CERAMIC: Case-Control Association Testing in Samples with Related Individuals, Based on Retrospective Mixed Model Analysis with Adjustment for Covariates

We consider the problem of genetic association testing of a binary trait in a sample that contains related individuals, where we adjust for relevant covariates and allow for missing data. We propose CERAMIC, an estimating equation approach that can be viewed as a hybrid of logistic regression and linear mixed-effects model (LMM) approaches. CERAMIC extends the recently proposed CARAT method to allow samples with related individuals and to incorporate partially missing data. In simulations, we show that CERAMIC outperforms existing LMM and generalized LMM approaches, maintaining high power and correct type 1 error across a wider range of scenarios. CERAMIC results in a particularly large power increase over existing methods when the sample includes related individuals with some missing data (e.g., when some individuals with phenotype and covariate information have missing genotype), because CERAMIC is able to make use of the relationship information to incorporate partially missing data in the analysis while correcting for dependence. Because CERAMIC is based on a retrospective analysis, it is robust to misspecification of the phenotype model, resulting in better control of type 1 error and higher power than that of prospective methods, such as GMMAT, when the phenotype model is misspecified. CERAMIC is computationally efficient for genomewide analysis in samples of related individuals of almost any configuration, including small families, unrelated individuals and even large, complex pedigrees. We apply CERAMIC to data on type 2 diabetes (T2D) from the Framingham Heart Study. In a genome scan, 9 of the 10 smallest CERAMIC p-values occur in or near either known T2D susceptibility loci or plausible candidates, verifying that CERAMIC is able to home in on the important loci in a genome scan.Author Summary: Case-control association testing has proven to be useful for identification of genetic variants that affect susceptibility to disease. One can expect to gain power for detecting such variants by including relevant covariates in the analysis, by accounting for any relatedness of sampled individuals, and by making use of partial information in the data. For analysis of continuously-varying traits, variations on linear mixed-model (LMM) approaches have proven effective at achieving some of these goals. However, for case-control or binary trait mapping, there remain significant challenges. Direct application of LMM approaches to binary traits suffers from power loss when covariate effects are strong, and existing generalized LMM approaches can perform poorly in the presence of trait model misspecification and partially missing data. We propose CERAMIC, a method for binary trait mapping, which is computationally feasible for large genome-wide studies, and which gains power over previous approaches by improved trait modeling, retrospective assessment of significance, accounting for sample structure, and making use of partially missing data. We illustrate this approach in genome-wide association mapping of type 2 diabetes in data from the Framingham Heart Study.