vbdm: fit a discrete mixture model

In vbdm: Variational Bayes Discrete Mixture Model

Description

Fits a discrete mixture model for rare variant association analysis. Uses an approximate variational Bayes coordinate ascent algorithm for a computationally efficient solution.

Usage

vbdm(y, G, X=NULL, thres=0.05, genotypes=TRUE,
     include.mean=TRUE, minor.allele=TRUE, impute="MEAN",
     eps=1e-4, scaling=TRUE, nperm=0, maxit=1000, hyper=c(2,2))

Arguments

y	A vector of continuous phenotypes.
G	A matrix of genotypes or variables of interest. Rows are treated as individuals and columns are treated as genotypes.
X	An optional matrix of covariates.
thres	If the matrix is of genotypes, then this specifies a minor allele frequency threshold. Variants with a MAF greater than this threshold are excluded from the analysis.
genotypes	This specifies whether or not to treat G as a matrix of genotypes. If it is treated as genotypes then it will be filtered based on thres, and there are more options for missing data imputations. The default genotype encoding is additive (e.g. genotypes are encoded as 0,1,2). Also if G is a genotype matrix vbdm will flip the encoding such that the homozygous major allele genotype is encoded as 0, the heterozygote as 1, and the homozygous minor allele genotype as 2 unless minor.allele=FALSE
include.mean	This specifies whether to add an interecept term to the model. If no covariates are provided it is automatically added, but if there are covariates provided it can be optional.
minor.allele	When minor.allele=TRUE and genotypes=TRUE the genotypes are flipped so that the major allele genotype is encoded as 0.
impute	If there is missing data in G this specifies the method with which to impute the missing data. There are two options impute="MEAN" which sets any missing genotype to the expected dosage given the MAF, or impute="MAJOR" which sets any missing genoypte to the homozygous genotype of the major allele. If the matrix is not treated as a genotype matrix (e.g. genotype=FALSE), then only impute="MEAN" will work. Also, missing data is not allowed in the covariates X.
eps	The tolerance for convergence of the coordinate ascent algorithm based on the change in the lower bound of the log marginal likeilhood.
scaling	Whether or not to scale the genotypes to have mean 0 and variance 1.
nperm	Optional parameter defining the number of null permutations of the vbdm likelihood ratio test. This can be used to generate an exact p-value
maxit	The maximum number of iterations allowed for the vbdm algorithm.
hyper	The hyperparameters for the prior defined over the mixing probability parameter. The first hyperparameter is the alpha parameter, and the second is the beta parameter.

Value

y	The phenotype vector passed to vbdm.
G	The genotype matrix passed to vbdm. Note that any variables that were dropped will be dropped from this matrix.
X	The covariate matrix passed to vbdm. Will include intercept term if it was added earlier.
keep	A vector of indices of the kept variables in G (if any were excluded based on thres)
pvec	The vector of estimated posterior probabilities for each variable in G.
gamma	A vector of additive covariate effect estimates.
theta	The estimated effect of the variables in G.
sigma	The estimated error variance.
prob	The estimated mixing parameter.
lb	The lower bound of the marginal log likelihood.
lbnull	The lower bound of the marginal log likelihood under the null model.
lrt	The approximate likelihood ratio test based on the lower bounds.
p.value	A p-value computed based on lrt with the assumption that lrt~chi^2_1
lbperm	If nperm>0, the lower bound of the fitted null permutations.
lrtperm	If nperm>0, the likelihood ratio test of the fitted null permutations.
p.value.perm	If nperm>0, the empirical p-value based on the fitted null permutations.

Author(s)

Benjamin A. Logsdon ([e-mail])

References

Logsdon, B.A., et al. (2014) A Variational Bayes Discrete Mixture Test for Rare Variant Association., Genetic Epidemiology, Vol. 38(1), 21-30 2014

See Also

vbdmR,burdenPlot

Examples

#generate some test data
library(vbdm)
set.seed(3)
n <- 1000
m <- 20
G <- matrix(rbinom(n*m,2,.01),n,m);
beta1 <- rbinom(m,1,.2)
y <- G%*%beta1+rnorm(n,0,1.3)

#with scaling:
res <- vbdm(y=y,G=G);
T5 <- summary(lm(y~rowSums(scale(G))))$coef[2,4];
cat('vbdm p-value:',res$p.value,'\nT5 p-value:',T5,'\n')
#vbdm p-value: 0.001345869 
#T5 p-value: 0.9481797 

#without scaling:
res <- vbdm(y=y,G=G,scaling=FALSE)
T5 <- summary(lm(y~rowSums(G)))$coef[2,4];
cat('vbdm p-value:',res$p.value,'\nT5 p-value:',T5,'\n')
#vbdm p-value: 0.0005315836 
#T5 p-value: 0.904476 

#run 100 permutations
set.seed(2)
res <- vbdm(y=y,G=G,scaling=FALSE,nperm=1e2);
cat('vbdm approximate p-value:',res$p.value,'\nvbdm permutation p-value <',res$p.value.perm,'\n');
#vbdm approximate p-value: 0.0005315836 
#vbdm permutation p-value: 0 

vbdm documentation built on May 2, 2019, 2:37 a.m.