Statistics for X‐chromosome associations

In a genome‐wide association study (GWAS), association between genotype and phenotype at autosomal loci is generally tested by regression models. However, X‐chromosome data are often excluded from published analyses of autosomes because of the difference between males and females in number of X chromosomes. Failure to analyze X‐chromosome data at all is obviously less than ideal, and can lead to missed discoveries. Even when X‐chromosome data are included, they are often analyzed with suboptimal statistics. Several mathematically sensible statistics for X‐chromosome association have been proposed. The optimality of these statistics, however, is based on very specific simple genetic models. In addition, while previous simulation studies of these statistics have been informative, they have focused on single‐marker tests and have not considered the types of error that occur even under the null hypothesis when the entire X chromosome is scanned. In this study, we comprehensively tested several X‐chromosome association statistics using simulation studies that include the entire chromosome. We also considered a wide range of trait models for sex differences and phenotypic effects of X inactivation. We found that models that do not incorporate a sex effect can have large type I error in some cases. We also found that many of the best statistics perform well even when there are modest deviations, such as trait variance differences between the sexes or small sex differences in allele frequencies, from assumptions.