betaFeedback

WorksMathematics

Bayesian Variable Selection for High-Dimensional Predictors with Missing Psychometric Outcomes

Abstract

High-dimensional, multimodal predictors and partially observed multivariate outcomes are common in psychometric research. However, existing regularization methods often do not accommodate hierarchical predictor structures and are primarily designed for univariate outcomes. We propose SHIM, a Bayesian framework for structured variable selection that combines hierarchical horseshoe shrinkage with a Bayesian treatment of missing outcomes. The framework jointly accommodates predictor hierarchies, dependence among outcomes, and incomplete multivariate responses. We establish theoretical properties of the proposed prior specification and evaluate SHIM through simulation studies. The results demonstrate that SHIM balances sensitivity with false-positive control while yielding accurate coefficient estimates and well-calibrated uncertainty quantification. We further apply SHIM to data from an Alzheimer's disease cohort to characterize associations between multimodal neuroimaging measures and multivariate neuropsychological outcomes and to generate posterior-based multiple imputations for downstream analyses of the relationships between fluid biomarkers and cognition. An R package, shim, is publicly available to facilitate implementation.

Introduction

Modern psychometric studies increasingly integrate multimodal data to characterize cognitive and behavioral functioning from diverse biological perspectives. In neurological and psychiatric research, for example, structural and functional neuroimaging, diffusion imaging, genomics, proteomics, and other molecular biomarkers have been collected alongside comprehensive neuropsychological assessments to investigate the complex mechanisms underlying cognitive impairment and disease progression [11, 27, 34]. Compared with analyses based on a single data source, multimodal studies may provide a richer and more comprehensive characterization of biological systems by leveraging complementary information across data modalities, potentially improving the identification and prediction of biomarkers associated with clinically meaningful psychometric outcomes [19, 42].

Meanwhile, the integration of multimodal data substantially increases the statistical complexity of psychometric studies. First, each modality may contain hundreds or thousands of candidate predictors, leading to high-dimensional regression problems in which the number of predictors may be comparable to or exceed the sample size. Conventional regression methods are therefore prone to unstable estimation, multicollinearity, and overfitting, inducing regularization and variable selection [8, 40]. Moreover, high-dimensional predictors collected from multimodal data often exhibit intrinsic hierarchical organization. For example, structural magnetic resonance imaging (MRI) measurements may be organized according to anatomical regions within imaging modalities, while molecular biomarkers may be grouped according to biological pathways or functional categories. Such hierarchical structures further induce complex correlation patterns among predictors and suggest that biologically related variables should be modeled jointly rather than treated as independent features. Existing regularization methods, however, typically ignore these multilevel relationships or accommodate only a single grouping structure, potentially limiting both statistical efficiency and scientific interpretability [1, 45, 46].

An additional challenge arises from the multivariate nature of psychometric outcomes. Modern psychometric studies routinely assess multiple cognitive or behavioral domains, such as memory, executive function, language, and attention, which reflect distinct yet correlated aspects of the same underlying neuropsychological process [43]. Modeling these outcomes jointly allows statistical procedures to borrow information across domains through their dependence, potentially improving estimation efficiency and statistical power relative to separate univariate analyses [7, 21, 47]. In addition, many biomarkers may exhibit effects that influence multiple psychometric outcomes simultaneously [18, 36]. Joint modeling therefore provides a coherent framework for identifying predictors with shared or outcome-specific associations while appropriately accounting for the correlation structure among outcomes. Despite the availability of multivariate extensions, many widely used high-dimensional variable-selection methods were originally developed for single-outcome regression settings and may fail to exploit dependence across multiple outcomes [10].

Existing methods have addressed different aspects of the statistical challenges arising from multimodal psychometric studies, although typically in isolation. To accommodate structured predictors, group-based regularization methods, including the group lasso and Bayesian sparse group selection, have been developed to exploit grouping information among predictors [44, 46]. More recently, hierarchical Bayesian shrinkage priors have been proposed to accommodate nested predictor structures through adaptive regularization at multiple levels [1, 45]. While these approaches improve variable selection by incorporating predictor structure, they have primarily focused on univariate responses and generally do not account for dependence among multiple psychometric outcomes. Multivariate Bayesian variable selection methods have also been developed to exploit correlations among multiple outcomes and identify predictors with shared effects across responses [7, 21, 24]. However, these methods typically treat predictors as independent or accommodate a single grouping structure. Finally, methods that have been devoted to handling incomplete outcomes through multiple imputation and joint modeling approaches are not designed for high-dimensional variable selection with hierarchically structured predictors [35, 39]. Therefore, there remains a need for a unified statistical framework that simultaneously accommodates multilevel predictor structures, multivariate psychometric outcomes, and incomplete outcome data.

To address these challenges, we propose SHIM (Shrinkage for Hierarchical predictors and Incomplete Multivariate outcomes), a Bayesian framework for variable selection with high-dimensional, multimodal predictors and partially observed multivariate psychometric outcomes. SHIM considers a hierarchical horseshoe prior for the regression coefficients to accommodate the multilevel structure of multimodal predictors while jointly modeling correlated psychometric outcomes through a multivariate regression model. Missing outcomes are treated as latent variables and inferred jointly with the model parameters within a Bayesian framework. In addition to direct posterior inference for the model parameters, posterior samples of the missing outcomes naturally yield a posterior-based multiple imputation representation for subsequent analyses under statistical models requiring completed outcome data. An open-source R package, shim, is available at https://github.com/zongyue-teng/shim to facilitate the implementation of SHIM.

The remainder of the manuscript is organized as follows. Section 2 introduces the SHIM framework, including the prior specification, posterior computation, posterior-based multiple-imputation strategy, pseudocode algorithm, and probabilistic justification of the prior specification. Section 3 presents a simulation study comprising eight primary scenarios that vary in predictor dimension, modality and group structure, and proportion of missing outcomes. This section also reports four sets of sensitivity analyses assessing the robustness of SHIM and an additional experiment evaluating its multiple-imputation performance. Section 4 presents two applications to an Alzheimer’s disease cohort, where the first examines associations between multimodal neuroimaging measures and multivariate outcomes, and the second evaluates multiple-imputation performance in a downstream analysis of the relationship between cerebrospinal fluid biomarkers and cognition. Finally, Section 5 summarizes the proposed method, discusses its limitations, and outlines directions for future research.

Method

Notations

Consider a study with nn independent individuals indexed by i=1,2,…,ni = 1, 2, \ldots, n. For each individual, let yi=(yi1,yi2,…,yiQ)T∈RQ\mathbf{y}_i = (y_{i1}, y_{i2}, \ldots, y_{iQ})^{\mathsf{T}} \in\mathbb{R}^{Q} denote a vector of psychometric outcomes measured across QQ domains. In many clinical and behavioral studies, some components of yi\mathbf{y}_i may be missing due to missed visits, inability to complete assessments, or study design [29, 38]. Let ri=(ri1,ri2,…,riQ)T\mathbf{r}_i = (r_{i1}, r_{i2}, \ldots, r_{iQ})^{\mathsf{T}} denote the corresponding missingness indicators, with riq=1r_{iq} = 1 if yiqy_{iq} is observed and riq=0r_{iq} = 0 otherwise. Accordingly, we are able to write yi=(yi,obs,yi,mis)\mathbf{y}_i = (\mathbf{y}_{i,\mathrm{obs}}, \mathbf{y}_{i,\mathrm{mis}}) to distinguish the observed and missing components.

For each individual ii, a high-dimensional predictor vector xi=(xi1,xi2,…,xiP)T∈RP\mathbf{x}_i = (x_{i1}, x_{i2}, \ldots, x_{iP})^{\mathsf{T}} \in\mathbb{R}^{P} is observed. Throughout the paper, predictors are assumed to be fully observed and standardized to have mean zero and unit variance. The feature dimension PP may be substantially larger than the sample size nn. In psychometric and biomedical applications, these predictors may represent high-dimensional features derived from one or multiple data modalities, such as regional brain imaging measurements, diffusion tract characteristics, or other high-dimensional biomarkers [11, 34]. In addition to the high-dimensional predictors, let zi=(zi1,zi2,…,ziL)T∈RL\mathbf{z}_i = (z_{i1}, z_{i2}, \ldots, z_{iL})^{\mathsf{T}} \in\mathbb{R}^{L} denote a vector of low-dimensional covariates, where LL is fixed and typically much smaller than both nn and PP. These covariates may include demographic and clinical variables, such as age, sex, education, and apolipoprotein E (APOE) genotype, which are conventionally adjusted for in psychometric and biomedical studies. Throughout the paper, these covariates are assumed to be fully observed and are not subject to hierarchical shrinkage.

Many high-dimensional predictors possess a natural hierarchical organization. For example, multimodal neuroimaging predictors may be organized according to imaging modality and anatomical region. To accommodate such structure, we assume that predictors are organized into three nested levels: (1) modality level indexed by m=1,2,…,Mm = 1, 2, \ldots, M, (2) group level within modality mm indexed by g=1,2,…,Gmg = 1, 2, \ldots, G_m, and (3) feature level within group gg indexed by k=1,2,…,Pmgk = 1, 2, \ldots, P_{mg}. Under this organization, the total number of predictors is

P=∑m=1M∑g=1GmPmg.P = \sum_{m=1}^{M} \sum_{g=1}^{G_m} P_{mg}.

The proposed method in this work is developed under this hierarchical structure, but the framework can be extended naturally to more complex hierarchical organizations. For notation convenience, predictors are subsequently re-indexed by k=1,2,…,Pk = 1, 2, \ldots, P, while m(k)m(k) and g(k)g(k) respectively identify the associated modality and group memberships.

Consider the multivariate regression model

yi=α+ΨTzi+BTxi+εi,\mathbf{y}_i = \boldsymbol{\alpha} + \boldsymbol{\Psi}^{\mathsf{T}}\mathbf{z}_i + \mathbf{B}^{\mathsf{T}}\mathbf{x}_i + \boldsymbol{\varepsilon}_i,

where α∈RQ\boldsymbol{\alpha} \in\mathbb{R}^{Q} is an intercept vector,

Ψ=(ψ1,ψ2,…,ψQ)∈RL×Q with ψq∈RL,\boldsymbol{\Psi} = (\boldsymbol{\psi}_1, \boldsymbol{\psi}_2, \ldots, \boldsymbol{\psi}_Q) \in\mathbb{R}^{L \times Q} \text{ with } \boldsymbol{\psi}_q \in\mathbb{R}^{L},

is the matrix of regression coefficients corresponding to the low-dimensional, non-regularized covariates,

B=(β1,β2,…,βQ)∈RP×Q with βq∈RP,\mathbf{B} = (\boldsymbol{\beta}_1, \boldsymbol{\beta}_2, \ldots, \boldsymbol{\beta}_Q) \in\mathbb{R}^{P \times Q} \text{ with } \boldsymbol{\beta}_q \in\mathbb{R}^{P},

is the matrix of regression coefficients corresponding to the high-dimensional predictors, and εi∼NQ(0,Σ)\boldsymbol{\varepsilon}_i \sim\mathcal{N}_Q(\mathbf{0}, \boldsymbol{\Sigma}) captures the residual dependence among outcomes.

The coefficient matrix B\mathbf{B} is assumed to be sparse, reflecting that only a subset of the high-dimensional predictors contributes meaningfully to the psychometric outcomes. In contrast, the low-dimensional covariates in zi\mathbf{z}_i are treated as adjustment variables and are not subject to shrinkage. The primary objective of the proposed framework is to estimate the regression coefficient matrix B\mathbf{B} and identify important predictors while accounting for both the hierarchical organization of the high-dimensional predictors and the multivariate dependence among the outcomes.

Multivariate Regression with Missing Outcomes

Conditional on the predictor and covariate vectors xi\boldsymbol{x}_i and zi\boldsymbol{z}_i and model parameters θ:=(α,Ψ,B,Σ)\boldsymbol{\theta} := (\boldsymbol{\alpha}, \boldsymbol{\Psi}, \boldsymbol{B}, \boldsymbol{\Sigma}), the proposed multivariate regression model assumes

yi∣xi,zi,θ∼NQ(α+ΨTzi+BTxi,Σ).(1)\boldsymbol{y}_i \mid\boldsymbol{x}_i,\boldsymbol{z}_i,\boldsymbol{\theta} \sim\mathcal{N}_Q\left(\boldsymbol{\alpha} + \boldsymbol{\Psi}^{\mathsf{T}}\boldsymbol{z}_i + \boldsymbol{B}^{\mathsf{T}}\boldsymbol{x}_i,\boldsymbol{\Sigma}\right). \tag*{(1)}

This formulation allows the covariance matrix Σ\boldsymbol{\Sigma} to capture residual correlations among psychometric domains and borrow information across outcomes when some outcomes are partially observed. Throughout the analysis, we assume that the missing data mechanism satisfies the missing at random (MAR) condition; that is

Pr⁡(ri∣yi,xi,zi)=Pr⁡(ri∣yi,obs,xi,zi),\Pr(\boldsymbol{r}_i \mid\boldsymbol{y}_i,\boldsymbol{x}_i,\boldsymbol{z}_i) = \Pr(\boldsymbol{r}_i \mid\boldsymbol{y}_{i,\mathrm{obs}},\boldsymbol{x}_i,\boldsymbol{z}_i),

under which the probability of missingness depends on the observed outcomes together with the observed covariates and predictors. Under this assumption, inference for θ\boldsymbol{\theta} can be based on the observed-data likelihood obtained by integrating over the missing outcomes. Specifically, the contribution of subject ii to the observed-data likelihood is

p(yi,obs∣xi,zi,θ)=∫p(yi∣xi,zi,θ) dyi,mis,p(\boldsymbol{y}_{i,\mathrm{obs}} \mid\boldsymbol{x}_i,\boldsymbol{z}_i,\boldsymbol{\theta}) = \int p(\boldsymbol{y}_i \mid\boldsymbol{x}_i,\boldsymbol{z}_i,\boldsymbol{\theta})\,\mathrm{d}\boldsymbol{y}_{i,\mathrm{mis}},

where p(yi∣xi,zi,θ)p(\boldsymbol{y}_i \mid\boldsymbol{x}_i,\boldsymbol{z}_i,\boldsymbol{\theta}) is specified in (1).

Rather than directly maximizing the observed-data likelihood, which requires computationally intensive numerical integration, we treat the missing outcomes yi,mis\boldsymbol{y}_{i,\mathrm{mis}} as latent variables and jointly infer them along with the model parameters [39]. Specifically, inference is based on the joint posterior distribution p(θ,Ymis∣Yobs,X)p(\boldsymbol{\theta},\boldsymbol{Y}_{\mathrm{mis}} \mid\boldsymbol{Y}_{\mathrm{obs}},\boldsymbol{X}), where Ymis={y1,mis,y2,mis,…,yn,mis}\boldsymbol{Y}_{\mathrm{mis}} = \{\boldsymbol{y}_{1,\mathrm{mis}},\boldsymbol{y}_{2,\mathrm{mis}},\ldots,\boldsymbol{y}_{n,\mathrm{mis}}\} and Yobs={y1,obs,y2,obs,…,yn,obs}\boldsymbol{Y}_{\mathrm{obs}} = \{\boldsymbol{y}_{1,\mathrm{obs}},\boldsymbol{y}_{2,\mathrm{obs}},\ldots,\boldsymbol{y}_{n,\mathrm{obs}}\} denote the collections of missing and observed outcomes, respectively. This latent-variable formulation propagates uncertainty associated with the missing outcomes directly into posterior inference for all model parameters. By jointly inferring the missing outcomes and model parameters within a unified Bayesian framework, this approach avoids potential incompatibility between separate imputation and analysis stages, providing coherent Bayesian inference for multivariate variable selection [13].

As a natural consequence of the joint Bayesian formulation, posterior draws of the missing outcomes additionally generate completed datasets, yielding a posterior-based multiple imputation representation of the missing data. Under the multivariate regression model in (1), posterior inference for the model parameters is performed directly from their joint posterior distribution and therefore does not require a separate multiple imputation stage. Instead, the completed datasets provide a coherent and convenient interface for downstream analyses using statistical models other than the proposed multivariate regression model that require fully observed outcome data. Additional details are presented in Section 2.5.

