GGtools (version 5.8.0)

pifdr: utility for computing plug-in FDR

Description

utility for computing plug-in FDR

Usage

pifdr( obs, perms, legacy=FALSE, trimToUnit = TRUE, ... )

Arguments

obs

observed association scores

perms

vector of association scores under permutation; length should be integer multiple of length(obs)

legacy

logical, if TRUE, use the approximate version of pifdr() available before 12/30/2013, with additional arguments if desired

trimToUnit

logical, if TRUE, values greater than 1 are replaced by 1. Such values can occur, for example, with relatively small sample sizes.

...

extra arguments passed if legacy is TRUE

Value

• vector of plug-in FDR estimates congruent to obs

Details

Revised 12/30/13 to employ hist() to rapidly bin the permuted values.

Use legacy=TRUE to obtain the approximate implementation, for which the following remarks held: ``As currently implemented the algorithm is quadratic in length(obs). While it is possible to get a unique FDR value for every element of obs, an approximate approach yields practically identical precision and by default this will be used for obs with length 2000 or more. In this case, approx is used with rule=2 to interpolate from the grid-based FDR estimates back to the data values.'' Additional parameters npts and applier may be supplied if legacy is set to TRUE.

npts defined the number of points spanning the range of obs to be used for a lossy grid-based computation only used if length(obs)>npts.

applier is to be an sapply-like function.

References

Hastie Tibshirani and Friedman Elements of Statistical Learning ch 18.7

Examples

set.seed(1234)
op = par(no.readonly=TRUE)
par(mfrow=c(2,2))
X = c(rchisq(30000,1),rchisq(300,10))
Y = rchisq(30300*3,1)
qqplot(Y, X, xlab="null", ylab="observed")
hist(pp <- pifdr(X,Y), xlab="plug-in FDR", main="")
library(multtest)
rawp = 1-pchisq(X, 1)
MT <- mt.rawp2adjp(rawp) 
MT2 = MT[[1]][order(MT[[2]]),]
plot(MT2[,"BH"], pp, xlab="BH FDR", ylab="plug-in FDR")
par(op)