R: Estimate (Local) False Discovery Rates For Diverse Test...
fdrtool
R Documentation
Estimate (Local) False Discovery Rates For Diverse Test Statistics
Description
fdrtool takes a vector of z-scores (or of correlations, p-values,
or t-statistics), and estimates for each case both the tail area-based Fdr
as well as the density-based fdr (=q-value resp. local false discovery rate).
The parameters of the null distribution are
estimated adaptively from the data (except for the case of p-values where
this is not necessary).
one of "normal" (default), "correlation", "pvalue".
This species the null model.
plot
plot a figure with estimated densities, distribution functions,
and (local) false discovery rates.
verbose
print out status messages.
cutoff.method
one of "fndr" (default), "pct0", "locfdr".
pct0
fraction of data used for fitting null model - only if cutoff.method="pct0"
color.figure
determines whether a color figure or a black and white
figure is produced (defaults to "TRUE", i.e. to color figure).
Details
The algorithm implemented in this function proceeds as follows:
A suitable cutoff point is determined. If cutoff.method
is "fndr" then first an approximate null model is fitted and
subsequently a cutoff point is sought with false nondiscovery
rate as small as possible (see fndr.cutoff).
If cutoff.method is "pct0"
then a specified quantile (default value: 0.75) of the data
is used as the cutoff point. If cutoff.method equals
"locfdr" then the heuristic of the "locfdr" package (version 1.1-6)
is employed to find the cutoff (z-scores and correlations only).
The parameters of the null model are estimated from the
data using censored.fit. This results
in estimates for scale parameters und and proportion
of null values (eta0).
Subsequently the corresponding p-values are computed, and
a modified grenander algorithm is employed
to obtain the overall density and distribution function
(note that this respects the estimated eta0).
Finally, q-values and local fdr values are computed for each case.
The assumed null models all have (except for p-values) one free
scale parameter. Note that the z-scores and the correlations
are assumed to have zero mean.
Value
A list with the following components:
pval
a vector with p-values for each case.
qval
a vector with q-values (Fdr) for each case.
lfdr
a vector with local fdr values for each case.
statistic
the specified type of null model.
param
a vector containing the estimated parameters (the null
proportion eta0 and the free parameter of the null model).
# load "fdrtool" library and p-values
library("fdrtool")
data(pvalues)
# estimate fdr and Fdr from p-values
data(pvalues)
fdr = fdrtool(pvalues, statistic="pvalue")
fdr$qval # estimated Fdr values
fdr$lfdr # estimated local fdr
# the same but with black and white figure
fdr = fdrtool(pvalues, statistic="pvalue", color.figure=FALSE)
# estimate fdr and Fdr from z-scores
sd.true = 2.232
n = 500
z = rnorm(n, sd=sd.true)
z = c(z, runif(30, 5, 10)) # add some contamination
fdr = fdrtool(z)
# you may change some parameters of the underlying functions
fdr = fdrtool(z, cutoff.method="pct0", pct0=0.9)