Hierarchical Horseshoe Prior Specification

To perform multilevel variable selection aligned with the hierarchical organization of the predictors, we propose SHIM (Shrinkage for Hierarchical predictors and Incomplete Multivariate outcomes), which employs a hierarchical horseshoe prior on the regression coefficient matrix B\boldsymbol{B}, motivated by grouped and hierarchical extensions of the classical horseshoe prior [1, 45]. Let βkq\beta_{kq} denote the regression coefficient corresponding to predictor kk and outcome qq. We introduce a hierarchical shrinking structure that encourages predictor-specific sparsity to be shared across outcomes while allowing outcome-specific deviations. Specifically, we assume

βkq∣τ,γm(k),δm(k),g(k),λk,ξkq∼N(0,τ2γm(k)2δm(k),g(k)2λk2ξkq2).(2)\beta_{kq} \mid\tau,\gamma_{m(k)},\delta_{m(k),g(k)},\lambda_k,\xi_{kq} \sim\mathcal{N}\left(0,\tau^2\gamma_{m(k)}^2\delta_{m(k),g(k)}^2\lambda_k^2\xi_{kq}^2\right). \tag*{(2)}

where τ>0\tau> 0 is a global shrinkage parameter controlling the overall sparsity across predictors; γm(k)>0\gamma_{m(k)} > 0 is a modality-level shrinkage parameter for the modality containing predictor kk, allowing entire modalities to be adaptively shrunk toward zero when weakly associated with the outcomes; δm(k),g(k)>0\delta_{m(k),g(k)} > 0 is a group-level shrinkage parameter for the group containing predictor kk, inducing regularization among related predictors within the same modality; λk>0\lambda_k > 0 is a predictor-level local shrinkage parameter shared across all outcomes; and ξkq>0\xi_{kq} > 0 is an outcome-specific local shrinkage parameter that allows individual coefficients to deviate from the shared sparsity pattern. These shrinkage parameters jointly induce regularization at the global, modality, group, predictor, and outcome levels. Note that the indices m(k)m(k) and g(k)g(k) identify the modality and the associated within-modality group of predictor kk, respectively.

We assign independent half-Cauchy priors to the shrinkage parameters:

τ∼C+(0,1),γm∼C+(0,1),δg∼C+(0,1),λk∼C+(0,1),ξkq∼C+(0,1),\tau\sim C^{+}(0,1), \quad\gamma_m \sim C^{+}(0,1), \quad\delta_g \sim C^{+}(0,1), \quad\lambda_k \sim C^{+}(0,1), \quad\xi_{kq} \sim C^{+}(0,1),

for all m=1,2,…,Mm = 1,2,\ldots,M, g=1,2,…,Gmg = 1,2,\ldots,G_m, k=1,2,…,Pk = 1,2,\ldots,P, and q=1,2,…,Qq = 1,2,\ldots,Q. This specification extends the classical horseshoe prior [8] to a hierarchical setting that enables adaptive shrinkage across multiple levels of predictor organization.

We would like to note that the shrinkage parameters in Equation (2) enter the model through a multiplicative variance decomposition, and are therefore not individually identifiable. Specifically, different configurations of (τ,γm(k),δm(k),g(k),λk,ξkq)(\tau, \gamma_{m(k)}, \delta_{m(k),g(k)}, \lambda_k, \xi_{kq}) may induce the same marginal variance for βkq\beta_{kq}, resulting in a non-unique decomposition of the overall shrinkage effect. Such non-identifiability is a common characteristic of global-local shrinkage priors and has been widely discussed in the literature [4, 5, 8]. However, this non-identifiability does not affect inference on the regression coefficients B\mathbf{B}, which are the primary quantities of interest. Accordingly, the shrinkage parameters should be viewed as auxiliary variables that jointly determine the amount of regularization through their product, rather than as separately interpretable model parameters.

Posterior Computation

Statistical inference is based on the joint posterior distribution

p(θ,Ymis,τ,γ,δ,λ,ξ∣Yobs,X,Z),p(\boldsymbol{\theta},\mathbf{Y}_{\mathrm{mis}},\tau,\boldsymbol{\gamma},\boldsymbol{\delta},\boldsymbol{\lambda},\boldsymbol{\xi} \mid\mathbf{Y}_{\mathrm{obs}},\mathbf{X},\mathbf{Z}),

where γ\boldsymbol{\gamma}, δ\boldsymbol{\delta}, λ\boldsymbol{\lambda}, and ξ\boldsymbol{\xi} collect the shrinkage parameters introduced in Section 2.3. Direct evaluation of this posterior distribution is analytically intractable due to the hierarchical shrinkage structure and the presence of missing outcomes. We therefore adopt a Bayesian data augmentation strategy [39] and construct a Markov chain Monte Carlo (MCMC) algorithm that iteratively samples from the full conditional distributions of each block of unknown quantities. Specifically, at each iteration, the algorithm first updates the missing outcomes Ymis\mathbf{Y}_{\mathrm{mis}} conditional on the current parameter values, followed by the regression parameters (α,Φ,B)(\boldsymbol{\alpha},\mathbf{\Phi},\mathbf{B}), the covariance matrix Σ\mathbf{\Sigma}, and the shrinkage parameters (τ,γm(k),δm(k),g(k),λk,ξkq)(\tau,\gamma_{m(k)},\delta_{m(k),g(k)},\lambda_k,\xi_{kq}). These updates constitute a Gibbs sampler with closed-form or conjugate distributions. For notation simplicity, iteration superscripts are omitted throughout this subsection.

For each individual ii, the conditional distribution of the missing outcome components yi,mis\mathbf{y}_{i,\mathrm{mis}} given the observed components yi,obs\mathbf{y}_{i,\mathrm{obs}} and the current parameter values follows directly from the multivariate normal model in Equation (1). Define μi=α+ΨTzi+BTxi\boldsymbol{\mu}_i = \boldsymbol{\alpha} + \mathbf{\Psi}^{\mathsf{T}}\mathbf{z}_i + \mathbf{B}^{\mathsf{T}}\mathbf{x}_i, and partition yi\mathbf{y}_i, μi\boldsymbol{\mu}_i, and Σ\mathbf{\Sigma} according to the observed and missing outcome components for subject ii:

yi=(yi,obsyi,mis),μi=(μi,oμi,m),Σ=(ΣooΣomΣmoΣmm)\mathbf{y}_i = \begin{pmatrix} \mathbf{y}_{i,\mathrm{obs}} \\ \mathbf{y}_{i,\mathrm{mis}} \end{pmatrix}, \qquad \boldsymbol{\mu}_i = \begin{pmatrix} \boldsymbol{\mu}_{i,o} \\ \boldsymbol{\mu}_{i,m} \end{pmatrix}, \qquad \mathbf{\Sigma} = \begin{pmatrix} \mathbf{\Sigma}_{oo} & \mathbf{\Sigma}_{om} \\ \mathbf{\Sigma}_{mo} & \mathbf{\Sigma}_{mm} \end{pmatrix}

where the subscripts oo and mm denote the observed and missing outcome components, respectively, for individual ii. The conditional distribution of yi,mis\mathbf{y}_{i,\mathrm{mis}} is therefore

yi,mis∣yi,obs,xi,zi,θ∼N(μi,m+ΣmoΣoo−1(yi,obs−μi,o),Σmm−ΣmoΣoo−1Σom).\mathbf{y}_{i,\mathrm{mis}} \mid\mathbf{y}_{i,\mathrm{obs}},\mathbf{x}_i,\mathbf{z}_i,\boldsymbol{\theta} \sim\mathcal{N}\left(\boldsymbol{\mu}_{i,m}+\boldsymbol{\Sigma}_{mo}\boldsymbol{\Sigma}_{oo}^{-1}(\mathbf{y}_{i,\mathrm{obs}}-\boldsymbol{\mu}_{i,o}),\boldsymbol{\Sigma}_{mm}-\boldsymbol{\Sigma}_{mo}\boldsymbol{\Sigma}_{oo}^{-1}\boldsymbol{\Sigma}_{om}\right).

Because the partition into observed and missing outcome components is subject-specific, the dimensions of yi,mis\mathbf{y}_{i,\mathrm{mis}} and the associated covariance blocks generally vary across individuals.

To update the regression coefficients, we stack the completed outcome data and perform a joint update. Denote Y=(y1,y2,…,yn)T∈Rn×Q\mathbf{Y}=(\mathbf{y}_1,\mathbf{y}_2,\ldots,\mathbf{y}_n)^{\mathsf{T}}\in\mathbb{R}^{n\times Q} and X=(x1,x2,…,xn)T∈Rn×P\mathbf{X}=(\mathbf{x}_1,\mathbf{x}_2,\ldots,\mathbf{x}_n)^{\mathsf{T}}\in\mathbb{R}^{n\times P}. The regression model in Equation (1) can be written in matrix form as

Y=1nαT+ZΨ+XB+E,\mathbf{Y}=\mathbf{1}_n\boldsymbol{\alpha}^{\mathsf{T}}+\mathbf{Z}\boldsymbol{\Psi}+\mathbf{X}\mathbf{B}+\mathbf{E},

where 1n\mathbf{1}_n is the nn-dimensional vector of ones, and the rows of E\mathbf{E} are independently distributed as NQ(0,Σ)\mathcal{N}_Q(\mathbf{0},\boldsymbol{\Sigma}). Next, let bk=(βk1,βk2,…,βkQ)T∈RQ\mathbf{b}_k=(\beta_{k1},\beta_{k2},\ldots,\beta_{kQ})^{\mathsf{T}}\in\mathbb{R}^Q, for k=1,2,…,Pk=1,2,\ldots,P, denote the coefficient vector associated with predictor kk across the QQ outcomes. Equivalently, bkT\mathbf{b}_k^{\mathsf{T}} is the kkth row of B\mathbf{B}. Conditional on the shrinkage parameters, these predictor-specific coefficient vectors are mutually independent and satisfy

bk∣τ,γm(k),δm(k),g(k),λk,ξk1,…,ξkQ∼NQ(0,Ωk),\mathbf{b}_k\mid\boldsymbol{\tau},\gamma_{m(k)},\delta_{m(k),g(k)},\lambda_k,\xi_{k1},\ldots,\xi_{kQ}\sim\mathcal{N}_Q(\mathbf{0},\boldsymbol{\Omega}_k),

with

Ωk=diag⁡(τ2γm(k)2δm(k),g(k)2λk2ξk12,τ2γm(k)2δm(k),g(k)2λk2ξk22,…,τ2γm(k)2δm(k),g(k)2λk2ξkQ2).\boldsymbol{\Omega}_k=\operatorname{diag}\left(\tau^2\gamma_{m(k)}^2\delta_{m(k),g(k)}^2\lambda_k^2\xi_{k1}^2,\tau^2\gamma_{m(k)}^2\delta_{m(k),g(k)}^2\lambda_k^2\xi_{k2}^2,\ldots,\tau^2\gamma_{m(k)}^2\delta_{m(k),g(k)}^2\lambda_k^2\xi_{kQ}^2\right).

Let vec⁡(⋅)\operatorname{vec}(\cdot) denote column-wise vectorization. Then vec⁡(BT)=(b1T,b2T,…,bPT)T\operatorname{vec}(\mathbf{B}^{\mathsf{T}})=(\mathbf{b}_1^{\mathsf{T}},\mathbf{b}_2^{\mathsf{T}},\ldots,\mathbf{b}_P^{\mathsf{T}})^{\mathsf{T}} is equivalent to stacking the rows of B\mathbf{B}. Conditional on the shrinkage parameters, the prior for B\mathbf{B} admits the equivalent representation

vec⁡(BT)∣τ,γ,δ,λ,ξ∼NPQ(0,DB),\operatorname{vec}(\mathbf{B}^{\mathsf{T}})\mid\boldsymbol{\tau},\boldsymbol{\gamma},\boldsymbol{\delta},\boldsymbol{\lambda},\boldsymbol{\xi}\sim\mathcal{N}_{PQ}(\mathbf{0},\mathbf{D}_{B}),

with DB=blockdiag⁡(Ω1,Ω2,…,ΩP)\mathbf{D}_{B}=\operatorname{blockdiag}(\boldsymbol{\Omega}_1,\boldsymbol{\Omega}_2,\ldots,\boldsymbol{\Omega}_P).

Combining the matrix form of the Gaussian complete-data likelihood with the Gaussian prior for vec⁡(BT)\operatorname{vec}(\mathbf{B}^{\mathsf{T}}), the conditional posterior distribution of vec⁡(BT)\operatorname{vec}(\mathbf{B}^{\mathsf{T}}) is

vec⁡(BT)∣others∼NPQ(μB,ΣB),\operatorname{vec}(\mathbf{B}^{\mathsf{T}})\mid\text{others}\sim\mathcal{N}_{PQ}(\boldsymbol{\mu}_B,\boldsymbol{\Sigma}_B),

with posterior covariance and mean respectively given by

ΣB={DB−1+(XTX)⊗Σ−1}−1,μB=ΣB(XT⊗Σ−1)vec⁡(Y−1nαT−ZΨ)T;\boldsymbol{\Sigma}_B=\left\{\mathbf{D}_B^{-1}+(\mathbf{X}^{\mathsf{T}}\mathbf{X})\otimes\boldsymbol{\Sigma}^{-1}\right\}^{-1},\qquad \boldsymbol{\mu}_B=\boldsymbol{\Sigma}_B(\mathbf{X}^{\mathsf{T}}\otimes\boldsymbol{\Sigma}^{-1})\operatorname{vec}\left(\mathbf{Y}-\mathbf{1}_n\boldsymbol{\alpha}^{\mathsf{T}}-\mathbf{Z}\boldsymbol{\Psi}\right)^{\mathsf{T}};

see Appendix A.1 for detailed derivations.

To update the intercept vector α\boldsymbol{\alpha} and the regression coefficient matrix Ψ\boldsymbol{\Psi} corresponding to the low-dimensional covariates that are not subject to shrinkage, we combine them into a single coefficient matrix. Specifically, define Z~=(1n,Z)∈Rn×(L+1)\widetilde{\mathbf{Z}}=(\mathbf{1}_n,\mathbf{Z})\in\mathbb{R}^{n\times(L+1)} and

Ψ~=(αTΨ)∈R(L+1)×Q,\widetilde{\boldsymbol{\Psi}}= \begin{pmatrix} \boldsymbol{\alpha}^{\mathsf{T}}\\ \boldsymbol{\Psi} \end{pmatrix} \in\mathbb{R}^{(L+1)\times Q},

where the first row of Ψ~\widetilde{\boldsymbol{\Psi}} corresponds to the intercept vector and the remaining rows correspond to the regression coefficients associated with the low-dimensional covariates. The multivariate regression model can therefore be written as

Y=Z~Ψ~+XB+E.\mathbf{Y}=\widetilde{\mathbf{Z}}\widetilde{\boldsymbol{\Psi}}+\mathbf{X}\mathbf{B}+\mathbf{E}.

Let ψ~q∈RL+1\tilde{\boldsymbol{\psi}}_q \in\mathbb{R}^{L+1} denote the qq-th column of Ψ~\tilde{\boldsymbol{\Psi}}, corresponding to the intercept and low-dimensional covariate coefficients for outcome qq. We assign independent Gaussian priors to the columns of Ψ~\tilde{\boldsymbol{\Psi}} as

ψ~q∼indNL+1(ψ~0q,V~0),\tilde{\boldsymbol{\psi}}_q \mathrel{\overset{\mathrm{ind}}{\sim}} \mathcal{N}_{L+1}\left(\tilde{\boldsymbol{\psi}}_{0q}, \tilde{\mathbf{V}}_0\right),

for q=1,2,…,Qq = 1, 2, \ldots, Q, where ψ~0q∈RL+1\tilde{\boldsymbol{\psi}}_{0q} \in\mathbb{R}^{L+1} denotes the prior mean vector for outcome qq and V~0∈R(L+1)×(L+1)\tilde{\mathbf{V}}_0 \in\mathbb{R}^{(L+1)\times(L+1)} is a symmetric positive definite prior covariance matrix. For notation simplicity, define Ψ~0=(ψ~01,ψ~02,…,ψ~0Q)∈R(L+1)×Q\tilde{\boldsymbol{\Psi}}_0 = (\tilde{\boldsymbol{\psi}}_{01}, \tilde{\boldsymbol{\psi}}_{02}, \ldots, \tilde{\boldsymbol{\psi}}_{0Q}) \in\mathbb{R}^{(L+1)\times Q}. It follows that the prior for Ψ~\tilde{\boldsymbol{\Psi}} can be expressed as

vec⁡(Ψ~T)∼N(L+1)Q(vec⁡(Ψ~0T),V~0⊗IQ).\operatorname{vec}\left(\tilde{\boldsymbol{\Psi}}^{\mathsf{T}}\right) \sim\mathcal{N}_{(L+1)Q}\left(\operatorname{vec}\left(\tilde{\boldsymbol{\Psi}}_0^{\mathsf{T}}\right), \tilde{\mathbf{V}}_0 \otimes I_Q\right).

