Dissertation titled 'Recovering Latent Factors in Data with Strong and Weak Signals: From Theory to Practice’ . This dissertation develops statistical methods and applications for high-dimensional biomedical data with latent structure. The first part introduces ShaC-PA, a two-stage Parallel Analysis procedure for factor selection that addresses eigenvalue shadowing and improves recovery of weaker but perceptible latent factors. The method is supported by theoretical guarantees, simulations, and applications to gene expression and metabolomics data. The second part studies latent confounder adjustment in metabolomic association analysis for atherosclerotic cardiovascular disease (ASCVD), showing how accounting for metabolite dependence can improve statistical discovery and risk prediction. The third part further investigates metabolomic profile heterogeneity across Life’s Essential 8 components and constructs a metabolomic age measure as a marker of cardiovascular risk. Together, these studies first address dimension estimation for latent structures and dependence, and then demonstrate how estimating and adjusting for such structures can improve biological interpretation and cardiovascular risk modeling in complex biomedical datasets.

Advisor:Dr.Chris McKennan