ud_init: Initialize Ultimate Deconvolution Model

In stephenslab/udr: Ultimate Deconvolution for Multivariate Normal Means

Initialize Ultimate Deconvolution Model

Description

Initialize an Ultimate Deconvolution model fit. See ud_fit for background and model definition.

Usage

ud_init(
  dat,
  V,
  n_rank1,
  n_unconstrained,
  U_scaled,
  U_rank1,
  U_unconstrained,
  control = list()
)

Arguments

dat	An n x m data matrix, in which each row of the matrix is an m-dimensional data point, or a “mash” object, for example created by mash_set_data. When dat is a matrix, it should have at least 2 rows and at least 2 columns.
V	Either an m x m matrix giving the residual covariance matrix for all n data points, or a list of m x m covariance matrices of length n.
n_rank1	A non-negative integer specifying the number of rank-1 covariance matrices included in the mixture prior. Initial estimates of the m x m rank-1 covariance matrices are generated at random. At most one of n_rank1 and U_rank1 should be provided. If neither are specified, rank-1 matrices will not be included.
n_unconstrained	A non-negative integer specifying the number of unconstrained covariance matrices included in the mixture prior. Initial estimates of the m x m covariance matrices are generated at random. At most one of n_unconstrained and U_unconstrained should be provided. If neither are specified, 4 random unconstrained matrices will be included.
U_scaled	A list specifying initial estimates of the scaled covariance matrices in the mixture prior. The default setting specifies two commonly used covariance matrices. In the case of a single scaled matrix, U_scaled may be a matrix instead of a list.
U_rank1	A list specifying initial estimates of the rank-1 matrices in the mixture prior. At most one of n_rank1 and U_rank1 should be provided. If U_rank1 is not given, the rank-1 covariates are initialized at random. In the case of a single rank-1 matrix, U_rank1 may be a matrix instead of list.
U_unconstrained	A list specifying initial estimates of the unconstrained matrices in the mixture prior. At most one of n_unconstrained and U_unconstrained should be provided. If U_unconstrained is not given, the matrices are initialized at random. In the case of a single unconstrained matrix, U_unconstrained may be a matrix instead of a list.
control	A list of parameters controlling the behaviour of the model initialization. See ud_fit for details.

Value

An Ultimate Deconvolution model fit. See ud_fit for details.

See Also

ud_fit

Examples

# Here we illustrate the different ways to initialize a UD model fit,
# with and without mashr. In the very simplest invocation, we supply a
# data matrix X.
set.seed(1)
X <- matrix(rnorm(2000),200,10)
fit <- ud_init(X)

# For more sophisticated usage, first create a mash data object with
# mash_set_data, then call ud_init. By setting alpha = 1, we invoke
# the EZ ("exchangeable z-scores") model, and with the model the
# residual covariance matrix is the same for all effects (rows of
# Bhat), and is the same as the V in the mash object.
library(mashr)
V <- matrix(0.1,10,10)
diag(V) <- 1
Shat <- sqrt(matrix(rexp(2000),200,10))
out <- simple_sims(200,10,Shat)
dat <- mash_set_data(out$Bhat,out$Shat,V = V,alpha = 1)
fit <- ud_init(dat)
all.equal(dat$V,fit$V,check.attributes = FALSE)

# By setting alpha = 0, we instead use the EE ("exchangeable effects")
# model. Now the residual covariance is different for each effect, and
# these covariances are now stored in a list with one element for
# every row of Bhat (or X).
dat <- mash_set_data(out$Bhat,out$Shat,V = V,alpha = 0)
fit <- ud_init(dat)
class(fit$V)

# Again we use the EE model, but since the standard errors are the
# same for all the effects (and the s.e.'s are equal to 1), the
# residual covariance matrix is now the same for all effects.
dat <- mash_set_data(out$Bhat,Shat = 1,V = V,alpha = 0)
fit <- ud_init(dat)
all.equal(dat$V,fit$V,check.attributes = FALSE)

# In this final example, we customize the mixture prior to include
# different numbers of unconstrained and rank-1 matrices.
fit <- ud_init(dat,n_rank1 = 2,n_unconstrained = 4)
names(fit$U)

stephenslab/udr documentation built on June 2, 2024, 3:04 a.m.