Combining the matrix form of the Gaussian complete-data likelihood with this Gaussian prior, the conditional posterior distribution of vec⁡(Ψ~T)\operatorname{vec}\left(\tilde{\boldsymbol{\Psi}}^{\mathsf{T}}\right) is

vec⁡(Ψ~T)∣others∼N(L+1)Q(μΨ~,ΣΨ~),\operatorname{vec}\left(\tilde{\boldsymbol{\Psi}}^{\mathsf{T}}\right) \mid\text{others} \sim\mathcal{N}_{(L+1)Q}\left(\boldsymbol{\mu}_{\tilde{\boldsymbol{\Psi}}}, \boldsymbol{\Sigma}_{\tilde{\boldsymbol{\Psi}}}\right),

with posterior covariance and mean respectively given by

ΣΨ~={V~0−1⊗IQ+(Z~TZ~)⊗Σ−1}−1,\boldsymbol{\Sigma}_{\tilde{\boldsymbol{\Psi}}} = \left\{\tilde{\mathbf{V}}_0^{-1} \otimes I_Q + \left(\tilde{\mathbf{Z}}^{\mathsf{T}}\tilde{\mathbf{Z}}\right) \otimes\boldsymbol{\Sigma}^{-1}\right\}^{-1},
μΨ~=ΣΨ~{(V~0−1⊗IQ)vec⁡(Ψ~0T)+(Z~T⊗Σ−1)vec⁡((Y−XB)T)}.\boldsymbol{\mu}_{\tilde{\boldsymbol{\Psi}}} = \boldsymbol{\Sigma}_{\tilde{\boldsymbol{\Psi}}}\left\{\left(\tilde{\mathbf{V}}_0^{-1} \otimes I_Q\right)\operatorname{vec}\left(\tilde{\boldsymbol{\Psi}}_0^{\mathsf{T}}\right) + \left(\tilde{\mathbf{Z}}^{\mathsf{T}} \otimes\boldsymbol{\Sigma}^{-1}\right)\operatorname{vec}\left((\mathbf{Y}-\mathbf{X}\mathbf{B})^{\mathsf{T}}\right)\right\}.

See Appendix A.2 for detailed derivations.

To update the residual covariance matrix, define the residual vector for subject ii, conditional on the current values of α\boldsymbol{\alpha}, ψ\boldsymbol{\psi}, B\mathbf{B}, and the completed outcome data, by ei=yi−α−ΨTzi−BTxi\mathbf{e}_i = \mathbf{y}_i - \boldsymbol{\alpha} - \boldsymbol{\Psi}^{\mathsf{T}}\mathbf{z}_i - \mathbf{B}^{\mathsf{T}}\mathbf{x}_i. We assign the inverse-Wishart prior Σ∼IW(ν0,S0)\boldsymbol{\Sigma} \sim\mathcal{IW}(\nu_0, \mathbf{S}_0), where ν0\nu_0 denotes the prior degrees of freedom and S0\mathbf{S}_0 is a symmetric positive definite matrix. Since the inverse-Wishart prior is conjugate to the multivariate normal likelihood [13], the posterior distribution of Σ\boldsymbol{\Sigma} is

Σ∣others∼IW(ν0+n,S0+∑i=1neieiT).\boldsymbol{\Sigma} \mid\text{others} \sim\mathcal{IW}\left(\nu_0+n,\mathbf{S}_0+\sum_{i=1}^{n}\mathbf{e}_i\mathbf{e}_i^{\mathsf{T}}\right).

Finally, the shrinkage parameters introduced in Section 2.3 are updated using the auxiliary-variable representation of the half-Cauchy prior proposed by [26]. Under this representation, each half-Cauchy prior is expressed as an inverse-gamma mixture. Together with the Gaussian prior for the regression coefficients in Equation (2), this leads to inverse-gamma conditional posterior distributions for the corresponding squared shrinkage parameters. Within each Gibbs sampling iteration, the global, modality-level, group-level, predictor-level, and outcome-specific shrinkage parameters are updated sequentially according to their respective conditional posterior distributions. The complete Gibbs sampling procedure is summarized in Algorithm 1.

Posterior-Based Multiple Imputation

As discussed in Section 2.2, inference for the model parameters is conducted directly from their marginal posterior distributions. Meanwhile, posterior samples of Ymis\mathbf{Y}_{\mathrm{mis}} also provide a natural posterior-based multiple imputation representation, which is particularly useful when the completed outcome data are subsequently analyzed using statistical models that differ from Equation (1).

Suppose that, after discarding the burn-in iterations, a total of T−TburnT-T_{\mathrm{burn}} posterior samples are retained. To construct multiple imputations, a subset of MM posterior draws (M≪T−TburnM \ll T-T_{\mathrm{burn}}) is selected,

Algorithm 1 Gibbs Sampler for Posterior Computation

  1. Initialize Ymis(0)\mathbf{Y}_{\mathrm{mis}}^{(0)}, α(0)\alpha^{(0)}, Ψ(0)\boldsymbol{\Psi}^{(0)}, B(0)\mathbf{B}^{(0)}, Σ(0)\boldsymbol{\Sigma}^{(0)}, and shrinkage parameters.

  2. for t=1,2,…,Tt = 1, 2, \ldots, T do

  3. Sample Ymis(t)\mathbf{Y}_{\mathrm{mis}}^{(t)} from its conditional multivariate normal distribution.

  4. Sample B(t)\mathbf{B}^{(t)} from its conditional multivariate normal distribution.

  5. Sample α(t)\alpha^{(t)} and Ψ(t)\boldsymbol{\Psi}^{(t)} jointly from their conditional multivariate normal distribution.

  6. Sample Σ(t)\boldsymbol{\Sigma}^{(t)} from its conditional inverse-Wishart distribution.

  7. Update the shrinkage parameters τ(t)\boldsymbol{\tau}^{(t)}, γ(t)\boldsymbol{\gamma}^{(t)}, δ(t)\boldsymbol{\delta}^{(t)}, λ(t)\boldsymbol{\lambda}^{(t)}, and ξ(t)\boldsymbol{\xi}^{(t)} from their respective conditional posterior distributions.

  8. end for

  9. Retain posterior samples after burn-in for inference.

for example, through systematic thinning. For m=1,…,Mm = 1, \ldots, M, the mmth selected posterior draw of the missing outcomes, Ymis(m)\mathbf{Y}_{\mathrm{mis}}^{(m)}, together with the observed outcomes, defines the completed dataset given by Y(m)=(Yobs,Ymis(m))\mathbf{Y}^{(m)} = (\mathbf{Y}_{\mathrm{obs}}, \mathbf{Y}_{\mathrm{mis}}^{(m)}). The completed datasets {Y(1),Y(2),…,Y(M)}\{\mathbf{Y}^{(1)}, \mathbf{Y}^{(2)}, \ldots, \mathbf{Y}^{(M)}\} may subsequently be analyzed using statistical models other than Equation (1) that require fully observed outcome data. These downstream analyses may additionally adjust for fully observed covariates, including those contained in Z\mathbf{Z}, depending on the scientific objective. In such settings, inference can be conducted using Rubin’s rules [35].

Remark 2.1. It is worth emphasizing that the posterior-based multiple imputation representation serves a different purpose from posterior inference under the proposed Bayesian model. For the multivariate regression model in Equation (1), inference for the model parameters is conducted directly from the joint posterior distribution, without requiring a separate multiple imputation stage. Consequently, parameter estimation and uncertainty quantification are performed within a single Bayesian framework, avoiding the need to specify separate imputation and analysis models. The completed datasets are instead intended for downstream analyses based on alternative statistical models that require fully observed outcome data.

Theoretical Properties of the SHIM Prior

We provide two theoretical characterizations of the SHIM prior. First, we characterize the dependence among coefficient magnitudes induced by the multiplicative shrinkage hierarchy. The following result shows that this dependence is determined by the number of shrinkage components shared by two coefficients, quantifying how the prior encourages related coefficients to exhibit similar magnitudes while retaining coefficient-specific variation.

Proposition 2.1. Consider a multiplicative hierarchical shrinkage prior with independent positive random variables {Hℓ}\{H_{\ell}\}. For coefficient aa, let Pa\mathcal{P}_{a} be the finite set of hierarchical components acting on that coefficient, and suppose

βa=saZa,sa=∏ℓ∈PaHℓ.\beta_a = s_a Z_a,\qquad s_a = \prod_{\ell\in\mathcal{P}_a} H_{\ell}.

*where ZaZ_a is a standard Gaussian variable that is independent of the shrinkage hierarchy. Further, the Gaussian variables are independent across distinct parameters. Assume 0<Var⁡{log⁡(Hℓ)}<∞0 < \operatorname{Var}\{\log(H_{\ell})\} < \infty for every ℓ\ell. Then, for two distinct coefficients a≠ba \ne b, we have

Cov⁡{log⁡∣βa∣,log⁡∣βb∣}=∑ℓ∈Pa∩PbVar⁡{log⁡(Hℓ)}andVar⁡{log⁡∣βa∣}=∑ℓ∈PaVar⁡{log⁡(Hℓ)}+π28.\operatorname{Cov}\{\log|\beta_a|,\log|\beta_b|\}=\sum_{\ell\in\mathcal{P}_a\cap\mathcal{P}_b}\operatorname{Var}\{\log(H_\ell)\} \quad\text{and}\quad \operatorname{Var}\{\log|\beta_a|\}=\sum_{\ell\in\mathcal{P}_a}\operatorname{Var}\{\log(H_\ell)\}+\frac{\pi^2}{8}.

The identities in Proposition 2.1 directly follow from

log⁡∣βa∣=∑ℓ∈Palog⁡(Hℓ)+log⁡∣Za∣\log|\beta_a|=\sum_{\ell\in\mathcal{P}_a}\log(H_\ell)+\log|Z_a|

and the independence of the primitive shrinkage components and Gaussian variables, together with the known result of Var⁡(log⁡∣Za∣)=π2/8\operatorname{Var}(\log|Z_a|)=\pi^2/8. A detailed derivation is omitted.

For SHIM, the prior scale associated with βkq\beta_{kq} is skq=τγm(k)δm(k),g(k)λkξkqs_{kq}=\tau\gamma_{m(k)}\delta_{m(k),g(k)}\lambda_k\xi_{kq}. For a standard half-Cauchy variable HH, we know Var⁡(log⁡H)=π2/4\operatorname{Var}(\log H)=\pi^2/4. Because every coefficient-specific scale contains five independent half-Cauchy components, we get

Var⁡{log⁡∣βkq∣}=5π24+π28=11π28.\operatorname{Var}\{\log|\beta_{kq}|\}=\frac{5\pi^2}{4}+\frac{\pi^2}{8}=\frac{11\pi^2}{8}.

In what follows, we have

Corr⁡{log⁡∣βkq∣,log⁡∣βk′q′∣}={8/11,k=k′, q≠q′,6/11,k≠k′, m(k)=m(k′), g(k)=g(k′),4/11,m(k)=m(k′), g(k)≠g(k′),2/11,m(k)≠m(k′).\operatorname{Corr}\{\log|\beta_{kq}|,\log|\beta_{k'q'}|\}= \begin{cases} 8/11, & k=k',\ q\ne q',\\ 6/11, & k\ne k',\ m(k)=m(k'),\ g(k)=g(k'),\\ 4/11, & m(k)=m(k'),\ g(k)\ne g(k'),\\ 2/11, & m(k)\ne m(k'). \end{cases}

These results imply that coefficients sharing more levels of the hierarchy have more strongly associated prior magnitudes. The shared global, modality, group, and predictor-specific scales induce this dependence, while the independent Gaussian variables and outcome-specific scales allow individual coefficients to retain distinct magnitudes and signs. This dependence provides a mechanism through which posterior learning about a shared scale can propagate across related coefficients.

We next characterize the marginal prior density induced when the conditional standard deviation of a Gaussian coefficient is the product of independent half-Cauchy variables. The result shows how the number of multiplicative components determines how rapidly the density diverges near zero and decays in the tails.

Theorem 2.1. Let K≥1K\ge1 be an integer, and define

H1,H2,…,HK∼indC+(0,1),SK=∏ℓ=1KHℓ.H_1,H_2,\ldots,H_K\overset{\mathrm{ind}}{\sim} C^+(0,1),\qquad S_K=\prod_{\ell=1}^{K}H_\ell.

Suppose β∣SK∼N(0,SK2)\beta\mid S_K\sim N(0,S_K^2) and let πK\pi_K denote the marginal density of β\beta. Then, as ∣β∣↓0|\beta|\downarrow0, we have

πK(β)∼C0,K{log⁡(1∣β∣)}KwithC0,K=(2/π)K2πK!.\pi_K(\beta)\sim C_{0,K}\left\{\log\left(\frac{1}{|\beta|}\right)\right\}^{K} \quad\text{with}\quad C_{0,K}=\frac{(2/\pi)^K}{\sqrt{2\pi}K!}.

Moreover, as ∣β∣→∞|\beta|\to\infty, we have

πK(β)∼C∞,K{log⁡∣β∣}K−1β2withC∞,K=(2/π)K2π(K−1)!.\pi_K(\beta)\sim C_{\infty,K}\frac{\{\log|\beta|\}^{K-1}}{\beta^2} \quad\text{with}\quad C_{\infty,K}=\frac{(2/\pi)^K}{\sqrt{2\pi}(K-1)!}.

The proof of Theorem 2.1 is provided in Appendix B. For SHIM, the coefficient-specific scale is the product of K=5K = 5 independent half-Cauchy variables. The theorem therefore gives

πSHIM(β)∼{415π52π{log⁡(1∣β∣)}5,as ∣β∣↓0,43π52π{log⁡∣β∣}4β2,as ∣β∣→∞.\pi_{\mathrm{SHIM}}(\beta) \sim\begin{cases} \dfrac{4}{15\pi^{5}\sqrt{2\pi}}\left\{\log\left(\dfrac{1}{|\beta|}\right)\right\}^{5}, & \text{as } |\beta| \downarrow0, \\[6pt] \dfrac{4}{3\pi^{5}\sqrt{2\pi}}\dfrac{\{\log|\beta|\}^{4}}{\beta^{2}}, & \text{as } |\beta| \to\infty. \end{cases}

Thus, the multilevel construction produces an integrable fifth-order logarithmic singularity at zero and a heavy polynomial tail modified by a fourth-order logarithmic factor. The density is continuous on R∖{0}\mathbb{R} \setminus\{0\} but diverges as ∣β∣↓0|\beta| \downarrow0, assigning substantial density to coefficients near zero. In the tails, πSHIM(β)\pi_{\mathrm{SHIM}}(\beta) has the same polynomial decay exponent as a Cauchy density, while the factor {log⁡∣β∣}4\{\log|\beta|\}^{4} makes it asymptotically heavier than a standard Cauchy density. These features encourage strong shrinkage of small coefficients while maintaining substantial prior support for large effects.

Simulations

Data Generation and Simulation Setup

This section presents simulation studies designed to evaluate the variable selection performance of the proposed method in comparison with several competing approaches. Motivated from the Vanderbilt Memory and Aging Project [27], we generated n=300n = 300 independent observations with P∈{200,400}P \in\{200, 400\} predictors and Q=2Q = 2 continuous outcomes. The predictors were organized according to a two-level hierarchical structure consisting of M=2M = 2 modalities, each containing G1=G2=4G_{1} = G_{2} = 4 groups of predictors. The two modalities contained equal numbers of predictors, and all groups were of equal size. That is, for P=200P = 200, each modality contained 100 predictors and each group contained 25 predictors; for P=400P = 400, each modality contained 200 predictors and each group contained 50 predictors.

To consider predictors with a hierarchical correlation structure, the rows of X\mathbf{X} were generated independently according to xi∼NP(0,ΣX)\mathbf{x}_{i} \sim N_{P}(0, \mathbf{\Sigma}_{X}). Specifically, the covariance between predictors kk and k′k' was specified as

Cov⁡(xik,xik′)=σmod2I{m(k)=m(k′)}+σgrp2I{m(k)=m(k′),g(k)=g(k′)}+σres2I{k=k′}.\operatorname{Cov}(x_{ik},x_{ik'}) = \sigma_{\mathrm{mod}}^{2}\mathbb{I}\{m(k)=m(k')\} + \sigma_{\mathrm{grp}}^{2}\mathbb{I}\{m(k)=m(k'),g(k)=g(k')\} + \sigma_{\mathrm{res}}^{2}\mathbb{I}\{k=k'\}.

where m(k)m(k) and g(k)g(k) respectively denote the modality and group membership of predictor kk, and I{⋅}\mathbb{I}\{\cdot\} denotes the standard indicator function. In addition, σmod2\sigma_{\mathrm{mod}}^{2} represents the shared modality-level variation, σgrp2\sigma_{\mathrm{grp}}^{2} represents the additional variation shared by predictors within the same group, and σres2\sigma_{\mathrm{res}}^{2} denotes the residual variance. Under this specification, predictors within the same group shared both modality- and group-level variation, predictors in different groups within the same modality shared only modality-level variation, and predictors from different modalities were independent. We further set σmod=0.8\sigma_{\mathrm{mod}} = 0.8, σgrp=0.6\sigma_{\mathrm{grp}} = 0.6, and σres=0.8\sigma_{\mathrm{res}} = 0.8, inducing correlations of approximately 0.6 between distinct predictors within the same group and 0.4 between predictors in different groups within the same modality. Following generation, each predictor was standardized to have sample mean zero and unit variance to mitigate the influence of predictor scale on variable selection [32].

