Statistical methods for handling missingness and confounding in UK biobank population brain imaging

Thesis

UK Biobank (UKB) brain imaging data is a one-of-a-kind resource for studying the links between the brain and demographic-, lifestyle- and genetic data. These links are often established via statistical modelling of the relation between Imaging Derived Phenotypes (IDPs) and non-IDPs (nIDPs). This type of analysis has been the source of multiple groundbreaking medical and epidemiological discoveries. There are, however, considerable challenges when establishing such links. This thesis focuses on solving two problems: missing data and confounding, for analysis linking IDPs and nIDPs. In the case of missing data, we identify the presence of Structured Missingness (SM) in the nIDPs, caused by the study design of UKB. We define a generative model for missing data in the nIDPs, and evaluate existing missing data methodology on synthetic data with access to the ground truth. Further, we develop methodology for performing Canonical Correlation Analysis (CCA) that is tailor-made to handle the type of missingness that we observed in the UKB brain imaging cohort. In the case of confounds, we create a fast and memory efficient pipeline for constructing UKB brain imaging confounds, generating a lower number of confounds compared to the previous, established pipeline (Alfaro-Almagro et. al. 2021). We also identify appropriate methodology for deconfounded inference when the number of confounds exceeds the number of subjects, enabling analysis of small subsets of subjects. The work presented in this thesis enables new types of analyses using the UKB brain imaging cohort, leading to deeper scientific understanding of the links between the human brain and everything around it.

Authors: Radosavljevic, L

• DOI: 10.5287/ora-9rq4eagj5
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: missing data; brain imaging; confounding; population health
• Subjects: Missing observations (Statistics); popul