MOST: Multivariate Outcome Score Test
Genetic association studies often collect data on multiple traits that are correlated. Discovery of genetic variants influencing multiple traits can lead to better understanding of the etiology of complex human diseases. Conventional univariate association tests may miss variants which have weak or moderate effects on individual traits. We propose several multivariate test statistics to complement univariate tests. Our framework covers both studies of unrelated individuals and family studies and allows any type/mixture of traits. We construct score-type statistics, which are computationally fast and numerically stable even in the presence of covariates and which can be combined efficiently across studies with different designs and arbitrary patterns of missing data. We also provide a strategy to determine genomewide significance that properly accounts for the linkage disequilibrium (LD) of genetic variants.
Our MOST will output three multivariate statistics, Q, T, and T’. The T test is more powerful when the genetic effect sizes are close to each other across different traits; the T’ test is more powerful when the standardized genetic effects are similar to each other; the Q test is more powerful when the genetic effects vary dramatically (possibly with opposite effects) for different traits. If one already knows that some of the traits are ‘healthy’ traits and the other traits are ‘unhealthy’ traits (for example, HDL versus LDL), the user should arrange the multiple traits in the plausibly same direction (this can be done by multiply each of the ‘unhealthy’ traits by -1). For the user’s convenience, MOST will also output the univariate analysis results for each trait. MOST is written in C.
MOST [-geno geno] [-ngeno ngeno] [-pheno pheno] [-npheno npheno] [-trait trait][-covar covar][-bin bin] [-monte monte]
In the above genotype file, FID refers to Family ID. For data with no family structures, one needs to simply copy the column of ID to the column of FID.
The example files are in the download.
1) If one is not interested in obtaining the Monte-Carlo threshold, then use the following command
MOST -geno geno.txt -ngeno 1000 -pheno pheno_and_covariates.txt -npheno 1000 -trait 2 -covar 2 -bin 1 0
-geno geno.txt” specifies the name of the genotype file.
-ngeno 1000” tells MOST that there are 1000 subjects in the genotype files.
-pheno” specifies the name of the phenotype file.
-npheno 1000” tells MOST that there are 1000 subjects in the phenotype files. MOST allows that the genotype file and phenotype file to have different number of subjects; it will sort out the subjects and use those overlapping subjects.
-traits 2” tells MOST that your pheno file contains 2 traits.
-covar 2” tells MOST that your pheno file contains 2 covariates.
-bin 1 0” tells MOST that the first trait is binary and the 2nd trait is continuous. If you have, say 4 traits and only the 3rd trait is binary, then you would specify
-bin 0 0 1 0“.
2) If one wishes to obtain the Monte-Carlo threshold, use the following command
MOST -geno geno.txt -ngeno 1000 -pheno pheno_and_covariates.txt -npheno 1000 -trait 2 -covar 2 -bin 1 0 -monte 1
-monte 1” tells MOST to invoke the Monte-Carlo procedure.
The example files are in the download.
The main results will be saved to Result.txt — this file contains the Q, T^2, T’^2, and the Z_k^2 for all the K traits. The p-values for these statistics are also contained in this file.
If the Monte-Carlo procedure was invoked, two files will be generated: Monte_U_and_V.txt — this file is organized in blocks, and each block contains the U-vector and the V-matrix for a SNP. Within each block, the first line is the SNP-ID, the 2nd line is the U-vector, and the remaining lines are for the V-matrix. (for information on U and V, see the Reference Paper.) This file is useful for Meta-analysis (if any).
T_max.txt — this file contains results from the Monte-Carlo procedure. The first column represents the simulation index. Each of the other columns represents the Monte-Carlo realization of a statistic. This file is useful for determining the Monte-Carlo Threshold for declaring genome-wide significance. We provide a simple R program to use the T_max.txt to get the Monte-Carlo Threshold.
Qianchuan He, Christy L. Avery, Dan-Yu Lin. (2013). A General Framework for Association Tests With Multivariate Traits. To be submitted.
MOST for 64-bit x86 based Linux [updated Jan 04, 2013]
Example files [updated Jan 04, 2013]
|1||January 4, 2013|| First version released|