The outcome was generated according to Y=1nαT+XB+EY = \mathbf{1}_{n}\alpha^{\mathsf{T}} + X B + E, where α\alpha was set to (0.4,0.4)T(0.4, 0.4)^{\mathsf{T}}, and the rows of EE were independently generated from NQ(0,ΣY)N_{Q}(0, \mathbf{\Sigma}_{Y}) with

ΣY=(10.30.31)\mathbf{\Sigma}_{Y} = \begin{pmatrix} 1 & 0.3 \\ 0.3 & 1 \end{pmatrix}

which corresponds to a residual correlation of 0.3 between the two outcomes. For simplicity, the simulation studies did not include low-dimensional adjustment covariates Z\mathbf{Z}. Since these covariates are treated as non-penalized components, the Gaussian update for Ψ\mathbf{\Psi} is conjugate and does not alter the hierarchical shrinkage mechanism or variable-selection properties of the proposed prior on B\mathbf{B}. This design choice also ensured a fair comparison among all competing methods described in Section 3.2, as not all methods allow adjustment covariates to be included as non-penalized terms.

The P×QP \times Q coefficient matrix B\mathbf{B} was assumed to be sparse. In the benchmark setting, four predictors were assigned nonzero coefficients with two signal strengths. Specifically, two predictors had coefficient vectors (1.2,1.2)T(1.2, 1.2)^{\mathsf{T}}, whereas the remaining two had coefficient vectors (0.6,0.6)T(0.6, 0.6)^{\mathsf{T}}. All other entries of B\mathbf{B} were set to zero. To evaluate the impact of the modality-level hierarchy, we considered two configurations for the allocation of the nonzero coefficients:

  1. Under the “1M2G” configuration, the four active predictors were distributed across two groups within the same modality, with one group containing the two stronger signals and the other containing the two weaker signals.

  1. Under the “2M2G” configuration, the active predictors were distributed across two groups belonging to two different modalities, with one group containing the two stronger signals and the other containing the two weaker signals.

Finally, missing values were imposed on the generated outcomes. In the benchmark setting, we considered a missing-at-random (MAR) mechanism, under which the probability that outcome qq was missing depended on a fully observed predictor through a logistic regression model. For each simulated dataset, the intercept of the logistic regression model was calibrated to achieve a target proportion πmiss∈{0.2,0.6}\pi_{\mathrm{miss}} \in\{0.2, 0.6\} of subjects with at least one missing outcome, corresponding to moderate and high levels of missingness, respectively.

Table 1 summarizes the eight benchmark simulation settings. Each setting was replicated 100 times. Additional implementation details for the Bayesian methods, including the number of burn-in iterations, posterior samples retained for inference, and MCMC convergence diagnostics, will be described subsequently. SHIM was implemented using the open-source R package shim, which also documents random number seeds and other information necessary to reproduce all simulations reported in this section.

DimensionP=200P = 200P=400P = 400
Configuration1M2G2M2G1M2G2M2G
πmiss\pi_{\mathrm{miss}}0.20.60.20.6
0.20.60.20.6
ScenarioS1S2S3S4
S5S6S7S8

Table 1. Table 1. Benchmark simulation settings (8 scenarios): Unless otherwise specified, all settings used n=300n = 300, a MAR missing-data mechanism, and four active predictors.

Methods for Comparison and Implementation Details

We compared SHIM with several competing methods representing different strategies for modeling multivariate outcomes and performing variable selection. First, to evaluate the benefit of jointly modeling multiple outcomes, we implemented a univariate counterpart of SHIM, denoted by SHIM_U, which applies the proposed hierarchical shrinkage model separately to each outcome while retaining the same multilevel shrinkage structure. We further included two modern Bayesian approaches: Group Inverse-Gamma Gamma (GIGG) [6] and Multivariate Bayesian Sparse Group Selection with Spike-and-Slab Priors (MBSGSSS) [24]. GIGG employs continuous shrinkage through inverse-gamma–gamma (IG–Gamma) priors, with the horseshoe prior arising as a special case. MBSGSSS is a multivariate extension of the Bayesian sparse group selection method based on spike-and-slab priors [44]. For all Bayesian methods, including SHIM, SHIM_U, GIGG, and MBSGSSS, we ran a total of 2,000 MCMC iterations and discarded the first 500 iterations as burn-in. These settings were chosen based on convergence diagnostics, in which trace plots of the active regression coefficients exhibited stable mixing after burn-in and autocorrelations decayed sufficiently rapidly. Representative trace plots are provided in Appendix C.

Lasso [40] and group lasso [46] were included as conventional frequentist regularization methods. Since both methods were originally developed for univariate outcomes, they were applied separately to each outcome. Lasso imposes coefficient-wise regularization and therefore performs variable selection at the individual level, whereas group lasso (glasso) imposes group-wise regularization and performs variable selection at the group level. Because both methods produce shrinkage estimates for the selected predictors, we further refitted an ordinary least squares (OLS) regression model using the selected predictors to obtain post-selection point estimates with reduced bias [3].

Except for SHIM and SHIM_U, the competing methods do not directly accommodate incomplete outcomes. These methods were therefore fitted using a common complete-case sample consisting of observations with all observed data. In contrast, SHIM and SHIM_U were fitted using all available observations through the proposed formulation for incomplete outcomes. Furthermore, GIGG, lasso, and glasso do not jointly model multivariate outcomes and were consequently fitted separately to each outcome. Finally, although glasso, GIGG, and MBSGSSS incorporate group-level structure, they do not accommodate the additional modality-level hierarchy considered by SHIM. Accordingly, only group membership information, but not modality membership information, was provided to these methods.

Variable selection was determined according to the estimation framework of each method. For continuous shrinkage methods, including SHIM, SHIM_U, and GIGG, a predictor was considered selected for a given outcome if the corresponding 95% marginal posterior credible interval excluded zero [6, 30]. For MBSGSSS, a predictor was considered selected if the posterior median of its regression coefficient was nonzero, following the general recommendation for spike-and-slab priors [24]. For lasso and glasso, predictors with nonzero estimated regression coefficients were considered selected. Table 2 summarizes the competing methods and their main characteristics.

MethodModelingShrinkage Hierarchy
SHIMMultivariateMulti-group
SHIM_UUnivariateMulti-group
LassoUnivariateIndividual
GlassoUnivariateGroup
GIGGUnivariateGroup
MBSGSSSMultivariateGroup

Table 2. Summary of the competing methods and their basic features.

Evaluation Criteria

Performance was evaluated with respect to variable selection and coefficient estimation. For all penalized methods (except MLM), we summarized variable-selection performance separately for each outcome using the true positive rate (TPR; sensitivity), the false positive rate (FPR), the false discovery rate (FDR; 1−precision1 - \text{precision}), and the F1 score. The TPR was defined as the proportion of truly active predictors that were correctly selected, whereas the FPR was defined as the proportion of truly inactive predictors that were incorrectly selected. The FDR was defined as the proportion of selected predictors that were false positives, and the F1 score was defined as the harmonic mean of sensitivity and precision.

We further evaluated coefficient estimation performance for the four active predictors. We calculated the absolute percentage bias and empirical standard deviation of the coefficient estimates across simulation replicates. The empirical standard deviation was used to assess the efficiency of the estimators. To evaluate the calibration of uncertainty quantification, we calculated the ratio of the empirical standard deviation to the average standard error across simulation replicates. Finally, coverage probability was defined as the proportion of 95% credible or confidence intervals containing the corresponding true coefficient and was compared with the nominal 95% level.

Simulation Results

Table 3 summarizes the variable-selection performance across the eight simulation settings. Overall, SHIM consistently achieved the best or nearly best performance across all scenarios, maintaining high TPR together with near-zero FPR and FDR, resulting in F1 scores close to one. Compared with its univariate counterpart SHIM_U, SHIM showed slightly improved TPR and F1 scores in the more challenging settings with higher missingness (S2, S4, S6, and S8), suggesting that jointly modeling correlated outcomes provides additional gains when information is limited. Among the competing methods, GIGG effectively controlled false positives but experienced a noticeable reduction in TPR under the more challenging scenarios, whereas MBSGSSS maintained high TPR and consistently achieved F1 scores comparable to those of SHIM. In contrast, lasso selected substantially more inactive predictors, resulting in substantially inflated FDR despite perfect or nearly perfect TPR, whereas glasso presented extremely high FPR and FDR and consequently poor F1 scores. Overall, these results demonstrate that SHIM provides an effective balance between sensitivity and false positive control, yielding robust variable-selection performance across a wide range of simulation settings.

S1S2
Outcome 1Outcome 2Outcome 1Outcome 2
TPRFPRFDRF1TPRFPRFDRF1TPRFPRFDRF1TPRFPRFDRF1
SHIM1.000.000.010.991.000.000.010.990.980.000.010.980.990.000.020.99
SHIM_U0.990.000.010.991.000.000.020.990.960.000.000.970.980.000.010.98
Lasso1.000.030.520.631.000.030.510.641.000.030.530.621.000.030.550.61
Glasso1.000.440.950.091.000.450.950.091.000.380.940.111.000.350.940.11
GIGG0.990.000.010.991.000.000.010.990.870.000.000.920.860.000.010.92
MBSGSSS1.000.000.030.981.000.000.030.990.990.000.060.971.000.000.060.97
S3S4
Outcome 1Outcome 2Outcome 1Outcome 2
TPRFPRFDRF1TPRFPRFDRF1TPRFPRFDRF1TPRFPRFDRF1
SHIM1.000.000.020.991.000.000.011.000.980.000.010.990.980.000.010.98
SHIM_U1.000.000.010.991.000.000.011.000.960.000.000.980.970.000.000.98
Lasso1.000.020.480.671.000.020.500.660.990.040.550.601.000.030.530.63
Glasso1.000.470.950.091.000.470.950.091.000.390.940.111.000.410.950.10
GIGG0.990.000.010.991.000.000.010.990.840.000.010.900.850.000.010.91
MBSGSSS1.000.000.010.991.000.000.020.990.990.000.040.980.990.000.040.97
S5S6
Outcome 1Outcome 2Outcome 1Outcome 2
TPRFPRFDRF1TPRFPRFDRF1TPRFPRFDRF1TPRFPRFDRF1
SHIM0.990.000.030.981.000.000.020.990.980.000.030.970.980.000.030.97
SHIM_U0.990.000.020.981.000.000.001.000.930.000.020.950.940.000.020.96
Lasso1.000.020.600.561.000.020.600.560.990.020.610.551.000.020.620.54
Glasso1.000.330.970.061.000.340.970.061.000.350.970.061.000.360.970.06
GIGG0.980.000.000.991.000.000.001.000.820.000.020.880.820.000.010.89
MBSGSSS1.000.000.020.991.000.000.020.990.990.000.050.970.990.000.060.96
S7S8
Outcome 1Outcome 2Outcome 1Outcome 2
TPRFPRFDRF1TPRFPRFDRF1TPRFPRFDRF1TPRFPRFDRF1
SHIM1.000.000.030.991.000.000.020.990.970.000.040.960.970.000.030.97
SHIM_U1.000.000.010.990.990.000.010.990.940.000.010.960.950.000.020.96
Lasso1.000.020.560.591.000.020.570.581.000.020.640.511.000.020.650.51
Glasso1.000.410.970.051.000.400.970.051.000.410.970.051.000.420.970.05
GIGG0.990.000.000.990.990.000.000.990.810.000.010.880.810.000.010.88
MBSGSSS1.000.000.020.991.000.000.020.990.990.000.040.970.990.000.040.97

Table 3. Average variable-selection performance over 100 simulation replicates. Results are reported separately for each outcome under the eight simulation settings. Performance is evaluated using the true positive rate (TPR), false positive rate (FPR), false discovery rate (FDR), and F1 score..

Figure 1 displays the distribution of the F1 scores across the 100 simulation replicates. Consistent with the average results in Table 3, SHIM achieved the highest median F1 scores with consistently small variability across nearly all simulation settings, implying stable variable-selection performance. MBSGSSS also demonstrated competitive performance, with F1 distributions that were generally comparable to those of SHIM. In contrast, GIGG showed a progressive decline in F1 scores as the simulation settings became more difficult, whereas lasso and glasso consistently yielded substantially lower F1 scores due to a large amount of false discoveries.

Distribution of F1 scores across 100 simulation replicates under the eight simulation settings

Figure 1. Distribution of F1 scores across 100 simulation replicates under the eight simulation settings.

In addition, Figure 2 presents the absolute percentage bias and coverage probabilities for the active regression coefficients. Across all methods, estimation was generally more accurate for the stronger signals (β1\beta_1 and β2\beta_2) than for the weaker signals (β3\beta_3 and β4\beta_4), particularly under high missingness (S2, S4, S6, and S8). SHIM, SHIM_U, and MBSGSSS achieved comparably low bias for the stronger coefficients. For the weaker signals, bias increased for all three methods, although SHIM consistently yielded lower bias than its univariate counterpart SHIM_U, particularly under the more challenging settings with larger predictor dimensions and higher missingness. MBSGSSS also demonstrated competitive performance and generally produced bias comparable to, or slightly smaller than, that of SHIM. In contrast, GIGG exhibited substantially larger bias for the weaker signals under high missingness, whereas lasso and glasso showed considerably larger bias across most simulation settings even though post-selection ordinary least squares estimation was implemented.

Figure 2. Figure 2. Absolute percentage bias (top) and coverage probability (bottom) for the four active coefficients across the eight simulation settings. Coefficients β1\beta_1 and β2\beta_2 correspond to the two stronger signals (true coefficient = 1.2), whereas β3\beta_3 and β4\beta_4 correspond to the two weaker signals (true coefficient = 0.6). For lasso and group lasso, the reported summaries are based on post-selection ordinary least squares estimates.

The coverage results followed similar patterns. SHIM, SHIM_U, and MBSGSSS generally achieved coverage probabilities close to the nominal 95% level across both outcomes, although some under-coverage occurred for the weaker coefficients under high missingness. Compared with SHIM_U, SHIM more consistently maintained nominal coverage for the weaker signals in the more challenging settings. GIGG achieved approximately nominal coverage for the stronger coefficients but exhibited noticeable under-coverage for some weaker coefficients, particularly under high missingness. Lasso showed systematic under-coverage across most simulation settings. Although group lasso frequently attained near-nominal coverage among the available estimates, these results were accompanied by substantially greater empirical variability and less reliable standard error calibration; additional results on the empirical standard deviation and the ratio of empirical to estimated standard errors are provided in Appendix D. Overall, SHIM provided accurate coefficient estimation together with well-calibrated interval estimation across a wide range of simulation settings.

Sensitivity Analysis

To evaluate the robustness of the proposed framework under alternative data-generating mechanisms, we considered four additional simulation settings obtained by modifying S6 in Section 3.1, detailed as follows:

  1. High outcome correlation (CORR): We increased the residual correlation between the two outcomes from 0.3 to 0.7. This setting evaluates the impact of stronger dependence among psychometric outcomes on variable selection and estimation.

  1. Missing not at random (MNAR): We considered an MNAR mechanism under which the probability that an outcome was missing depended on its potentially unobserved value together with the value of the other outcome. The residual correlation between the two outcomes was set to 0.7. This setting assesses the robustness of the competing methods when the MAR assumption is violated.

  1. Outcome-specific sparsity (OUT): We allowed the two outcomes to have different sets of active predictors. Specifically, the two stronger signals had coefficient vectors (1.2,0)⊤(1.2, 0)^\top and (0,1.2)⊤(0, 1.2)^\top, whereas the two weaker signals had coefficient vectors (0.6,0)⊤(0.6, 0)^\top and (0,0.6)⊤(0, 0.6)^\top. This setting evaluates the ability of the method to distinguish outcome-specific associations from predictors with shared effects across outcomes.

  1. Ultra-high dimensionality (ULTRA): We increased the number of predictors to P=600P = 600, with two modalities containing 300 predictors each and each modality further divided into four groups of 75 predictors. This setting evaluates the scalability of the competing methods under increasingly high-dimensional predictors.

Except for the modifications described above, all remaining data-generating features were identical to those of S6, including πmiss=0.6\pi_{\mathrm{miss}} = 0.6. The same variable-selection and coefficient-estimation metrics used in the primary simulation study were calculated for each sensitivity setting.

The variable-selection results under CORR, MNAR, and ULTRA were generally consistent with those from the primary simulations. SHIM, SHIM_U, and MBSGSSS continued to achieve comparably high F1 scores, whereas GIGG exhibited lower TPR values, and lasso and glasso continued to yield inflated FDR values; see Table 9 in Appendix E for details. Under OUT, however, MBSGSSS no longer performed comparably to SHIM. As shown in Table 4, its average F1 scores decreased substantially, primarily because of increased FDR. This deterioration likely reflects the predictor-driven selection structure of MBSGSSS, under which sparsity is largely shared across outcomes and therefore may not adequately distinguish predictors associated with only a subset of responses. In contrast, SHIM incorporates outcome-specific local shrinkage and maintained strong selection performance under this particular scenario.

