BLiP: Given samples from a posterior or a list of candidate groups,...

In amspector100/blipr: A Bayesian Linear Program for resolution-adaptive signal detection

Given samples from a posterior or a list of candidate groups, BLiP performs resolution-adaptive signal detection to maximize power while controlling (e.g.) the FDR.

Description

Note: when working with image data or a continuous set of locations, consider using BLiP_cts.

Usage

BLiP(
  samples = NULL,
  cand_groups = NULL,
  weight_fn = "inverse_size",
  error = "fdr",
  q = 0.1,
  max_pep = 1,
  deterministic = T,
  verbose = F,
  perturb = T,
  max_iters = 100,
  search_method = "binary",
  solver = NULL
)

Arguments

samples	(N,p)-shaped matrix of posterior samples where a nonzero value indicates the presence of a signal.
cand_groups	A list of lists, where the inner lists must have a "group" attribute, corresponding to the features in the group and a "pep" attribute, corresponding to a posterior error probability.
weight_fn	How to weight discoveries. Can be one of 'inverse_size' or 'log_inverse_size' or a function which takes a candidate group as an input and returns a weight.
error	Bayesian error rate to control: one of "fwer", "pfer", "fdr", "local_fdr".
q	The level at which to control the Bayesian FWER/PFER/FDR/local FDR.
max_pep	Never select any group with a pep greater than max_pep.
deterministic	Whether or not BLiP should return a deterministic solution.
verbose	If TRUE, gives occasional progress reports.
perturb	If TRUE, adds a tiny (random) perturbation to the weights to ensure the existence of a unique optimal solution.
max_iters	Maximum number of binary-search iterations for FWER when.
search_method	For FWER control, how to find the optimal parameter for the LP. Either 'binary' (defalt) or 'none'.
solver	The solver to use within CVXR. By default, will use Gurobi, CBC, or ECOS (in that order), depending on whether they are installed.

Value

A list of candidate groups, which asserts that each group contains a signal.

Examples

# Example 1: sparse linear regression
set.seed(123); n <- 100; p <- 200
data <- blipr::generate_regression_data(n=n, p=p)
# sample from the posterior, e.g., using NPrior
nprior <- NPrior::NPrior_run(
  X=data$X, y=data$y, N=5000, prior='SpSL-G'
)
# run blip on posterior samples
detections <- blipr::BLiP(samples=t(nprior$ThetaSamples), q=0.1, error='fdr')

# Example 2: Running BLiP directly on candidate groups
cand_groups <- list(
  list(group=c(1), pep=0.1),
  list(group=c(2), pep=0.5),
  list(group=c(1,2), pep=0.01)
 )
detections <- blipr::BLiP(cand_groups=cand_groups, q=0.1, error='fdr')

amspector100/blipr documentation built on June 22, 2022, 2:43 a.m.