vbsr: fit a linear model with variational Bayes spike penalty

In vbsr: Variational Bayes Spike Regression Regularized Linear Models

Description

Fit a linear model via a fast coordinate variational Bayes algorithm. Applicable to linear and logistic regression, and solves the problem on either a path of the spike (l0) parameter or at a fixed value based on the data-dimensions.

Usage

vbsr(y,
  	X,
		ordering_mat=NULL,
		eps=1e-6,
		exclude=NULL,
		add.intercept=TRUE,
		maxit = 1e4,
		n_orderings = 10,
    family = "normal",
		scaling = TRUE,
		return_kl = TRUE,
		estimation_type = "BMA",
		bma_approximation = TRUE,
		screen = 1.0,
		post=0.95,
		already_screened = 1.0,
		kl = 0.99,
		l0_path=NULL,
    cleanSolution=FALSE)

Arguments

y	response variable. Normally distributed errors for family="normal". For family="binomial" should be coded as a vector of 0's and 1's.
X	Design matrix, an n x m matrix, with rows as observations
ordering_mat	Optionally specified coordinate update ordering matrix. Must be in matrix form with columns as permutation vectors of length m, and there must be n_orderings columns.
eps	Tolerance used to determine convergence of the algorithm based on the lower bound.
exclude	An optional indicator vector of length m of 0's and 1's indicating whether to penalize a particular variable or not (0=penalize, 1=unpenalized)
add.intercept	A boolean variable indicating whether or not to include an unpenalized intercept variable.
maxit	The maximum number of iterations to run the algorithm for a given solution to a penalized regression problem.
n_orderings	The number of random starts used.
family	The type of error model used. Currently supported modes are family="normal" and family="binomial"
scaling	A boolean variable indicating whether or not to scale the columns of X to have mean zero and variance one.
return_kl	A boolean variable indicating whether or not to return an analysis of the null distributed features in the data-set as a function of the penalty parameter.
estimation_type	The type of estimation to perform based on the number of unique solution identified to the penalized regression problem. Valid values are estimation_type="BMA" and estimation_type="MAXIMAL"

.

bma_approximation	A boolean variable indicating whether to compute a full correction to the z statistic. WARNING can make the algorithm very computationally intensive for highly multi-modal posterior surfaces.
screen	P-value to do marginal screening. Default is to not do marginal prescreening (e.g marginal p-value of 1.0)
post	Choice of penalty parameter such that a feature will have a posterior probability of 0.95 if it passes a Bonferroni correction in the multivariate model. Default is post=.95. More conservative approach would be post=0.5
already_screened	If features are already screened, the marginal p-value used for screening.
kl	The inner percentiles of the distribution to compute the Kullback-Leibler overfitting statistic. Only works for analysis when directly specifying a path of penalization parameter (e.g. l0_path). For default kl=0.99 the KL-statistic is used for the statistics between the 1%-99% of the distribution.
l0_path	The path of penalty parameters to solve the spike regression problem. If post is specified, this is computed automatically.
cleanSolution	This parameter determines whether a given solution is further filtered using an unpenalized model. If cleanSolution=TRUE, then the features that are significant after a Bonferroni correction given the p-values from the vbsr regression model are then tested in an unpenalized linear regression model. The p-values and z-statistics are updated using the Wald test from the unpenalized linear regression model for the features that were selected.

Details

The solutions to the spike penalized regression model are fit with a coordinate variational Bayes algorithm based on the l0_path values of the spike hyper-parameter.

Value

A list with all the results of the vbsr analysis.

beta	The expected value of the penalized regression coefficients.
alpha	The estimated value of the unpenalized regression coefficients.
z	The Z-statistic for each penalized regression coefficient
pval	The p-values based on the asymptotic normal assumption of the Z-statistics
post	The posterior probabilities of each of the regression coefficients
l0	The penalty parameters used to solve the penalized regression problem
modelEntropy	The entropy of the identified approximate posterior probability distribution over model space.
modelProb	The approximate posterior probability distribution over the identified model space.
kl_index	If a path solution was run with the KL diagnostic statistic then the points in the path where the KL statistic is nearest the min, the mean, the min + 1 s.e., and the mean +1 s.e.
kl	The KL statistic computed across the path
kl_min	The minimum KL statistic identified along the path
kl_mean	The expected KL statistic given the number of features identified

Author(s)

Benjamin A. Logsdon

References

Logsdon, B.A, G.E. Hoffman, and J.G. Mezey (2010) A variational Bayes algorithm for fast and accurate multiple locus genome-wide association analysis, http://www.biomedcentral.com/1471-2105/11/58, BMC Bioinformatics, Vol. 11(1), 58

Logsdon, B.A., G.E. Hoffman, and J.G. Mezey, (2012). Mouse obesity network reconstruction with a variational Bayes algorithm to employ aggresive false positive control, http://www.biomedcentral.com/1471-2105/13/53/, BMC Bioinformatics, Vol. 13(1), 53

Logsdon, B.A., C.L. Carty, A.P. Reiner, J.Y. Dai, and C. Kooperberg (2012). A novel variational Bayes multiple locus Z-statistic for genome-wide association studies with Bayesian model averaging. Bioinformatics, Vol. 28(13), 1738-1744

See Also

compute_KL

Examples

   n <- 100;
   m <- 500;
   ntrue <- 10;
   e <- rnorm(n);
   X <- matrix(rnorm(n*m),n,m);
   tbeta <- sample(1:m,ntrue);
   beta <- rep(0,m);
   beta[tbeta]<- rnorm(ntrue,0,.3);
   y <- X%*%beta;
   y <- y+e;


   res<- vbsr(y,X,family="normal");

vbsr documentation built on May 2, 2019, 9:27 a.m.