MethodOutcome 1 TPROutcome 1 FPROutcome 1 FDROutcome 1 F1Outcome 2 TPROutcome 2 FPROutcome 2 FDROutcome 2 F1
SHIM0.960.000.060.940.960.000.040.95
SHIM_U0.950.000.010.960.940.000.010.95
Lasso1.000.010.480.650.990.010.530.60
Glasso0.970.330.980.030.980.370.990.03
GIGG0.840.000.000.890.800.000.010.86
MBSGSSS0.950.010.510.640.940.010.520.63

Table 4. Average variable-selection performance under the outcome-specific sparsity (OUT) setting over 100 simulation replicates. Performance is evaluated using the true positive rate (TPR), false positive rate (FPR), false discovery rate (FDR), and F1 score.

CORRMNAR
Outcome 1Outcome 2Outcome 1Outcome 2
TPRFPRFDRF1TPRFPRFDRF1TPRFPRFDRF1TPRFPRFDRF1
SHIM0.980.000.010.980.980.000.010.980.940.000.010.960.950.000.010.97
SHIM_U0.940.000.000.960.940.000.020.960.920.000.030.940.910.000.020.94
Lasso0.990.020.610.551.000.020.620.540.980.020.630.520.980.020.650.50
Glasso1.000.350.970.061.000.370.970.061.000.350.970.061.000.360.970.06
GIGG0.820.000.010.890.820.000.010.890.770.000.010.850.770.000.010.86
MBSGSSS0.960.000.060.950.960.000.060.950.950.000.060.940.950.000.060.94
OUTULTRA
Outcome 1Outcome 2Outcome 1Outcome 2
TPRFPRFDRF1TPRFPRFDRF1TPRFPRFDRF1TPRFPRFDRF1
SHIM0.960.000.060.940.960.000.040.950.950.000.040.950.960.000.040.95
SHIM_U0.950.000.010.960.940.000.010.950.910.000.010.940.920.000.010.95
Lasso1.000.010.480.650.990.010.530.600.990.020.650.500.990.020.660.50
Glasso0.970.330.980.030.980.370.990.031.000.330.980.041.000.320.980.04
GIGG0.840.000.000.890.800.000.010.860.800.000.010.870.800.000.010.88
MBSGSSS0.950.010.510.640.940.010.520.630.980.000.040.960.980.000.040.97

Table 9. Average variable-selection performance over 100 simulation replicates. Results are reported separately for each outcome under the four sensitivity analysis settings. Performance is evaluated using the true positive rate (TPR), false positive rate (FPR), false discovery rate (FDR), and F1 score.

Coefficient-estimation patterns were generally similar to those in the primary simulations, although the OUT setting more clearly differentiated SHIM, SHIM_U, and MBSGSSS. As shown in Table 5, MBSGSSS exhibited greater bias for weak signals, higher empirical variability, and more pronounced under-coverage. Its ratios of empirical to average estimated standard errors were also noticeably above one, suggesting that the estimated standard errors understated the sampling variability. This pattern may reflect difficulty in accommodating outcome-specific associations when selection of MBSGSSS is primarily structured at the predictor level. In contrast, SHIM and SHIM_U generally provided accurate estimates with low empirical variability, although both showed increased bias for weak signals. SHIM also demonstrated a modest advantage over SHIM_U in coverage and standard error calibration. Among the remaining methods, GIGG continued to exhibit greater bias and under-coverage for weak signals, lasso showed substantially greater bias and under-coverage across active coefficients, and glasso produced obviously greater empirical variability.

MethodPB (%) β1\beta_1PB (%) β2\beta_2PB (%) β3\beta_3PB (%) β4\beta_4ESD β1\beta_1ESD β2\beta_2ESD β3\beta_3ESD β4\beta_4ESD / SE β1\beta_1ESD / SE β2\beta_2ESD / SE β3\beta_3ESD / SE β4\beta_4CP (%) β1\beta_1CP (%) β2\beta_2CP (%) β3\beta_3CP (%) β4\beta_4
SHIM2.67.40.120.140.941.039693
SHIM_U2.58.10.110.140.901.099490
Lasso13.324.80.170.141.311.177369
Glasso3.513.30.420.461.031.139487
GIGG3.315.80.160.191.041.139390
MBSGSSS3.813.60.150.221.011.379388
SHIM3.811.60.120.141.001.049491
SHIM_U3.612.60.120.141.021.099390
Lasso14.928.50.150.161.241.376864
Glasso5.77.00.820.502.051.249888
GIGG4.621.90.130.220.961.329485
MBSGSSS4.719.60.120.230.921.449485

Table 5. Estimation performance for the active predictors under OUT. Coefficients β1\beta_1 and β2\beta_2 correspond to the two stronger signals (true coefficient = 1.2), whereas β3\beta_3 and β4\beta_4 correspond to the two weaker signals (true coefficient = 0.6). Coefficients β1\beta_1 and β3\beta_3 are only associated with Outcome 1, whereas coefficients β2\beta_2 and β4\beta_4 are only associated with Outcome 2.

Across the other sensitivity settings, SHIM generally remained competitive in terms of bias and maintained relatively low empirical variability, particularly when the outcomes were highly correlated. Its performance deteriorated under MNAR, however, with increased bias and under-coverage. Under ULTRA, SHIM continued to perform reasonably well, although MBSGSSS appeared comparatively stronger. Full results and further discussion are provided in Table 9 and Table 10 of Appendix E.

Table 10. Estimation performance for the active predictors under CORR, MNAR and ULTRA.
CORROutcome 1
MethodPB (%) β1\beta_1PB (%) β2\beta_2PB (%) β3\beta_3PB (%) β4\beta_4ESD β1\beta_1ESD β2\beta_2ESD β3\beta_3ESD β4\beta_4ESD / SE β1\beta_1ESD / SE β2\beta_2ESD / SE β3\beta_3ESD / SE β4\beta_4CP (%) β1\beta_1CP (%) β2\beta_2CP (%) β3\beta_3CP (%) β4\beta_4
SHIM3.01.95.514.90.110.110.120.120.920.970.950.9395959893
SHIM_U2.72.06.216.70.120.120.130.140.960.961.001.0495979191
Lasso15.914.825.033.90.180.140.140.161.211.051.031.2175747967
Glasso1.516.612.828.60.440.790.450.531.061.891.121.3795959395
GIGG3.62.912.625.40.180.140.200.211.110.951.131.2291959086
MBSGSSS2.82.64.416.20.150.130.190.211.020.931.201.3594939591
CORROutcome 2
MethodPB (%) β1\beta_1PB (%) β2\beta_2PB (%) β3\beta_3PB (%) β4\beta_4ESD β1\beta_1ESD β2\beta_2ESD β3\beta_3ESD β4\beta_4ESD / SE β1\beta_1ESD / SE β2\beta_2ESD / SE β3\beta_3ESD / SE β4\beta_4CP (%) β1\beta_1CP (%) β2\beta_2CP (%) β3\beta_3CP (%) β4\beta_4
SHIM3.23.36.212.40.120.110.130.120.990.981.040.9493939493
SHIM_U3.13.56.114.10.130.120.140.151.051.011.051.0794939489
Lasso15.517.026.231.40.170.150.150.151.161.081.071.1477698174
Glasso3.712.68.013.30.660.910.520.531.622.291.311.3486959295
GIGG3.04.612.922.60.160.140.210.221.010.951.151.2996919183
MBSGSSS2.63.65.213.90.150.130.190.220.990.951.221.3792949187
MNAROutcome 1
MethodPB (%) β1\beta_1PB (%) β2\beta_2PB (%) β3\beta_3PB (%) β4\beta_4ESD β1\beta_1ESD β2\beta_2ESD β3\beta_3ESD β4\beta_4ESD / SE β1\beta_1ESD / SE β2\beta_2ESD / SE β3\beta_3ESD / SE β4\beta_4CP (%) β1\beta_1CP (%) β2\beta_2CP (%) β3\beta_3CP (%) β4\beta_4
SHIM6.65.910.520.50.110.120.130.150.931.001.001.1092919483
SHIM_U6.36.410.321.30.120.120.140.170.971.011.071.1993929183
Lasso21.322.832.339.10.160.140.150.181.191.011.141.3055546756
Glasso8.78.513.04.80.370.370.480.360.931.001.280.9698988598
GIGG9.310.823.631.50.170.130.220.231.120.881.261.3889898376
MBSGSSS8.19.614.720.50.140.120.220.221.030.891.351.3688898688
MNAROutcome 2
MethodPB (%) β1\beta_1PB (%) β2\beta_2PB (%) β3\beta_3PB (%) β4\beta_4ESD β1\beta_1ESD β2\beta_2ESD β3\beta_3ESD β4\beta_4ESD / SE β1\beta_1ESD / SE β2\beta_2ESD / SE β3\beta_3ESD / SE β4\beta_4CP (%) β1\beta_1CP (%) β2\beta_2CP (%) β3\beta_3CP (%) β4\beta_4
SHIM6.46.614.818.10.120.120.140.131.020.981.061.0190919189
SHIM_U6.77.216.219.70.130.120.160.161.040.951.111.1788918986
Lasso20.222.533.836.50.180.150.160.161.271.121.171.2154496965
Glasso6.411.01.23.20.370.410.430.410.911.091.121.0392929098
GIGG8.811.124.227.90.170.150.210.211.100.981.221.2185838383
MBSGSSS7.710.015.818.50.150.130.210.221.050.981.311.3690878688
ULTRAOutcome 1
MethodPB (%) β1\beta_1PB (%) β2\beta_2PB (%) β3\beta_3PB (%) β4\beta_4ESD β1\beta_1ESD β2\beta_2ESD β3\beta_3ESD β4\beta_4ESD / SE β1\beta_1ESD / SE β2\beta_2ESD / SE β3\beta_3ESD / SE β4\beta_4CP (%) β1\beta_1CP (%) β2\beta_2CP (%) β3\beta_3CP (%) β4\beta_4
SHIM3.62.69.811.10.140.140.180.151.051.091.281.0591898991
SHIM_U3.92.512.211.20.140.130.190.151.081.071.361.0792908693
Lasso16.216.734.729.00.170.150.170.141.121.131.251.0375637176
GIGG3.03.221.216.30.160.160.250.231.021.051.471.2792908187
MBSGSSS1.72.410.40.90.150.140.210.151.021.021.451.0592938895
ULTRAOutcome 2
MethodPB (%) β1\beta_1PB (%) β2\beta_2PB (%) β3\beta_3PB (%) β4\beta_4ESD β1\beta_1ESD β2\beta_2ESD β3\beta_3ESD β4\beta_4ESD / SE β1\beta_1ESD / SE β2\beta_2ESD / SE β3\beta_3ESD / SE β4\beta_4CP (%) β1\beta_1CP (%) β2\beta_2CP (%) β3\beta_3CP (%) β4\beta_4
SHIM2.40.914.04.60.120.140.170.140.931.071.191.0296948895
SHIM_U2.11.115.74.60.120.130.180.140.901.091.251.0494918796
Lasso17.716.636.924.40.160.170.160.151.101.221.151.0673746682
GIGG3.42.024.68.80.170.160.240.221.031.071.361.2393919191
MBSGSSS2.61.412.72.70.140.130.200.150.981.001.401.0093949095

Table 10.

Posterior-Based Multiple Imputation for Downstream Analysis

To illustrate the use of posterior-based multiple imputation for downstream analysis, we conducted an additional simulation in which a partially observed variable WW was subsequently used as a predictor of a fully observed outcome YY. Data were generated according to

W=αW1n+Xβ+ZψW+εWandY=αY1n+κW+ZψY+εYW = \alpha_W 1_n + X\beta+ Z\psi_W + \varepsilon_W \qquad\text{and} \qquad Y = \alpha_Y 1_n + \kappa W + Z\psi_Y + \varepsilon_Y

with εW,εY∼Nn(0,In)\varepsilon_W,\varepsilon_Y \sim N_n(0,I_n), αW=αY=0.4\alpha_W = \alpha_Y = 0.4, and κ=0.8\kappa= 0.8 representing the association of interest. The high-dimensional predictors XX and their coefficients were generated following S6 (from Section 3.1), while ZZ contained one standardized continuous covariate and one binary covariate with associated coefficients set to ψW=(0.5,0.3)⊤\psi_W = (0.5, 0.3)^\top and ψY=(0.6,0.4)⊤\psi_Y = (0.6, 0.4)^\top. Missing values were imposed on WW under the MAR mechanism with a target missing proportion of 0.6. This setting is motivated by applications in which a partially observed biomarker, such as a cerebrospinal fluid biomarker, is subsequently related to a fully observed psychometric outcome.

SHIM was fitted jointly to WW and YY conditional on XX and ZZ, and M=20M = 20 completed datasets were obtained from posterior draws of the missing values of WW. Within each completed dataset, the downstream model regressed YY on WW and ZZ using ordinary least squares, and inference for κ\kappa was combined across imputations using Rubin’s rules. The procedure was repeated over 100 simulation replicates to evaluate estimation performance of κ\kappa.

Table 6 summarizes the downstream estimation performance of posterior-based multiple imputation. For the association of interest (κ\kappa), the mean estimate showed low bias, and the average estimated standard error closely matched the empirical variability, although the confidence intervals exhibited modest under-coverage. The effects of the continuous and binary covariates (ψY1\psi_{Y_1} and ψY2\psi_{Y_2}, respectively) were also estimated with reasonably low bias, with estimated standard errors comparable to their empirical counterparts and coverage probabilities close to the nominal 95% level. Overall, these results suggest that posterior-based multiple imputation can support reasonably accurate downstream estimation and uncertainty quantification for both the association of interest and the covariate effects, even in this challenging setting with 60% missing data.

ParameterTrue ValueEstimateESDESD / SECP (%)
κ\kappa0.80.780.030.9590
ψY1\psi_{Y_1}0.60.640.100.9894
ψY2\psi_{Y_2}0.40.410.160.9495

Table 6. Downstream estimation performance of posterior-based multiple imputation across 100 simulation replicates, where κ\kappa denotes the association of interest, whereas ψY1\psi_{Y_1} and ψY2\psi_{Y_2} denote the associations of the continuous and binary covariates with the outcome, respectively.

Application

Data Description

We applied SHIM to baseline data from the legacy cohort of the Vanderbilt Memory and Aging Project (VMAP), a longitudinal observational study designed to investigate the relationships among vascular health, brain aging, and cognitive impairment in older adults [27]. VMAP collected comprehensive demographic, clinical, neuropsychological, fluid-biomarker, and multimodal neuroimaging data. The analytic sample included 288 participants after excluding those with incomplete demographic information. Demographic and genetic characteristics of the analytic sample are summarized in Table 7.

CharacteristicOverall (n=288n = 288)
Age, years72.8±7.472.8 \pm7.4
Sex (female), %43.4
Education, years15.9±2.715.9 \pm2.7
Race/Ethnicity (NHW), %87.2
APOE-ε4 carrier (positive), %35.4

Table 7. Demographic and genetic characteristics of the VMAP analytic sample. NHW: non-Hispanic white; APOE-ε4: apolipoprotein E epsilon 4.

We considered regional gray matter (GM) volume and cerebral blood flow (CBF) as complementary neuroimaging predictors. VMAP participants underwent brain magnetic resonance imaging at the Vanderbilt University Institute of Imaging Science using a 3T Philips Achieva system (Best, the Netherlands). Regional GM volumes were derived from T1-weighted magnetic resonance images, which were skull-stripped using SynthStrip [17] and parcellated into anatomical regions of interest (ROIs) using NiChart_DLMUSE [2]. Regional CBF estimates were derived from pseudo-continuous arterial spin-labeling (pCASL) images and corrected for partial-volume effects. Cerebrospinal fluid (CSF) Aβ1−42A\beta_{1-42} was measured using the Fujirebio INNOTEST assay, and was additionally considered as a biomarker associated with cerebral amyloid pathology [18]. Detailed protocols for neuroimaging acquisition and processing and CSF collection are provided in Appendix F.

VMAP participants completed a comprehensive neuropsychological assessment covering episodic memory, executive function, language and information processing speed. The present analyses used VMAP composite scores for episodic memory and executive function. These standardized scores were constructed from item-level data using bifactor latent-variable models in which each measure loaded on both a general domain factor and a test-specific factor [28]. The episodic memory composite incorporated learning, recall, and recognition measures from the California Verbal Learning Test, Second Edition (CVLT-II), and the Biber Figure Learning Test (BFLT). The executive function composite incorporated the Delis–Kaplan Executive Function System (D-KEFS) Color–Word Inhibition, Tower, and Letter–Number Switching tests, together with Letter Fluency (FAS).

