6.4\times 10^{-9} ) and identifies the optimal proteomic composite mediating age’s effect on dementia. Keywords: closed-form estimator; composite mediator; global test; high-dimensional mediation; indirect effect; mediation index 1 Introduction Many contemporary studies involve a treatment or exposure \mathbf{A} whose effect on an outcome \mathbf{Y} is transmitted through a large set of candidate mediators \mathbf{X}\in\mathbb{R}^{n\times p} , such as gene expression profiles or protein abundances. Mediation analysis asks through which intermediaries this effect operates—knowledge that can point to therapeutic targets, modifiable risk factors, or mechanisms linking environmental exposures to disease. The mediated effect of a treatment often reflects the collective activity of many molecular features rather than any single one: genes act in regulatory pathways, proteins function within signalling complexes, and metabolites participate in coordinated biochemical reactions. A composite score \mathbf{M}=\mathbf{X}\mathbf{w} is therefore a natural target—one that aggregates the joint mediation signal into a single interpretable mediator and captures structure that single-feature analyses miss. A particularly pressing instance arises in aging research, where large-scale proteomics platforms now enable construction of proteomic aging clocks —composite protein scores that estimate biological age and predict a wide range of diseases including Alzheimer’s, heart failure, and cancer ( Oh et al. 2023 ; Oh et al. 2025 ; Argentieri et al. 2024 ; Tian et al. 2025 ; Wang et al. 2025 ) . These composites are increasingly used to track biological aging and stratify disease risk, and are natural candidates for monitoring the effects of anti-aging interventions ( Oh et al. 2023 ; Oh et al. 2025 ; Argentieri et al. 2024 ; Tian et al. 2025 ; Wang et al. 2025 ) . Yet existing clocks are constructed by regressing chronological age on proteins alone, without reference to any disease outcome, maximising age tracking at the expense of disease relevance. The mediation framing resolves this tension directly: chronological age remodels the proteome, and those proteomic changes are the mechanism through which age increases disease risk ( Oh et al. 2023 ; Tian et al. 2025 ) . The composite mediator of interest is a protein score that simultaneously tracks biological aging and transmits that signal to disease risk—yet this joint optimisation has not previously been pursued as a direct problem, and neither single-criterion approach guarantees that the resulting composite jointly optimises both properties. The mediation objective. We model mediation through three structural equations on centred variables ( Baron and Kenny 1986 ) : \mathbf{M}=\alpha\,\mathbf{A}+\bm{\varepsilon}_{1},\qquad\mathbf{Y}=\gamma\,\mathbf{A}+\beta\,\mathbf{M}+\bm{\varepsilon}_{2},\qquad\mathbf{Y}=\tau\,\mathbf{A}+\bm{\varepsilon}_{3}, (1) where \alpha , \beta , \gamma , \tau are path coefficients (treatment \to mediator, direct, mediator \to outcome, total) and \bm{\varepsilon}_{1},\bm{\varepsilon}_{2},\bm{\varepsilon}_{3} are mean-zero errors. We seek the weight vector \mathbf{w}^{*} that maximises the population indirect effect for the composite mediator \mathbf{M}=\mathbf{X}\mathbf{w} : \mathbf{w}^{*}_{+}\;=\;\operatorname{arg\,max}_{\mathbf{w}\in\mathbb{R}^{p}}\;\alpha(\mathbf{w})\,\beta(\mathbf{w}), (2) estimated in practice by \operatorname{arg\,max}_{\mathbf{w}}\hat{\alpha}(\mathbf{w})\hat{\beta}(\mathbf{w}) . This finds the concordant composite mediator: the direction along which treatment increases the mediator and the mediator increases the outcome (both path coefficients share the same sign). A complementary quantity is the suppression composite mediator, \mathbf{w}^{*}_{-}\;=\;\operatorname{arg\,min}_{\mathbf{w}\in\mathbb{R}^{p}}\;\alpha(\mathbf{w})\,\beta(\mathbf{w}), (3) which captures directions where treatment and outcome signals pull in opposite directions ( \alpha(\mathbf{w})\beta(\mathbf{w})<0 ). Since flipping the sign of \mathbf{w} changes the signs of both \alpha(\mathbf{w}) and \beta(\mathbf{w}) simultaneously, their product is unchanged; we orient each solution so that \hat{\alpha}(\mathbf{w}^{*}_{\pm})\geq 0 , after which the sign of \hat{\beta} identifies concordant ( \hat{\beta}>0 ) versus suppression ( \hat{\beta}<0 ) mediation. Running both recovers a complete picture of how treatment effects are transmitted through the mediator panel. The remainder of the paper focuses on the concordant estimator \mathbf{w}^{*}_{+} without loss of generality: as shown in Section 2 , \mathbf{w}^{*}_{-} is obtained by negating the sufficient statistic \mathbf{z} , so all theoretical results and algorithms apply to both problems by a trivial sign change at no additional cost. Related Work Causal mediation analysis. The modern causal framework for mediation, developed by Robins & Greenland ( Robins and Greenland 1992 ) and Pearl ( Pearl 2001 ) and extended by Imai et al. ( Imai et al. 2010 ) and VanderWeele ( VanderWeele 2015 ) , defines the natural indirect effect (NIE) and natural direct effect (NDE) in terms of potential outcomes, with identification resting on sequential ignorability and the cross-world consistency assumption. Under a linear structural equation model these reduce to the path-coefficient product \alpha\beta ( Pearl 2001 ) , which is the estimand we maximise. Semiparametric efficiency bounds for the NIE are derived in Tchetgen Tchetgen & Shpitser ( Tchetgen Tchetgen and Shpitser 2012 ) ; our contribution operates within the parametric linear model and solves the previously unaddressed problem of optimally choosing the composite mediator within this class. High-dimensional composite mediator methods. Principal component mediation analysis (PCMA, ( Huang and Pan 2016 ) ) and its sparse variant SPCMA ( Zhao et al. 2020 ) construct composite scores as principal components of the treatment-residualised mediators. These methods are designed for dimension reduction and screening: the PCA criterion maximises residual variance in \mathbf{X} , which is independent of both \mathbf{A} and \mathbf{Y} . They do not optimise any mediation criterion and are therefore solving a different problem from ours; a leading principal component need not align with either mediation path. The most direct predecessor to MaxIE is the Directions of Mediation (DM, ( Chen et al. 2018 ) ), which finds a composite mediator by maximising the joint LSEM log-likelihood via an iterative eigenvalue algorithm. Both DM and MaxIE seek a single weight vector capturing the mediation pathway, but DM uses a likelihood surrogate and requires iterative optimisation, whereas MaxIE directly maximises \hat{\alpha}(\mathbf{w})\hat{\beta}(\mathbf{w}) in closed form. Moreover, DM provides no inferential framework for the global null that no composite mediator of either sign exists ( H_{0} : ( 14 )); the cosine test derived here fills this gap. Screening and selection methods. A large literature screens individual mediators via penalisation: HIMA ( Zhang et al. 2016 ) , HDMA ( Gao et al. 2019 ) , Zhou et al. ( Zhou et al. 2020 ) , and Pathway LASSO ( Zhao and Luo 2022 ) recover active individual mediators, but any composite score they produce is a byproduct of regularisation rather than the result