Covariance-insured screening

He, Kevin; Kang, Jian; Hong, Hyokyoung Grace; Zhu, Ji; Liu, Yanming; Lin, Huazhen; Han, Xu; Li, Yang

doi:10.1016/j.csda.2018.09.001

Cited by 12 publications

(14 citation statements)

References 37 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…To incorporate the correlation information, He et al () proposed compartmentalizing covariates into blocks so that variables from distinct blocks are less correlated. The algorithm starts from the idea of thresholding.…”

Section: Methodsmentioning

confidence: 99%

“…The success of SIS relies on a fundamental assumption that the true association between the individual predictors and the response can be inferred from their marginal associations. To account for the violation of this assumption, recent researches have expressed a growing interest in conducting multivariate screenings (Cho & Fryzlewicz, ; Cui, Li, & Zhong, ; Hall & Miller, ; He et al, ; Jin, Zhang, & Zhang, ; Kang, Hong, & Li, ; Li, Peng, Zhang, & Zhu, ; Wang & Leng, ; Zhu, Li, Li, & Zhu, ). In particular, according to the reports by the authors, the high‐dimensional ordinary‐least squares projection (HOLP; Wang & Leng, ) substantially improves variable screening accuracy of SIS for many scenarios with theoretical supports.…”

Section: Introductionmentioning

confidence: 99%

“…The partition‐based screening (PartS; Kang et al, ) can integrate prior grouping knowledge into the variable screening with solid theoretical foundation and it achieves a better performance in spatial variable screening with neuroimaging applications. Similarly, covariance‐insured screening (CIS; He et al, ) can effectively take advantage of the sparse block diagonal covariance structure of the ultrahigh‐dimensional predictors (if any) and produce more accurate variable screening.…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

A selective overview of feature screening methods with applications to neuroimaging data

Han

Kang

2018

WIREs Computational Stats

Self Cite

View full text Add to dashboard Cite

In neuroimaging studies, regression models are frequently used to identify the association of the imaging features and clinical outcome, where the number of imaging features (e.g., hundreds of thousands of voxel-level predictors) much outweighs the number of subjects in the studies. Classical best subset selection or penalized variable selection methods that perform well for low-or moderate-dimensional data do not scale to ultrahigh-dimensional neuroimaging data. To reduce the dimensionality, variable screening has emerged as a powerful tool for feature selection in neuroimaging studies. We present a selective review of the recent developments in ultrahigh-dimensional variable screening, with a focus on their practical performance on the analysis of neuroimaging data with complex spatial correlation structures and high-dimensionality. We conduct extensive simulation studies to compare the performance on selection accuracy and computational costs between the different methods. We present analyses of resting-state functional magnetic resonance imaging data in the Autism Brain Imaging Data Exchange study. correlated covariates, imaging data analysis, linear regression, variable screening 1 | INTRODUCTION Recent advances in neuroimaging technology have generated high-resolution brain imaging data that can measure brain functions and structures with increasing accuracy. This provides unprecedented opportunities for researchers to precisely identify the important brain regions that are strongly associated with certain clinical symptoms, which will have a great impact on public health and precision medicine. To this end, a class of regression models has been widely used, where the response variable is the clinical outcome of interest and the predictors include imaging features. We refer to this model as scalar-on-image regression. Performing variable selection in scalar-on-image regression directly identifies the brain regions of interest. However, in a typical neuroimaging study, the three-dimensional brain image may involve up to millions of voxels, while the number of subjects is usually in a range of hundreds to thousands. Thus, for voxel-level selection in the scalar-on-image regression, the number of predictors is often on the exponential order of the sample size. Due to the computational infeasibility, some classical variable selection methods, such as the best subset selection, are not directly applicable to this setting. Similarly, regularization-based variable selection methods have been extensively studied during the past decades, and have many

show abstract

Section: Methodsmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

A selective overview of feature screening methods with applications to neuroimaging data

Han

Kang

2018

WIREs Computational Stats

Self Cite

View full text Add to dashboard Cite

show abstract

“…An accurate fine-mapping model requires exact knowledge of the LD pattern among the SNPs within a GWAS locus [4,7,8,13,14]. Generally, the LD pattern refers to two aspects: i) the local aspect, pairwise LD scores between SNPs which are directly available, and ii) the global aspect, the (latent) network topological structure (e.g., the community structure) of the LD matrix.…”

Section: Introductionmentioning

confidence: 99%

“…Generally, the LD pattern refers to two aspects: i) the local aspect, pairwise LD scores between SNPs which are directly available, and ii) the global aspect, the (latent) network topological structure (e.g., the community structure) of the LD matrix. A large body of literature on multivariate statistics has shown that the accuracy of variable selection relies on the accurate knowledge of the network structure of the dependence between predictors [14][15][16][17][18][19]. However, the network structure of the LD pattern measured using the commonly used haplotype block methods cannot accurately reveal the latent network structure of the LD pattern.…”

Section: Introductionmentioning

confidence: 99%

Genetic Fine-mapping with Dense Linkage Disequilibrium Blocks: genetics of nicotine dependence

Hatch

et al. 2020

Preprint

View full text Add to dashboard Cite

Fine-mapping is an analytical step to perform causal prioritization of the polymorphic variants on a trait-associated genomic region observed from genome-wide association studies (GWAS). The prioritization of causal variants can be challenging due to the linkage disequilibrium (LD) patterns among hundreds to thousands of polymorphisms associated with a trait. We propose a novel ℓ0 graph norm shrinkage algorithm to select causal variants from dense LD blocks consisting of highly correlated SNPs that may not be proximal or contiguous. We extract dense LD blocks and perform regression shrinkage to calculate a prioritization score to select a parsimonious set of causal variants. Our approach is computationally efficient and allows performing fine-mapping on thousands of polymorphisms. We demonstrate its application using a large UK Biobank (UKBB) sample related to nicotine addiction. Our results suggest that polymorphic variances in both neighboring and distant variants can be consolidated into dense blocks of highly correlated loci. Simulations were used to evaluate and compare the performance of our method and existing fine-mapping algorithms. The results demonstrated that our method outperformed comparable fine-mapping methods with increased sensitivity and reduced false-positive error rate regarding causal variant selection. The application of this method to smoking severity trait in UKBB sample replicated previously reported loci and suggested the causal prioritization of genetic effects on nicotine dependency.Author summaryDisentangling the complex linkage disequilibrium (LD) pattern and selecting the underlying causal variants have been a long-term challenge for genetic fine-mapping. We find that the LD pattern within GWAS loci is intrinsically organized in delicate graph topological structures, which can be effectively learned by our novel ℓ0 graph norm shrinkage algorithm. The extracted LD graph structure is critical for causal variant selection. Moreover, our method is less constrained by the width of GWAS loci and thus can fine-map a massive number of correlated SNPs.

show abstract