Off-campus UNL users: To download campus access dissertations, please use the following link to log into our proxy server with your NU ID and password. When you are done browsing please remember to return to this page and log out.
Non-UNL users: Please talk to your librarian about requesting this dissertation through interlibrary loan.
Methods to Incorporate Global Information in Local Models to Offset Sample Size Constraints
The application of local models has proven to be successful in many areas. The localmodels allow flexible modeling of an otherwise heterogeneous population by identifyingpockets of homogeneity. So, identifying such homogeneous groups and buildingseparate local models on them often increase the performance of statistical models.However, compared to developing a global model- where the structures are not fullyexplained, building local models suffer from a crucial drawback, i.e., lack of sufficientdata in each relatively homogeneous group. This lack of group-specific sample canlead to lack of precision in the estimates of group-specific parameters. To circumventthis issue and thus improve the performance of the local model, information can beborrowed from the global structure. In this dissertation work, utilizing three studies,we have shown ways to incorporate global information in local models to offset samplesize constraints. We offer the general theme of the studies conducted in this dissertation workin chapter 1. In chapter 2, we develop two robust frameworks to analyze the dose timeresponse curves in the precision medicine setting. Here the information acrosssubjects (cell-lines) is borrowed to predict the dose-response curves for new cell-linesat different time points. We demonstrate that the borrowing of information leadsto narrower point-wise predictive intervals for the dose-response curves. Chapter 3focuses on gene selection under time-varying transcriptomic data in the context of thenutritional content of rice. We couple global shrinkage with a local Stochastic SearchVariable Selection (SSVS) model to perform a computationally efficient group-specificgene selection. In chapter 4, we discuss boosting the performance of classificationmodels through metric learning and data sub-setting. We demonstrate that if weidentify stable local structures in high dimensional labeled data, a metric learningassisted clustering algorithm performs at least as good as a classifier. We anticipatethat, in molecular structure-based drug screening, our algorithm can induce isometryin structure -activity relationship in a transformed topological space.
Kesawan, Ramesh Aravind, "Methods to Incorporate Global Information in Local Models to Offset Sample Size Constraints" (2022). ETD collection for University of Nebraska-Lincoln. AAI29323193.