local_mfdr: Estimate local mFDR for all features

In pbreheny/ncvreg: Regularization Paths for SCAD and MCP Penalized Regression Models

Estimate local mFDR for all features

Description

local_mfdr() is called by summary.ncvreg(), which typically offers a more convenient interface to users. If, however, you are working with local mfdrs programmatically rather than interactively, you probably want to use local_mfdr(), which skips the sorting, filtering, and print formatting of summary.ncvreg().

Usage

local_mfdr(
  fit,
  lambda,
  X = NULL,
  y = NULL,
  method = c("ashr", "kernel"),
  sigma,
  ...
)

Arguments

fit	A fitted ncvreg or ncvsurv object.
lambda	The value of lambda at which inference should be carried out.
X, y	The design matrix and response used to fit the model; in most cases, it is not necessary to provide X and y as they are returned by ncvreg, but see the returnX argument in ncvreg().
method	What method should be used to calculate the local fdr? Options are ashr (which tends to be more accurate) and kernel (which requires no additional packages). The default is to use ashr if the package is installed.
sigma	For linear regression models, users can supply an estimate of the residual standard deviation. The default is to use RSS / DF, where degrees of freedom are approximated using the number of nonzero coefficients.
...	Additional arguments to ashr::ash() if using method='ashr'.

Value

If all features are penalized, then the object returns a data frame with one row per feature and four columns:

• Estimate: The coefficient estimate from the penalized regression fit

• z: A test statistic that approximately follows a standard normal distribution under the null hypothesis that the feature is marginally independent of the outcome

• mfdr: The estimated marginal local false discovery rate

• Selected: Features with nonzero coefficient estimates are given an asterisk

If some features are penalized and others are not, then a list is returned with two elements: pen.vars, which consists of the data frame described above, and unpen.vars, a data frame with four columns: Estimate, SE, Statistic, and p.value. The standard errors and p-values are based on a classical lm/glm/coxph model using the effect of the penalized features as an offset.

See Also

summary.ncvreg()

Examples

# Linear regression
data(Prostate)
fit <- ncvreg(Prostate$X, Prostate$y)
local_mfdr(fit, 0.1)

fit <- ncvreg(Prostate$X, Prostate$y, penalty.factor=rep(0:1, each=4))
local_mfdr(fit, 0.1)

# Logistic regression
data(Heart)
X <- Heart$X
y <- Heart$y
fit <- ncvreg(X, y, family='binomial')
local_mfdr(fit, 0.1)

# Cox regression
data(Lung)
X <- Lung$X
y <- Lung$y
fit <- ncvsurv(X, y)
local_mfdr(fit, 0.1)

pbreheny/ncvreg documentation built on April 25, 2024, 5:22 a.m.