Browsing by Subject "Bayesian variable selection"

A primary goal of human genetics research is the investigation of associations between genetic variants and diseases. Due to the high number of genetic variants, sophisticated statistical methods for high dimensional data are required. A genome-wide association study (GWAS) is the initial analysis used to measure the marginal associations between genetic variants and biological traits, but because it ignores correlation between variants, identification of truly causal variants remains difficult. Fine-mapping refers to the statistical methods that aim to identify causal variants from GWAS results by incorporating information about correlation between variants. One such fine-mapping method is FINEMAP, a widely used Bayesian variable selection model. To make computations efficient, FINEMAP assumes a constant sample size for the measured genetic variants, but in a meta-analysis that combines data from several studies, this assumption may not hold. This results in miscalibration of the FINEMAP model with meta-analyzed data. In this thesis, a novel extension for FINEMAP is developed, named FINEMAP-MISS. With an additional inversion of the variants' correlation matrix and other less demanding computational adjustments, FINEMAP-MISS makes it possible to fine-map meta-analyzed GWAS data. To test the effectiveness of FINEMAP-MISS, genetic data from the UK Biobank is used to generate sets of simulated data, where a single variant has a non-zero effect on the generated trait. For each simulated dataset, a meta-analysis with missing information is emulated, and fine-mapping is performed with FINEMAP and FINEMAP-MISS. The results verify that with missing data FINEMAP-MISS clearly performs better than FINEMAP in identification of causal variants. Additionally, with missing data the posterior probability estimates provided by FINEMAP-MISS are properly calibrated, whereas the estimates by FINEMAP exhibit miscalibration. FINEMAP-MISS enables the use of fine-mapping for meta-analyzed genetic studies, allowing for greater power in the detection of causal genetic variants.