The Validity and Statistical Power of the Case-Only Study Design for Interaction Analysis : Gene-Gene Interaction and the Role of Genotype Imputation in Gene-Environment Interaction
In search of the origin of complex human diseases such as inflammatory bowel disease (IBD) and Parkinson disease (PD), not only are genetic and environmental factors thought to play a role, but gene-gene (G×G) and gene-environment (G×E) interactions may also contribute to the disease etiology. However, examining these interactions is challenging, as high statistical power is needed in order to detect them, especially when the effects are small and single nucleotide polymorphisms (SNPs) with low minor allele frequencies (MAFs) are examined. In epidemiological studies, the traditional case-control (CC) design is often employed, however, it often does not achieve the necessary statistical power for interaction analysis. The case-only (CO) study design proves to be of great use in these circumstances, as it not only obviates the need for controls, but given the same number of cases, it is statistically more powerful than the CC study design. However, two key assumptions must be fulfilled in order for the CO study design to be valid: (i) the disease of interest must be sufficiently rare, and (ii) the two risk factors (gene and environment in case of G×E interaction, both genetic in G×G interaction) must be independent in the general population. Nevertheless, the practical implementation of the CO study design in the context of G×G interaction analysis remained unexplored. Another aspect that increases the statistical power to detect interaction effects is the number of observations available for the analysis. Thus, combined data from the largest consortia comprising of numerous centers and thousands of cases gives the highest possible chance of detecting interactions to date. Depending on the center, a different genotyping chip is often used resulting in different genotyped SNPs. Genotype imputation uses a reference database to impute missing data and thus allows to gain information on numerous SNPs and make the analysis of data from different centers possible. It is a standard procedure in genome-wide-association-studies which analyse genetic main effects (MEs). The reference base used for genotype imputation is population based and assumed to consist of healthy individuals, therefore, their linkage disequilibrium (LD) structure may differ from diseased cases, particularly in areas with MEs. Thus, whether genotype imputation has an impact on the validity and statistical power of statistical tests for G×E interactions in CO studies would be a useful asset in the analysis, yet was unknown. This thesis examined two aspects of interaction analysis in the CO study design. First, whether imputing data from a reference base consisting of healthy individuals into diseased cases has consequences for the downstream G×E interaction analysis. The results showed, that imputation does not work well in areas with MEs and low minor allele frequencies of SNPs. The lower the LD to neighbouring SNPs was, the more the MAF resembled the reference base controls than the cases from the used dataset. This imputation bias further led to a loss of statistical power in the G×E interaction analysis. The second aspect of this thesis is the practical implementation of G×G interaction analysis in which SNPs were considered as proxies for genes. The (ii) assumption of independence of both factors is problematic in G×G interactions due to LD. Moreover, computational issues arise due to the large number of possible genome-wide interaction pairs that, given more than one center, need to be calculated separately for each. Thus, a method was proposed that practically implements G×G interaction analysis. The method includes, among others aspects, analysing SNPs on different chromosomes or chromosome arms to fulfil the (ii) assumption and focusing on SNPs with known MEs in order to reduce the computational burden. The largest available datasets for IBD and PD to date were used for the analysis of G×G interactions for these complex diseases. While the G×G interaction analysis for IBD found G×G interactions to be scarce, it yielded 10 unique significant G×G interactions for PD after multiple test correction. The findings of this thesis will add to an improved understanding of G×E and G×G interaction analysis in the CO study design. It points out areas of caution when examining G×E interaction using imputed data. Furthermore, this work shows how G×G interaction can be implemented in a statistically sound and computationally efficient manner. This could lead to further G×G interaction analyses, opening doors to more in-depth knowledge on the etiology of complex human diseases.
Use and reproduction:
No license. The provisions of the German Copyright Act (UrhG) apply.
Please note that individual components of the publication may be subject to other licensing or copyright conditions.