Analysis Procedure

We conducted two multivariate analyses. The first jointly considered memory and executive function, two related but distinct cognitive domains, to identify neuroimaging associations that were shared across or specific to each domain. The second jointly considered memory and CSF amyloid Aβ1−42A\beta_{1-42}. This analysis had two objectives: to examine how the neuroimaging predictors were associated with cognitive function and AD pathology, and to investigate the downstream association between memory and CSF Aβ1−42A\beta_{1-42}. Specific for the second analysis, because CSF Aβ1−42A\beta_{1-42} was unavailable for 53.1% of the participants, SHIM was used to generate posterior-based multiple imputations of the missing data. The completed datasets were then used to estimate the association between memory and the CSF biomaker while accounting for uncertainty arising from the missing data.

The predictor hierarchy consisted of two neuroimaging modalities: regional GM and CBF, each containing 98 predictors. Within each modality, predictors were further organized into five anatomical groups, including frontal, limbic, occipital, parietal, and temporal. This hierarchy was supplied to SHIM to guide hierarchical shrinkage and variable selection. All neuroimaging predictors were standardized prior to model implementation.

All analyses adjusted for age, sex, APOE-ε4APOE\text{-}\varepsilon4 carrier status, race and ethnicity, and years of education. Intracranial volume (ICV) was also included as a covariate in analyses to account for individual differences in head size. All adjusting covariates were included in SHIM as unpenalized variables. Participants in the analytic sample (of 288) did not have missing data in the neuroimaging predictors.

In the first analysis, we applied SHIM and its univariate counterpart, SHIM_USHIM\_U, to episodic memory and executive function. SHIM_USHIM\_U was included because it was among the most competitive competing alternative in the simulation study and accommodated the same unpenalized adjusting covariates as SHIM. MBSGSSS was not included in the analysis since it was not able to directly accommodate unpenalized covariates.

In the second analysis, we applied posterior-based multiple imputation by SHIM, complete-case analysis, and multiple imputation by chained equations (MICE) [41]. For MICE, the default method was used, and for each imputation, 20 completed datasets were generated. The imputation model included memory, neuroimaging predictors, and all adjusting covariates by convention [25]. Within each completed dataset, memory was regressed on standardized CSF Aβ1−42A\beta_{1-42} with adjustment for the demographic and clinical covariates described above. The resulting estimates and variances were combined using Rubin’s rules [35].

Results

In the first analysis, SHIM identified positive associations between GM in the left parahippocampal gyrus and episodic memory (β^=0.243\hat{\beta}=0.243, 95% CI [credible interval]: 0.038–0.410), and between GM in the left inferior frontal gyrus, pars triangularis and executive function (β^=0.163\hat{\beta} = 0.163, 95% CI: 0.011–0.277). SHIMU_{\mathrm{U}} identified only the association between left parahippocampal GM and episodic memory (β^=0.254\hat{\beta} = 0.254, 95% CI: 0.007–0.421). No regional CBF measure met the selection criterion under either approach. Overall, SHIM selected two of the 392 possible associations between the 196 regional imaging measures and the two cognitive outcomes, both involving GM volume. This parsimonious result suggests that, after accounting jointly for the correlated imaging measures and covariates, posterior evidence was concentrated in a small number of anatomically specific regions. The selected regions were broadly consistent with previous neuroimaging evidence. Parahippocampal gray matter has been implicated in episodic memory in older adults, and reduced parahippocampal volume has been observed among individuals with memory impairment [9, 20]. Similarly, previous studies have linked inferior frontal GM volume to executive-function performance [22, 31]. Notably, the executive-function association was identified under multivariate SHIM but not under SHIM_U, underscoring the potential benefit of borrowing information across related cognitive outcomes in multivariate models.

In the second analysis, the estimated downstream association between episodic memory and CSF Aβ1−42A\beta_{1-42} is summarized in Table 8. The pooled analysis based on SHIM posterior imputations indicated a positive association, with higher CSF Aβ1−42A\beta_{1-42} associated with better episodic-memory performance (β^=0.178\hat{\beta} = 0.178, 95% CI [confidence interval]: 0.044–0.313, p=0.01p = 0.01). This relationship is consistent with the established AD biomarker literature, where lower CSF Aβ1−42A\beta_{1-42} reflects greater cerebral amyloid deposition and is associated with poorer cognitive outcomes [15, 23, 33]. The complete case analysis produced a similar point estimate but a greater standard error and a wider confidence interval, resulting in statistical insignificance at the nominal level. The lower precision indicated the loss of information from excluding the 153 participants without observed CSF measurements. MICE produced a counterintuitive result. With 196 correlated imaging predictors but only 135 observed CSF measurements, MICE was potentially sensitive to collinearity and imputation model specification. In contrast, SHIM jointly modeled the outcomes while applying hierarchical shrinkage to the imaging predictors and generating posterior imputations of the missing CSF measurements. In this application, the integrated approach produced an association estimate consistent with prior biological evidence and greater precision than the complete case analysis.

MethodSample SizeEstimate95% CIP-value
SHIM2880.178(0.044, 0.313)0.010
Complete case1350.178(−0.001, 0.357)0.054
MICE288−0.006(−0.014, 0.001)0.099

Table 8. Estimated association between episodic memory and CSF Aβ1−42A\beta_{1-42} using posterior-based multiple imputation by SHIM, complete case analysis, and MICE. CI: confidence interval.

Discussion

In this paper, we proposed SHIM, a Bayesian framework for structured variable selection in studies involving high-dimensional, multi-modal predictors and partially observed multivariate psychometric outcomes. Such settings increasingly arise when multiple related behavioral or cognitive constructs are studied alongside neuroimaging, genetic, or other high-dimensional auxiliary information. SHIM combines hierarchical horseshoe shrinkage with Bayesian treatment of the missing data, allowing the predictor hierarchy, dependence among the outcomes, and incomplete outcome data to be accommodated within a unified framework. The theoretical results formalize two properties of the proposed prior. Coefficients that are closer in the predictor hierarchy have more strongly associated prior magnitudes, and the marginal coefficient density combines substantial concentration near zero with heavy tails. Across the primary simulations and sensitivity analyses, SHIM generally achieved a balance between sensitivity and false positive control while providing accurate coefficient estimates and reasonably calibrated uncertainty quantification, although its performance deteriorated when the MAR assumption was violated. The VMAP application further illustrated two practical uses of the framework: borrowing information across related cognitive outcomes to identify neuroimaging associations and generating posterior-based multiple imputations for a downstream analysis involving an incompletely observed CSF biomarker.

The present study has several limitations that warrant further investigation. First, although SHIM accommodates missing outcomes, it currently requires the high-dimensional predictors and adjusting covariates to be fully observed, which may be restrictive in practice. Second, inference relies on the MAR assumption. The increased bias and under-coverage observed in the MNAR sensitivity setting indicate that performance may deteriorate when this assumption is severely violated. Third, the current formulation is primarily based on continuous outcomes modeled using a multivariate normal distribution. Extensions to binary, ordinal and mixed-type outcomes would broaden its applicability to psychometric research. Finally, SHIM requires the predictor hierarchy to be specified in advance, and its performance and interpretation may depend on how well that hierarchy reflects the underlying scientific structure. Future methodological work could therefore address incomplete predictors, non-Gaussian or latent psychometric outcomes, nonignorable missingness mechanisms, and more flexible or data-adaptive predictor hierarchies.

Establishing high-dimensional posterior theory for SHIM also remains an important direction for future research. The shared modality-, group-, and predictor-level scales induce dependence across the coefficient matrix, requiring joint prior-concentration arguments for hierarchically structured sparse signals rather than coordinate-wise calculations. The residual covariance matrix must also be learned jointly with the regression surface, while the observed-data likelihood under MAR varies across missingness patterns. A possible strategy is to combine general prior-concentration and testing frameworks for posterior contraction [14] with techniques developed for high-dimensional regression under continuous shrinkage priors [37] and multivariate regression with an unknown covariance matrix [48]. Establishing variable-selection consistency and valid uncertainty quantification would additionally require appropriate sparsity and signal-strength conditions [30, 37]. Once joint posterior contraction for the regression and covariance parameters has been established under the observed-data likelihood, corresponding contraction of the posterior predictive imputation distribution may follow from continuity of the conditional multivariate normal distribution. These developments would provide a fuller theoretical foundation for structured Bayesian variable selection and imputation in high-dimensional psychometric studies.

Funding Statement. This research was supported by grants from the NIH (R01AG034962).

Competing Interests. T.J.H. serves on the Scientific Advisory Board of Circular Genomics, as Deputy Editor of Alzheimer’s & Dementia: Translational Research & Clinical Interventions, and as Section Editor of Alzheimer’s & Dementia. A.L.J is an advisor for Medtronic, and Chair of the Observational Study Monitoring Board for the Diverse-VCID: White Matter Lesion Etiology of Dementia in Diverse Populations Study.

Data Availability Statement. The VMAP data analyzed in this study are not publicly available. Qualified investigators may apply for access through the VMAP Data Sharing Portal at https://vmapdata.org/. Requests are subject to review and approval by the VMAP scientific committee and require completion of a data-use agreement. Code used to implement the analyses reported in this article is available at https://github.com/zongyue-teng/shim.

Ethical Standards. The research meets all ethical guidelines, including adherence to the legal requirements of the study country.

Author Contributions. Conceptualization: Z.T.; S.M.; T.J.H.; P.Z. Data curation: T.J.H.; A.L.J. Formal Analysis: Z.T. Funding Acquisition: A.L.J. Methodology: Z.T; S.M; P.Z. Software: Z.T.; P.Z. Supervision: P.Z. Validation: Z.T.; P.Z. Visualization: Z.T. Writing original draft: Z.T.; P.Z. Writing review & editing: Z.T.; S.M.; T.J.H.; A.L.J.; P.Z. All authors approved the final submitted draft.

Derivation of Gaussian Full Conditional Distributions

This appendix provides derivations of the Gaussian full conditional distributions for the regression coefficients and intercept vector used in the Gibbs sampler, i.e., Algorithm 1. The remaining updates either follow directly from standard conjugate Bayesian analyses (e.g., the inverse-Wishart update for the residual covariance matrix) or from the auxiliary-variable representation of the half-Cauchy prior [26], and are therefore omitted. Throughout this appendix, “others” denotes conditioning on the current values of all remaining unknown quantities.

Regression Coefficients

Denote

w=vec⁡{(Y−1nαT−ZΨT)T}andb=vec⁡(BT).\mathbf{w} = \operatorname{vec}\left\{\left(\mathbf{Y} - \mathbf{1}_n\boldsymbol{\alpha}^{\mathsf{T}} - \mathbf{Z}\mathbf{\Psi}^{\mathsf{T}}\right)^{\mathsf{T}}\right\} \qquad\text{and}\qquad \mathbf{b} = \operatorname{vec}\left(\mathbf{B}^{\mathsf{T}}\right).

Using the “vec trick” identity [16] given by

vec⁡(M1M2M3)=(M3T⊗M1)vec⁡(M2),\operatorname{vec}(\mathbf{M}_1\mathbf{M}_2\mathbf{M}_3) = \left(\mathbf{M}_3^{\mathsf{T}} \otimes\mathbf{M}_1\right)\operatorname{vec}(\mathbf{M}_2),

where ⊗\otimes denotes the Kronecker product, we re-express the w\mathbf{w} as

w=vec⁡{(Y−1nαT−ZΨT)T}=vec⁡(BTXT)+vec⁡(ET)=(X⊗IQ)b+vec⁡(ET),\mathbf{w} = \operatorname{vec}\left\{\left(\mathbf{Y} - \mathbf{1}_n\boldsymbol{\alpha}^{\mathsf{T}} - \mathbf{Z}\mathbf{\Psi}^{\mathsf{T}}\right)^{\mathsf{T}}\right\} = \operatorname{vec}\left(\mathbf{B}^{\mathsf{T}}\mathbf{X}^{\mathsf{T}}\right) + \operatorname{vec}\left(\mathbf{E}^{\mathsf{T}}\right) = \left(\mathbf{X} \otimes\mathbf{I}_Q\right)\mathbf{b} + \operatorname{vec}\left(\mathbf{E}^{\mathsf{T}}\right),

which suggests w∣b,Σ∼NnQ(Hb,R)\mathbf{w} \mid\mathbf{b}, \mathbf{\Sigma} \sim N_{nQ}(\mathbf{H}\mathbf{b}, \mathbf{R}) with H:=X⊗IQ\mathbf{H} := \mathbf{X} \otimes\mathbf{I}_Q and R:=In⊗Σ\mathbf{R} := \mathbf{I}_n \otimes\mathbf{\Sigma}.

Since the prior distribution of b\mathbf{b} is specified as b∼NPQ(0,DB)\mathbf{b} \sim N_{PQ}(\mathbf{0}, \mathbf{D}_B), the posterior density is given by

p(b∣others)∝p(w∣b,Σ)p(b)p(\mathbf{b} \mid\text{others}) \propto p(\mathbf{w} \mid\mathbf{b}, \mathbf{\Sigma})p(\mathbf{b})
∝exp⁡{−12(w−Hb)TR−1(w−Hb)−12bTDB−1b}\propto\exp\left\{-\frac{1}{2}(\mathbf{w} - \mathbf{H}\mathbf{b})^{\mathsf{T}}\mathbf{R}^{-1}(\mathbf{w} - \mathbf{H}\mathbf{b}) - \frac{1}{2}\mathbf{b}^{\mathsf{T}}\mathbf{D}_B^{-1}\mathbf{b}\right\}
∝exp⁡{−12bT(HTR−1H+DB−1)b+bTHTR−1w},\propto\exp\left\{-\frac{1}{2}\mathbf{b}^{\mathsf{T}}\left(\mathbf{H}^{\mathsf{T}}\mathbf{R}^{-1}\mathbf{H} + \mathbf{D}_B^{-1}\right)\mathbf{b} + \mathbf{b}^{\mathsf{T}}\mathbf{H}^{\mathsf{T}}\mathbf{R}^{-1}\mathbf{w}\right\},

which leads to

b∣others∼NPQ(μB,ΣB),\mathbf{b} \mid\text{others} \sim N_{PQ}(\boldsymbol{\mu}_B, \mathbf{\Sigma}_B),

with

ΣB=(HTR−1H+DB−1)−1andμB=ΣBHTR−1w\mathbf{\Sigma}_B = \left(\mathbf{H}^{\mathsf{T}}\mathbf{R}^{-1}\mathbf{H} + \mathbf{D}_B^{-1}\right)^{-1} \qquad\text{and}\qquad \boldsymbol{\mu}_B = \mathbf{\Sigma}_B\mathbf{H}^{\mathsf{T}}\mathbf{R}^{-1}\mathbf{w}

according to the multivariate normal distribution kernel. Finally, plugging H\mathbf{H} and R\mathbf{R} in the expressions of ΣB\mathbf{\Sigma}_B and μB\boldsymbol{\mu}_B to get

ΣB=((X⊗IQ)T(In⊗Σ)−1(X⊗IQ)+DB−1)−1=((XTX)⊗Σ−1+DB−1)−1\mathbf{\Sigma}_B = \left(\left(\mathbf{X} \otimes\mathbf{I}_Q\right)^{\mathsf{T}}\left(\mathbf{I}_n \otimes\mathbf{\Sigma}\right)^{-1}\left(\mathbf{X} \otimes\mathbf{I}_Q\right) + \mathbf{D}_B^{-1}\right)^{-1} = \left(\left(\mathbf{X}^{\mathsf{T}}\mathbf{X}\right) \otimes\mathbf{\Sigma}^{-1} + \mathbf{D}_B^{-1}\right)^{-1}

and

μB=ΣB(X⊗IQ)T(In⊗Σ)−1w=ΣB(XT⊗Σ−1)w,\boldsymbol{\mu}_B = \mathbf{\Sigma}_B\left(\mathbf{X} \otimes\mathbf{I}_Q\right)^{\mathsf{T}}\left(\mathbf{I}_n \otimes\mathbf{\Sigma}\right)^{-1}\mathbf{w} = \mathbf{\Sigma}_B\left(\mathbf{X}^{\mathsf{T}} \otimes\mathbf{\Sigma}^{-1}\right)\mathbf{w},

which completes the derivations.

Intercept Vector and Covariate Coefficients

To derive the joint full conditional distribution of the intercept and covariate coefficient matrix Ψ~\widetilde{\mathbf{\Psi}}, consider

u=vec⁡{(Y−XB)T}andψ~=vec⁡(Ψ~T).\mathbf{u} = \operatorname{vec}\left\{\left(\mathbf{Y} - \mathbf{X}\mathbf{B}\right)^{\mathsf{T}}\right\} \qquad\text{and}\qquad \widetilde{\boldsymbol{\psi}} = \operatorname{vec}\left(\widetilde{\mathbf{\Psi}}^{\mathsf{T}}\right).

Similarly, we can re-express u\mathbf{u} as

u=vec⁡(Ψ~TZ~T)+vec⁡(ET)=(Z~⊗IQ)ψ~+vec⁡(ET),\mathbf{u} = \operatorname{vec}\left(\widetilde{\mathbf{\Psi}}^{\mathsf{T}}\widetilde{\mathbf{Z}}^{\mathsf{T}}\right) + \operatorname{vec}\left(\mathbf{E}^{\mathsf{T}}\right) = \left(\widetilde{\mathbf{Z}} \otimes\mathbf{I}_Q\right)\widetilde{\boldsymbol{\psi}} + \operatorname{vec}\left(\mathbf{E}^{\mathsf{T}}\right),

which suggests u∣ψ~,Σ∼NnQ(Gψ~,R)\mathbf{u} \mid\widetilde{\boldsymbol{\psi}}, \mathbf{\Sigma} \sim N_{nQ}(\mathbf{G}\widetilde{\boldsymbol{\psi}}, \mathbf{R}) with G:=Z~⊗IQ\mathbf{G} := \widetilde{\mathbf{Z}} \otimes\mathbf{I}_Q and R=In⊗Σ\mathbf{R} = \mathbf{I}_n \otimes\mathbf{\Sigma}. Given the specified prior for Ψ~\widetilde{\mathbf{\Psi}}, denote ψ~0=vec⁡(Ψ~0)\widetilde{\boldsymbol{\psi}}_0 = \operatorname{vec}(\widetilde{\mathbf{\Psi}}_0) and Dψ~=V~0⊗IQ\mathbf{D}_{\widetilde{\boldsymbol{\psi}}} = \widetilde{\mathbf{V}}_0 \otimes I_Q for notation simplicity.

The posterior density is given by

p(ψ~∣others)∝p(u∣ψ~,Σ)p(ψ~)p(\widetilde{\boldsymbol{\psi}} \mid\text{others}) \propto p(\mathbf{u} \mid\widetilde{\boldsymbol{\psi}}, \boldsymbol{\Sigma})p(\widetilde{\boldsymbol{\psi}})
∝exp⁡{−12(u−Gψ~)TR−1(u−Gψ~)−12(ψ~−ψ~0)TDψ~−1(ψ~−ψ~0)}\propto\exp\left\{-\frac{1}{2}(\mathbf{u}-\mathbf{G}\widetilde{\boldsymbol{\psi}})^{\mathsf{T}}\mathbf{R}^{-1}(\mathbf{u}-\mathbf{G}\widetilde{\boldsymbol{\psi}})-\frac{1}{2}(\widetilde{\boldsymbol{\psi}}-\widetilde{\boldsymbol{\psi}}_0)^{\mathsf{T}}\mathbf{D}_{\widetilde{\boldsymbol{\psi}}}^{-1}(\widetilde{\boldsymbol{\psi}}-\widetilde{\boldsymbol{\psi}}_0)\right\}
∝exp⁡{−12ψ~T(GTR−1G+Dψ~−1)ψ~+ψ~T(GTR−1u+Dψ~−1ψ~0)}.\propto\exp\left\{-\frac{1}{2}\widetilde{\boldsymbol{\psi}}^{\mathsf{T}}\left(\mathbf{G}^{\mathsf{T}}\mathbf{R}^{-1}\mathbf{G}+\mathbf{D}_{\widetilde{\boldsymbol{\psi}}}^{-1}\right)\widetilde{\boldsymbol{\psi}}+\widetilde{\boldsymbol{\psi}}^{\mathsf{T}}\left(\mathbf{G}^{\mathsf{T}}\mathbf{R}^{-1}\mathbf{u}+\mathbf{D}_{\widetilde{\boldsymbol{\psi}}}^{-1}\widetilde{\boldsymbol{\psi}}_0\right)\right\}.

which corresponds to

ψ~∣others∼N(L+1)Q(μψ~,Σψ~).\widetilde{\boldsymbol{\psi}} \mid\text{others} \sim\mathcal{N}_{(L+1)Q}\left(\boldsymbol{\mu}_{\widetilde{\boldsymbol{\psi}}},\boldsymbol{\Sigma}_{\widetilde{\boldsymbol{\psi}}}\right).

with

Σψ~=(GTR−1G+Dψ~−1)−1andμψ~=Σψ~(GTR−1u+Dψ~−1ψ~0).\boldsymbol{\Sigma}_{\widetilde{\boldsymbol{\psi}}}=\left(\mathbf{G}^{\mathsf{T}}\mathbf{R}^{-1}\mathbf{G}+\mathbf{D}_{\widetilde{\boldsymbol{\psi}}}^{-1}\right)^{-1} \qquad\text{and}\qquad \boldsymbol{\mu}_{\widetilde{\boldsymbol{\psi}}}=\boldsymbol{\Sigma}_{\widetilde{\boldsymbol{\psi}}}\left(\mathbf{G}^{\mathsf{T}}\mathbf{R}^{-1}\mathbf{u}+\mathbf{D}_{\widetilde{\boldsymbol{\psi}}}^{-1}\widetilde{\boldsymbol{\psi}}_0\right).

Plugging in for G\mathbf{G}, R\mathbf{R}, Dψ~\mathbf{D}_{\widetilde{\boldsymbol{\psi}}}, u\mathbf{u} and ψ~0\widetilde{\boldsymbol{\psi}}_0 leads to

Σψ~=((Z~⊗IQ)T(In⊗Σ)−1(Z~⊗IQ)+(V~0⊗IQ)−1)−1=((Z~TZ~)⊗Σ−1+V~0−1⊗IQ)−1.\boldsymbol{\Sigma}_{\widetilde{\boldsymbol{\psi}}}=\left((\widetilde{\mathbf{Z}}\otimes I_Q)^{\mathsf{T}}(I_n\otimes\boldsymbol{\Sigma})^{-1}(\widetilde{\mathbf{Z}}\otimes I_Q)+(\widetilde{\mathbf{V}}_0\otimes I_Q)^{-1}\right)^{-1}=\left((\widetilde{\mathbf{Z}}^{\mathsf{T}}\widetilde{\mathbf{Z}})\otimes\boldsymbol{\Sigma}^{-1}+\widetilde{\mathbf{V}}_0^{-1}\otimes I_Q\right)^{-1}.

and

μψ~=Σψ~{(Z~T⊗Σ−1)vec⁡((Y−XB)T)+(V~0−1⊗IQ)vec⁡(Ψ~0T)}.\boldsymbol{\mu}_{\widetilde{\boldsymbol{\psi}}}=\boldsymbol{\Sigma}_{\widetilde{\boldsymbol{\psi}}}\left\{(\widetilde{\mathbf{Z}}^{\mathsf{T}}\otimes\boldsymbol{\Sigma}^{-1})\operatorname{vec}\left((\mathbf{Y}-\mathbf{X}\mathbf{B})^{\mathsf{T}}\right)+(\widetilde{\mathbf{V}}_0^{-1}\otimes I_Q)\operatorname{vec}(\widetilde{\mathbf{\Psi}}_0^{\mathsf{T}})\right\}.

This completes the derivations.

Proof of Theorem 2.1

Proof. For x>0x>0, define the one-sided Mellin transform of the symmetric density πK\pi_K by

MK(s)=∫0∞xs−1πK(x) dx.\mathcal{M}_K(s)=\int_0^\infty x^{s-1}\pi_K(x)\,\mathrm{d}x.

By symmetry, we get

MK(s)=12E(∣β∣s−1).\mathcal{M}_K(s)=\frac{1}{2}\mathbb{E}\left(|\beta|^{s-1}\right).

Write

β=SKZ,SK=∏ℓ=1KHℓ,\beta=S_KZ,\qquad S_K=\prod_{\ell=1}^{K}H_\ell,

where Z∼N(0,1)Z\sim\mathcal{N}(0,1) and H1,…,HK∼indC+(0,1)H_1,\ldots,H_K\stackrel{\mathrm{ind}}{\sim}\mathcal{C}^{+}(0,1) are mutually independent. It follows that

MK(s)=12E(∣Z∣s−1)∏ℓ=1KE(Hℓs−1).\mathcal{M}_K(s)=\frac{1}{2}\mathbb{E}\left(|Z|^{s-1}\right)\prod_{\ell=1}^{K}\mathbb{E}\left(H_\ell^{s-1}\right).

For a standard half-Cauchy variable, we know

E(Hℓs−1)=2π∫0∞hs−11+h2 dh=csc⁡(πs2)\mathbb{E}\left(H_\ell^{s-1}\right)=\frac{2}{\pi}\int_0^\infty\frac{h^{s-1}}{1+h^2}\,\mathrm{d}h=\csc\left(\frac{\pi s}{2}\right)

with 0<Re⁡(s)<20 < \operatorname{Re}(s) < 2. For Z∼N(0,1)Z \sim\mathcal{N}(0,1), we have

E(∣Z∣s−1)=2(s−1)/2Γ(s/2)π\mathbb{E}\left(|Z|^{s-1}\right) = \frac{2^{(s-1)/2}\Gamma(s/2)}{\sqrt{\pi}}

with Re⁡(s)>0\operatorname{Re}(s) > 0. Therefore, we get

MK(s)=2(s−1)/2Γ(s/2)2π{csc⁡(πs2)}K.(3)\mathcal{M}_{K}(s) = \frac{2^{(s-1)/2}\Gamma(s/2)}{2\sqrt{\pi}}\left\{\csc\left(\frac{\pi s}{2}\right)\right\}^{K}. \tag*{(3)}

for 0<Re⁡(s)<20 < \operatorname{Re}(s) < 2, which is the fundamental strip of MK\mathcal{M}_{K}. For convenience, define

AK=12π(2π)K.A_{K} = \frac{1}{\sqrt{2\pi}}\left(\frac{2}{\pi}\right)^{K}.

We first consider the behavior near zero. As s→0s \to0, we have

2(s−1)/2=2−1/2{1+O(s)},2^{(s-1)/2} = 2^{-1/2}\{1 + O(s)\},
Γ(s/2)=2s{1+O(s)},\Gamma(s/2) = \frac{2}{s}\{1 + O(s)\},
csc⁡(πs2)=2πs{1+O(s2)}.\csc\left(\frac{\pi s}{2}\right) = \frac{2}{\pi s}\{1 + O(s^{2})\}.

Substitution into (3) gives

MK(s)=AKs−(K+1){1+O(s)}.\mathcal{M}_{K}(s) = A_{K}s^{-(K+1)}\{1 + O(s)\}.

Thus, MK\mathcal{M}_{K} has a pole of order K+1K+1 at the left boundary s=0s = 0. The converse mapping theorem for Mellin transforms [12] therefore gives

πK(x)∼AKK!{log⁡(1x)}K,\pi_{K}(x) \sim\frac{A_{K}}{K!}\left\{\log\left(\frac{1}{x}\right)\right\}^{K},

as x↓0x \downarrow0.

We next consider the tail behavior. Set w=2−sw = 2-s. As w→0w \to0, we have

2(s−1)/2=2(1−w)/2=2{1+O(w)},2^{(s-1)/2} = 2^{(1-w)/2} = \sqrt{2}\{1 + O(w)\},
Γ(s/2)=Γ(1−w/2)=1+O(w),\Gamma(s/2) = \Gamma(1-w/2) = 1 + O(w),
csc⁡(πs2)=csc⁡(πw2)=2πw{1+O(w2)}.\csc\left(\frac{\pi s}{2}\right) = \csc\left(\frac{\pi w}{2}\right) = \frac{2}{\pi w}\{1 + O(w^{2})\}.

Hence, we get

MK(2−w)=AKw−K{1+O(w)}.\mathcal{M}_{K}(2-w) = A_{K}w^{-K}\{1 + O(w)\}.

Equivalently, MK(s)\mathcal{M}_{K}(s) has a pole of order KK at the right boundary s=2s = 2. A second application of the converse mapping theorem yields

πK(x)∼AK(K−1)!(log⁡x)K−1x2,\pi_{K}(x) \sim\frac{A_{K}}{(K-1)!}\frac{(\log x)^{K-1}}{x^{2}},

as x→∞x \to\infty. Symmetry of πK\pi_{K} allows xx to be replaced by ∣β∣|\beta|.

It remains to verify the analytic conditions required by the converse mapping theorem. The expression in (3) extends meromorphically beyond its fundamental strip. Moreover, for fixed real σ\sigma on a vertical line avoiding its poles, Stirling’s formula and the standard asymptotics of the cosecant function give

∣Γ(σ+it2)∣=O((1+∣t∣)σ/2−1/2e−π∣t∣/4)and∣csc⁡(π(σ+it)2)∣=O(e−π∣t∣/2)\left|\Gamma\left(\frac{\sigma+\mathrm{i}t}{2}\right)\right|=O\left((1+|t|)^{\sigma/2-1/2}e^{-\pi|t|/4}\right) \quad\text{and} \quad\left|\csc\left(\frac{\pi(\sigma+\mathrm{i}t)}{2}\right)\right|=O\left(e^{-\pi|t|/2}\right)

as ∣t∣→∞|t| \to\infty. Thus, MK(σ+it)M_K(\sigma+\mathrm{i}t) decays exponentially on such vertical lines, which completes the proof. □

Example Trace Plots for Bayesian Methods

Figure 3 presents representative trace plots and for all four Bayesian methods under one benchmark simulation setting. Similar convergence behavior was observed across the remaining simulation settings.

Figure 3. Example trace plots of the four active regression coefficients for the four Bayesian methods under Scenario S6, based on 2,000 MCMC iterations with the first 500 iterations discarded as burn-in.

Additional Estimation Results

For completeness, this appendix reports two additional estimation results that were not presented in the main text. Figure 4 shows the empirical standard deviation of the coefficient estimates and the ratio of the empirical standard deviation to the average estimated standard error. Overall, SHIM, SHIM_U, and MBSGSSS exhibited comparable empirical standard deviations, with slightly larger variability observed for the weaker coefficients under the more challenging simulation settings. GIGG demonstrated obviously greater variability for the weaker signals, whereas glasso generally produced the largest empirical standard deviations, particularly when both predictor dimensionality and missingness were high. The empirical-to-estimated standard error ratios for SHIM remained close to one across nearly all simulation settings, indicating good agreement between empirical sampling variability and model-based uncertainty estimates. Similar behavior was observed for SHIM_U and MBSGSSS, whereas larger deviations from one were found for lasso, glasso, and GIGG in the more challenging settings.

Figure 4. Empirical standard deviation (top) and empirical-to-estimated standard error (bottom) for the four active coefficients across the eight simulation settings. Coefficients β1\beta_1 and β2\beta_2 correspond to the two stronger signals (true coefficient = 1.2), whereas β3\beta_3 and β4\beta_4 correspond to the two weaker signals (true coefficient = 0.6). For lasso and group lasso, the reported summaries are based on post-selection ordinary least squares estimates.

Full Results for Sensitivity Analysis

Table 9 provides the full variable-selection results. As discussed in Section 3.5, SHIM and SHIM_U maintained high F1 scores across all sensitivity settings, with SHIM consistently showing higher TPR values than SHIM_U. MBSGSSS was comparable to SHIM and SHIM_U under most settings except OUT, where its F1 scores deteriorated substantially, primarily because of increased FDR. GIGG continued to exhibit lower TPR values, whereas lasso and glasso yielded inflated FDR values.

Table 10 provides the full coefficient-estimation results. Across settings, GIGG demonstrated greater bias and under-coverage for the weak signals, lasso showed persistent bias and under-coverage across the active coefficients, and glasso exhibited substantially greater empirical variability and poorer standard error calibration. In particular, for ULTRA, no valid post-selection OLS estimates were available for glasso because the number of selected predictors was consistently at least as large as the complete-case sample size.

Under CORR, SHIM, SHIM_U, and MBSGSSS produced comparably low bias for the strong signals but greater bias for some weak signals. However, SHIM and SHIM_U exhibited lower empirical standard deviations than MBSGSSS and the other competing methods, particularly for the weak signals, with SHIM generally showing slightly smaller values than SHIM_U. Their empirical-to-average estimated standard error ratios were generally close to one, and coverage was close to the nominal 95% level. In contrast, MBSGSSS showed noticeably poorer standard error calibration and some under-coverage for the weak signals.

Similar patterns were observed under MNAR, although bias increased and coverage deteriorated for most methods. Nevertheless, SHIM and SHIM_U retained competitive bias and relatively low empirical variability. In contrast, MBSGSSS exhibited greater bias even for the strong signals, together with more pronounced standard error miscalibration and under-coverage for the weak signals.

Under ULTRA, SHIM, SHIM_U, and MBSGSSS performed comparably in terms of bias and empirical variability, with low bias for the strong signals but greater bias for some weak signals. All three methods showed some standard error miscalibration and under-coverage for the weak signals. MBSGSSS generally exhibited the greatest miscalibration, whereas SHIM remained competitive in calibration and coverage.

Neuroimaging and CSF Acquisition Details

T1-weighted magnetization-prepared rapid gradient-echo (MPRAGE) images were acquired with a repetition time (TR) of 8.9 ms, an echo time (TE) of 4.6 ms, and an isotropic spatial resolution of 1×1×11 \times1 \times1 mm3^{3}. The images were skull-stripped using SynthStrip [17] and parcellated into anatomical regions of interest (ROIs) using NiChart_DLMUSE [2]. Regional GM volumes were derived from the resulting anatomical parcellations.

Pseudo-continuous arterial spin-labeling (pCASL) label-control image pairs were acquired to quantify CBF. Acquisition parameters included a labeling duration of 1.65 s, a post-labeling delay of 1.525 s, a spatial resolution of 3×3×73 \times3 \times7 mm3^{3}, a TR of 4,000 ms, and a TE of 13 ms. Motion correction was

performed using 3dvolreg in AFNI (version 23.1.10). Regional CBF estimates were derived for the anatomical ROIs and corrected for partial-volume effects to reduce the influence of regional atrophy on the estimated perfusion values.

Cerebrospinal fluid collection was an optional component of the VMAP protocol. Participants undergoing CSF collection completed a morning fasting lumbar puncture using a Sprotte 25-gauge spinal needle. Samples were immediately mixed and centrifuged at 2,000 g and 4°C for 10 minutes, after which the supernatant was aliquoted into polypropylene tubes and stored at −80°C. CSF Aβ1−42A\beta_{1-42} concentrations were measured in batch using the Fujirebio INNOTEST β\beta-AMYLOID(1−42)_{(1-42)} enzyme-linked immunosorbent assay. The reported intra-assay coefficients of variation were below 10%.

References

  1. [1]Ethan M. Alt, Anil Anderson, Qing Li, Jia Hua, Amarjot Kaur, and Joseph G. Ibrahim, Hierarchical grouped horseshoe priors for subgroup identification and estimation, Statistics in Medicine 44 (2025), no. 20–22, e70246.DOI
  2. [2]Vishnu M. Bashyam, Guray Erus, Yuhan Cui, Di Wu, Guyjoon Hwang, Alexander Getka, Ashish Singh, George Aidinis, Kyunglook Baik, Randa Melhem, Elizabeth Mamourian, Jimit Doshi, Ashwini Davison, Ilya M. Nasrallah, Christos Davatzikos, Alzheimer’s Disease Neuroimaging Initiative, and iSTAGING Consortium, Dlmuse: Robust brain segmentation in seconds using deep learning, Radiology: Artificial Intelligence 7 (2025), no. 6, e240299.
  3. [3]Alexandre Belloni and Victor Chernozhukov, Least squares after model selection in high-dimensional sparse models, Bernoulli 19 (2013), no. 2, 521–547.
  4. [4]Anindya Bhadra, Jyotishka Datta, Yunfan Li, and Nicholas G. Polson, Horseshoe regularisation for machine learning in complex and deep models, International Statistical Review 88 (2020), no. 2, 302–320.arxiv.org/abs/1904.10939
  5. [5]Anindya Bhadra, Jyotishka Datta, Nicholas G. Polson, and Brandon Willard, Lasso meets horseshoe: A survey, Statistical Science 34 (2019), no. 3, 405–427.
  6. [6]Jonathan Boss, Jyotishka Datta, Xin Wang, Sung Kyun Park, Jian Kang, and Bhramar Mukherjee, Group inverse-gamma gamma shrinkage for sparse linear models with block-correlated regressors, Bayesian Analysis 19 (2024), no. 3, 785–814.DOI
  7. [7]Philip J. Brown, Marina Vannucci, and Tom Fearn, Multivariate Bayesian variable selection and prediction, Journal of the Royal Statistical Society: Series B (Statistical Methodology) 60 (1998), no. 3, 627–641.DOI
  8. [8]Carlos M. Carvalho, Nicholas G. Polson, and James G. Scott, The horseshoe estimator for sparse signals, Biometrika 97 (2010), no. 2, 465–480.DOI
  9. [9]Carmen Echávarri, Pauline Aalten, Harry B. M. Uylings, Heidi I. L. Jacobs, Pieter Jelle Visser, Ed H. B. M. Gronenschild, Frans R. J. Verhey, and Saartje Burgmans, Atrophy in the parahippocampal gyrus as an early biomarker of Alzheimer’s disease, Brain Structure and Function 215 (2011), no. 3, 265–271.
  10. [10]Jianqing Fan and Jinchi Lv, A selective overview of variable selection in high dimensional feature space, Statistica Sinica 20 (2010), no. 1, 101–148.pubmed.ncbi.nlm.nih.gov/21572976
  11. [11]Yong Fan, Nematollah Batmanghelich, Chris M. Clark, Christos Davatzikos, and the Alzheimer’s Disease Neuroimaging Initiative, Spatial patterns of brain atrophy in MCI patients, identified via high-dimensional pattern classification, predict subsequent cognitive decline, Neuroimage 39 (2008), no. 4, 1731–1743.pubmed.ncbi.nlm.nih.gov/18053747
  12. [12]Philippe Flajolet, Xavier Gourdon, and Philippe Dumas, Mellin transforms and asymptotics: Harmonic sums, Theoretical Computer Science 144 (1995), 3–58.DOI
  13. [13]Andrew Gelman, John B. Carlin, Hal S. Stern, David B. Dunson, Aki Vehtari, and Donald B. Rubin, Bayesian data analysis, 3 ed., Chapman & Hall/CRC Press, Boca Raton, FL, 2013.
  14. [14]Subhashis Ghosal and Aad van der Vaart, Convergence rates of posterior distributions for noniid observations, The Annals of Statistics 35 (2007), no. 1, 192–223.DOI
  15. [15]Oskar Hansson, Henrik Zetterberg, Peder Buchhave, Elisabet Londos, Kaj Blennow, and Lennart Minthon, Association between CSF biomarkers and incipient Alzheimer’s disease in patients with mild cognitive impairment: a follow-up study, The Lancet Neurology 5 (2006), no. 3, 228–234.pubmed.ncbi.nlm.nih.gov/16488378
  16. [16]David A. Harville, Kronecker products and the vec and vech operators, Matrix Algebra From a Statistician’s Perspective (David A. Harville, ed.), Springer, New York, NY, 1997, pp. 337–378.
  17. [17]Andrew Hoopes, Jocelyn S. Mora, Adrian V. Dalca, Bruce Fischl, and Malte Hoffmann, SynthStrip: Skull-stripping for any brain image, NeuroImage 260 (2022), 119474.DOI
  18. [18]Clifford R. Jack Jr., David A. Bennett, Kaj Blennow, Maria C. Carrillo, Billy Dunn, Samantha Budd Haeberlein, David M. Holtzman, William Jagust, Frank Jessen, Jason Karlawish, Enchi Liu, Jose Luis Molinuevo, Thomas Montine, Creighton Phelps, Katherine P. Rankin, Christopher C. Rowe, Philip Scheltens, Eric Siemers, Heather M. Snyder, Reisa Sperling, Contributors, Cerise Elliott, Eliezer Masliah, Laurie Ryan, and Nina Silverberg, NIA-AA research framework: Toward a biological definition of Alzheimer’s disease, Alzheimer’s & Dementia 14 (2018), no. 4, 535–562.DOI
  19. [19]Adrienne Kline, Hanyin Wang, Yikuan Li, Saya Dennis, Meghan Hutch, Zhenxing Xu, Fei Wang, Feixiong Cheng, and Yuan Luo, Multimodal machine learning in precision health: A scoping review, npj Digital Medicine 5 (2022), 171.DOI
  20. [20]Ylva K öhncke, Sandra Düzel, Myriam C Sander, Ulman Lindenberger, Simone Kühn, and Andreas M Brandmaier, Hippocampal and parahippocampal gray matter structural integrity assessed by multimodal imaging is associated with episodic memory in old age, Cerebral Cortex 31 (2021), no. 3, 1464–1477.DOI
  21. [21]Debamita Kundu, Riten Mitra, and Jeremy T. Gaskins, Bayesian variable selection for multioutcome models through shared shrinkage, Scandinavian Journal of Statistics 48 (2021), no. 1, 295–320.DOI
  22. [22]Ann J. Lee, Jessica H. Stark, and Scott M. Hayes, Baseline frontoparietal gray matter volume predicts executive function performance in aging and mild cognitive impairment at 24-month follow-up, Journal of Alzheimer’s Disease 100 (2024), no. 1, 357–374.
  23. [23]Ge Li, Steven P. Millard, Elaine R. Peskind, Jing Zhang, Chang-En Yu, James B. Leverenz, Cynthia Mayer, Jane S. Shofer, Murray A. Raskind, Joseph F. Quinn, Douglas R. Galasko, and Thomas J. Montine, Cross-sectional and longitudinal relationships between cerebrospinal fluid biomarkers and cognitive function in people without cognitive impairment from across the adult life span, JAMA Neurology 71 (2014), no. 6, 742–751.DOI
  24. [24]Benoit Liquet, Kerrie L. Mengersen, Anthony N. Pettitt, and Matthew Sutton, Bayesian variable selection regression of multivariate responses for group data, Bayesian Analysis 12 (2017), no. 4, 1039–1067.
  25. [25]Roderick J. A. Little and Donald B. Rubin, Statistical analysis with missing data, 3 ed., John Wiley & Sons, Hoboken, NJ, 2019.
  26. [26]Enes Malakic and Daniel F. Schmidt, A simple sampler for the horseshoe estimator, IEEE Signal Processing Letters 23 (2016), no. 1, 179–182.
  27. [27]Elizabeth E. Moore, Dandan Liu, Corey W. Brown, Hailey A. Kresge, Deepak K. Gupta, Kimberly R. Pechman, Lisa A. Mendes, L. Taylor Davis, Katherine A. Gifford, Adam W. Anderson, Thomas J. Wang, Bennett A. Landman, Timothy J. Hohman, and Angela L. Jefferson, Lower cardiac output is associated with neurodegeneration among older adults with normal cognition but not mild cognitive impairment, Brain Imaging and Behavior 15 (2020), 2040–2050.DOI
  28. [28]Shubhbrata Mukherjee, Seo-Eun Choi, Michael L. Lee, Phoebe Scollard, Emily H. Trittschuh, Jesse Mez, Andrew J. Saykin, Laura E. Gibbons, R. Elizabeth Sanders, Andrew F. Zaman, Merilee A. Teylan, Walter A. Kukull, Lisa L. Barnes, David A. Bennett, Andrea Z. Lacroix, Eric B. Larson, Michael Cuccaro, Shannon Mercado, Logan Dumitrescu, Timothy J. Hohman, and Paul K. Crane, Cognitive domain harmonization and coc calibration in studies of older adults, Neuropsychology 37 (2023), no. 4, 409–423.DOI
  29. [29]Vesa Närhi, Seppo Laaksonen, Risto Hietala, Timo Ahonen, and Heikki Lyyti, Treating missing data in a clinical neuropsychological dataset—data imputation, The Clinical Neuropsychologist 15 (2009), no. 4, 343–358.
  30. [30]Stéphanie van der Pas, Botond Szabó, and Aad van der Vaart, Uncertainty quantification for the horseshoe (with discussion), Bayesian Analysis 12 (2017), no. 4, 1221–1274.DOI
  31. [31]Alba Pérez-Cordón, Gemma Monté-Rubio, Angela Sanabria, Octavio Rodriguez-Gomez, Sergi Valero, Carla Abdelnour, Marta Marquié, Ana Espinosa, Gemma Ortega, Isabel Hernandez, Maite Rosende-Roca, Liliana Vargas, Ana Mauleón, Silvia Gil, Juan Pablo Tartari, Francisco Lormeña, Francisco Campos, Assumpta Vivas, Marta Gomez-Chiari, Alba Benaque, Agustin Ruiz, Luis Tárraga, Mercè Boada, and Montserrat Alegret, Subtle executive deficits are associated with higher brain amyloid burden and lower cortical volume in subjective cognitive decline: the FACEHBI cohort, Scientific Reports 10 (2020), no. 1, 17721.
  32. [32]Juho Piironen and Aki Vehtari, Sparsity information and regularization in the horseshoe and other shrinkage priors, Electronic Journal of Statistics 11 (2017), no. 2, 5018–5051.DOI
  33. [33]Lorena Rami, Juan Fortea, Beatriz Bosch, Cristina Solé-Padullés, Albert Lladó, Alex Iranzo, Raquel Sánchez-Valle, and Jose Luis Molinuevo, Cerebrospinal fluid biomarkers and memory present distinct associations along the continuum from healthy subjects to AD patients, Journal of Alzheimer’s Disease 23 (2011), no. 2, 319–326.DOI
  34. [34]Jaroslav Rokicki, Thomas Wolfers, Wibeke Nordhøy, Natalia Tesli, Daniel S. Quintana, Dag Alnæs, Genevieve Richard, Ann-Marie G. de Lange, Martina J. Lund, Linn Nørbom, Ingrid Agartz, Ingrid Melle, Terje Nærland, Geir Selbæk, Karin Persson, Jan E. Nordvik, Emanuel Schwar, Ole A. Andreassen, Tobias Kaufmann, and Lars T. Westlye, Multimodal imaging improves brain age prediction and reveals distinct abnormalities in patients with psychiatric and neurological disorders, Human Brain Mapping 42 (2021), no. 6, 1714–1726.
  35. [35]Donald B. Rubin, Multiple imputation for nonresponse in surveys, John Wiley & Sons, Inc., Hoboken, NJ, 1987.
  36. [36]Martina Sattlecker, Steven J. Kiddle, Stephen Newhouse, Petroula Proitsi, Sally Nelson, Stephen Williams, Caroline Johnston, Richard Killick, Andrew Simmons, Eric Westman, Angela Hodges, Hilkka Soininen, Iwona Kłoszewska, Patrizia Mecocci, Magda Tsolaki, Bruno Vellas, Simon Lovestone, AddNeuroMed Consortium, and Richard J. B. Dobson, Alzheimer’s disease biomarker discovery using SOMAscan multiplexed protein technology, Alzheimer’s & Dementia 10 (2014), no. 6, 724–734.DOI
  37. [37]Qifan Song and Faming Liang, Nearly optimal Bayesian shrinkage for high-dimensional regression, Science China Mathematics 66 (2023), 409–442.
  38. [38]Daniel Suen and Yen-Chi Chen, Modeling missing at random neuropsychological test scores using a mixture of binomial product experts, Psychometrika 91 (2026), no. 1, 223–246.
  39. [39]Martin A. Tanner and Wing Hung Wong, The calculation of posterior distributions by data augmentation, Journal of the American Statistical Association 82 (1987), no. 398, 528–540.
  40. [40]Robert Tibshirani, Regression shrinkage and selection via the Lasso, Journal of the Royal Statistical Society: Series B (Methodological) 58 (1996), no. 1, 267–288.
  41. [41]Stef van Buuren and Karin Groothuis-Oudshoorn, mice: Multivariate imputation by chained equations in R, Journal of Statistical Software 45 (2011), no. 3, 1–67.
  42. [42]Janani Venugopalan, Li Tong, Hamid Reza Hassanzadeh, and May D. Wang, Multimodal deep learning models for early detection of Alzheimer’s disease stage, Scientific Reports 11 (2021), 3254.DOI
  43. [43]Sandra Weintraub, Lilah Besser, Hiroko H. Dodge, Merilee Teylan, Steven Ferris, Felicia C. Goldstein, Bruno Giordani, Joel Kramer, David Loewenstein, Dan M arson, Dan Mungas, David Salmon, Kathleen Welsh-Bohmer, Xiao-Hua Zhou, Steven D. Shirk, Alireza Atri, Walter A. Kukull, Creighton Phelps, and John C. Morris, Version 3 of the Alzheimer disease centers’ neuropsychological test battery in the uniform data set (UDS), Alzheimer Disease & Associated Disorders 32 (2018), no. 1, 10–17.DOI
  44. [44]Xiaofan Xu and Malay Ghosh, Bayesian variable selection and estimation for group lasso, Bayesian Analysis 10 (2015), no. 4, 909–936.DOI
  45. [45]Zemei Xu, Daniel F. Schmidt, Enes Malakic, Guoqi Qian, and John L. Hopper, Bayesian grouped horseshoe regression with application to additive models, Proceedings of AI 2016: Advances in Artificial Intelligence (Cham, Switzerland) (B yeong Ho Kang and Quan Bai, eds.), Springer, 2016, pp. 299–240.
  46. [46]Ming Yuan and Yi Lin, Model selection and estimation in regression with grouped variables, Journal of the Royal Statistical Society: Series B (Statistical Methodology) 68 (2006), no. 1, 49–67.DOI
  47. [47]Arnold Zellner, An efficient method of estimating seemingly unrelated regressions and tests for aggregation bias, Journal of the American Statistical Association 57 (1962), no. 298, 348–368.
  48. [48]Ruoyang Zhang and Malay Ghosh, Ultra high-dimensional multivariate posterior contraction rate under shrinkage priors, Journal of Multivariate Analysis 187 (2022), 104835.

Paper details

vlnc:W_8sdzu884
Version
v1 · latest
Published
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents